This is the fourth article in the Briefing Room series examining major failures of digital health companies. The previous three articles covered Olive AI, IBM Watson Health, and Carbon Health. All three are content for my paid subscribers.

It takes a lot of work to research and write an article like this. Please consider becoming a paid subscriber to support the work.

I have a limited-time 50% off offer on the 1-year paid subscription available to you here.

Get 50% off for 1 year

Each healthcare technology company was built on a different premise, in a slightly different market segment, with a different funding structure. And yet, each illustrates a remarkably consistent pattern of underlying reasons for their failure.

As you’ll see, Babylon Health is no different.

But first, a short piece of cultural history…

The Trophy of the Conqueror

In Persian national identity, Babylon is not a symbol of decadence. It is a trophy.

When Cyrus the Great marched into Babylon in 539 BCE, he did not need to besiege it. He walked through the gates because the city had already lost faith in its own leadership. The city did not fall to an assault from outside. It fell to an internal collapse of credibility. Cyrus then cast himself as the restorer. He freed the captives, rebuilt the temples, and inherited a civilization.

The king he replaced was a man named Nabonidus. He is the other figure in this story, and he requires a proper introduction.

Nabonidus was the last king of Babylon, and he was not, by most accounts, a cruel or stupid man. He was, in fact, deeply intellectual. His problem was a very specific kind of overconfidence: he became so absorbed in his own vision and convictions that he stopped maintaining the institutional relationships his authority depended on. Babylon’s state religion was built around Marduk, the city’s patron deity.

Nabonidus preferred a different god entirely, the moon deity Sin, and he pursued that personal framework with the certainty of a man who had stopped requiring the agreement of the people around him. He left Babylon for extended periods, sometimes years at a stretch, delegating governance to his son Belshazzar while he pursued his own agenda elsewhere.

The priesthood of Marduk, whose endorsement was the foundation of any Babylonian king’s legitimacy, found itself marginalized by a ruler whose grand vision had no room for their institutional role.

When Cyrus arrived from Persia with his army, there was almost nothing left to defend. The priests did not resist the Persians. They welcomed him, because Cyrus, at least, would restore the rituals and respect the temple. Nabonidus had already surrendered Babylon’s internal legitimacy long before a single Persian soldier reached the walls.

Cyrus won Babylon. Nabonidus lost it.

Now, if we know one thing about history, the history books are written by the victor, so we must add that color to the story, but stick with me for a moment as we pivot to Babylon Health.

The Persian at the Gates

Ali Parsa is an Iranian-born. He was born in Rasht, in Gilan Province, and left at 16 when the Iranian Revolution made such a departure necessary. He came to Britain, taught himself the qualifications he needed, studied at University College London, spent years in investment banking with Merril Lynch and Goldman Sachs, before turning his attention to healthcare to build companies.

In 2013, he founded his technology fuelled healthcare services company, which he named Babylon Health.

For an Iranian founder, that name carries weight it does not carry for others. Babylon, in the Persian tradition, is the city your ancestors defeated. It is the city Cyrus walked into without a fight and liberated. Naming a healthcare company after it is, whether conscious or not, a statement of intent.

Parsa was positioning himself as the ‘Persian at the gates’ of a broken system, about to walk in without a siege, because the old order had already lost the people’s faith it was supposed to serve. He was going to liberate them through the democratizing force of the web and, later, artificial intelligence.

That framing give color to every interview, every funding round, and every government endorsement Babylon Health attracted over the following decade. It makes a compelling story.

But in the end, it was the wrong story. And the reason it was the wrong story is that Parsa was’t Cyrus the Great.

In the end, he looked closer to Nabonidus.

Cyrus, in the Babylonian record, is the restorer who enters a failing order and wins the legitimacy to inherit it. Nabonidus is the ruler who loses touch with the institutional core of his own Babylon. As you’ll see, that is why the historical metaphor fits the Babylon Health story so well.

The real failure was not merely commercial. It was a failure of legitimacy, of overreaching visionary confidence, of hubris, all while underestimating the durable power of operational and technical reality, clinical credibility, and market trust.

The Founding Promise

Ali Parsa founded Babylon Health in London in 2013. He was not new to ambitious NHS-adjacent ventures. A decade earlier, he had co-founded Circle Health, a private hospital operator that grew to become one of Europe’s largest independent NHS service providers. Circle took over Hinchingbrooke NHS Trust in 2012, in what was, at the time, the first private-sector takeover of an NHS hospital. The contract ran at a deficit. Hinchingbrooke was handed back to NHS management in 2015.

Read more

I recently connected with the COO of a large, well-known commercial entity. He and I have known each other for a bunch of years. I remember he was one of the first very smart people who introduced me to the market-driven product management methodology known as the Pragmatic Framework. What he said next was shocking to me.

That shook me, as I have always known him to be a rational, scientific guy. But suddenly he’s all on board with,

never mind the quality, feel the volume.

Right now, somewhere, a product team is demoing a feature that took three days to build. It looks impressive. The stakeholders are nodding. Nobody in the room is asking whether the user it was built for actually has the problem it purports to solve.

That question used to get asked, not because product teams were wiser, but because the build took six weeks and someone had to justify the spend. The friction was the filter. The filter is gone.

Discernment

At its root, discernment is the capacity to see clearly what is actually there, not what you hope is there, fear is there, or have been told is there.

It’s the cognitive and moral discipline of separating signal from noise, and then having the courage to act on what the signal tells you, even if it’s inconvenient. It’s distinct from analysis.

Applied to a product feature decision, it operates like this:

1. Whose voice is loudest isn’t whose voice matters most.

A vocal customer, an excited sales rep, a CEO who just saw a competitor demo — these create pressure that *feels* like signal. Discernment asks: does this request represent a pattern in the market, or a personality in the room?

2. Feasibility and desirability are not the same as viability.

You *can* build it. Users *want* it. Neither answers whether you *should*. Discernment holds all three lenses simultaneously and refuses to let enthusiasm in one dimension crowd out honest assessment in the others.

3. What problem is this feature actually solving — and for whom?

Not the stated problem. The actual one. Discernment presses past the surface request into the underlying job-to-be-done, and then asks the harder question: is solving *that* job our job?

4. What does saying yes cost that isn’t on the estimate?

Every feature carries a second price tag that rarely appears in sprint planning: maintenance burden, UX complexity, support load, and architectural constraint on future decisions. Discernment sees the long tail, not just the build cost.

5. Reversibility.

Can you undo this? Shipped features create expectations. Expectations become dependencies. Dependencies become obligations. Discernment treats a feature decision as closer to a hire than a task — you’re not just adding capability, you’re making a commitment.

If discernment were a filter, it would sound like this:

Five rights. Most features that should be killed pass two or three. Discernment holds out for all five, or at a minimum, requires you to consciously own which ones you’re compromising and why.

What Vibe Coding Has Done to This

AI-assisted development has collapsed the cost of building without touching the cost of being wrong. That’s the asymmetry nobody is pricing in.

When a feature took six weeks of engineering time, the friction of the build was itself a discernment mechanism. You had to want it badly enough to justify the spend. Disagreements surfaced. Corners got examined. The cost created a forcing function for judgment.

At vibe-coding speed, that friction disappears. With it goes the incidental discipline it was providing. You can now ship in days what used to take months. Which means the only thing standing between a half-formed idea and a production feature is the judgment of whoever typed the prompt.

And that judgment is, in most organizations, completely underdeveloped for this moment.

The result is a signal-to-noise ratio of roughly 1-in-5 to 1-in-10 features delivering any meaningful impact. That’s not a failure rate. It’s evidence that no discernment filter is operating upstream of the build.

Buy me a coffee?

What’s filling the vacuum where judgment should be:

• Velocity theatre — shipping fast as a proxy for delivering value

• Feature FOMO — if the competitor has it, we need it, regardless of fit

• Prompt optimism — the ease of generation creates an illusion of validation

• Demo culture — things that look impressive in a demo get greenlit; things that solve real problems but demo poorly do not

The result is exactly what the user community experiences: a bloated feature surface area, most of which they never asked for, don’t understand, and quietly route around. And their resistance isn’t irrational. It’s a learned response to repeated disappointment.

The Absorption Limit

Organizations and users have a finite capacity to absorb change, learn new workflows, and integrate new capabilities into actual practice. Vibe coding at scale is a direct assault on that limit.

When you ship ten features to buy one or two meaningful ones, you’re not just wasting build capacity. You’re **spending down user trust and attention** on features that don’t land — which means when the genuinely impactful ones arrive, the user community has already developed the antibodies.

This is how shelfware compounds. Each low-impact feature makes the next feature slightly harder to adopt, regardless of its merit. The product surface grows. The adoption rate shrinks. And somewhere in a QBR, a slide appears showing feature utilization data that nobody in the room wants to discuss honestly.

The Discernment Litmus Test

For leaders who suspect their roadmap is a hallucination.

The discipline needed isn’t slowing down the build; that ship has sailed. It’s inserting a judgment gate upstream of the prompt. Not a process. A confrontation.

These five tests exist to force a choice between looking busy and being effective.

1. The Ghost Test — The Right Problem.

Is the problem we are solving an observed, bleeding-neck frustration in the market, or is it a prompt-friendly hallucination? If you cannot point to where the user is currently struggling without your help, you are not solving a problem. You are just filling a bucket.

2. The Receipts Test — The Right User.

Can you name five actual customers who will have a measurably better day because of this? If the answer is “Personas” or “The Market,” you are shipping to a ghost audience. Real discernment requires real names.

3. The Demo Test — The Right Solution.

Are you building this because it elegantly solves the problem, or because it looks impressive in a thirty-second Loom for the Board? Discernment is the capacity to kill a feature that demos like magic but functions like a labyrinth.

4. The Antibody Test — The Right Time.

Does the user have the cognitive surplus to learn this today? Every feature is a tax on user attention. If your velocity is faster than your user’s absorption, you are not delivering value. You are performing an assault.

5. The Identity Test — The Right Fit.

Just because the cost of the build is now zero does not mean the cost to your brand is zero. Does this feature reinforce who you are, or is it Feature FOMO dressed up in a new prompt? If it is not your job to solve this, saying yes is a form of identity theft.

In a vibe coding culture, these questions are routinely treated as friction — things that slow the team down.

But slowing down the build was never the goal. It was the accidental byproduct of having to think and then act deliberately.

Now that the thinking is optional, or compressed, these five tests are the only thing standing between a coherent product and the Sorcerer’s Apprentice flood.

The litmus test for leadership is simple:

What are you more afraid of?

An empty roadmap, or a bloated and useless one?

Answers in the comments, please!

Leave a comment

Keep havering

This is breaking news stuff that I am getting out and making free to everybody.

On April 4, Anthropic sent an email to subscribers.

Starting at noon Pacific, users could no longer apply their Claude subscription limits to OpenClaw and other third-party tools. If they wanted to continue, they would need to pay extra — on top of the subscription they were already paying. Anthropic’s head of Claude Code noted that the policy applies to all third-party harnesses and would be rolled out to more tools shortly.

That email surprised a lot of people, but not to anyone paying attention.

But it was a signal the market had been waiting for.

Functionality constraints have arrived.

This is no longer a Forecast. It is now!

Four weeks. One case study.

The timeline deserves to be read slowly because it is the argument.

March 13. Anthropic launches what it calls a “limited-time promotion” — doubling usage limits for subscribers who use Claude during off-peak hours. Framed as a gift. It was a shift in infrastructure management. Read it again. Carefully!

March 26. Anthropic tightens peak-hour session limits across all plans — free, Pro, and Max. Users burn through their five-hour session windows faster during core working hours. Approximately 7% of users hit walls they hadn’t encountered before. An Anthropic engineer posts the announcement on X and recommends shifting “token-intensive background jobs to off-peak hours.”

I personally experienced this during a working session that lasted a mere 45 minutes before I got the dreaded “limits” message and the offer to buy more.

March 31. Anthropic acknowledges publicly that users of Claude Code are “hitting usage limits way faster than expected.” The company calls it the team’s top priority.

April 4. Third-party tool access removed from all subscription tiers. The subscription that covered your tools no longer covers all your tools. Pay separately or do without.

That is not a pattern of growing pains. It is a pattern of deliberate constraint management that escalates in scope and transparency over three weeks. Each step framed as technical necessity, while the commercial architecture underneath becomes more explicit and more patent.

Buy me a Coffee?

The drug dealer analogy writes itself

The first hit is designed to be easy. Low friction. Full capability. Unremarkable monthly cost. The goal is not margin at this tier — the goal is habit. Workflow integration. Muscle memory.

Once the dependency is established, the economics of the arrangement can be renegotiated, and the user stays not because the contract says so, but because the alternative is rebuilding everything somewhere else.

We lived this before in the past 18 months. It’s the streaming wars.

Netflix at $7.99 trained an entire generation to expect on-demand content at the price of a paperback.

It eliminated alternatives, built an irreplaceable habit, and then, once the habit was established and the competition had followed it into the same warm, pleasant pool of subscription payment to escape the evil cable companies and DirecTV, began the systematic repricing of the arrangement.

• Prices rose.

• Password sharing became a violation.

• Tiers multiplied.

• Content fragmented.

Users stayed because the switching cost was no longer the subscription fee. It was the ritual.

The AI subscription market is running the same play. The mechanism is slightly different. The commercial logic is identical.

What these companies are actually doing

Plan-level profitability is not publicly disclosed. But profitability is not the right framing.

The right framing is, it’s an incentive structure.

These companies are acquiring user habits at scale, at a price low enough to generate no resistance. They are building workflow dependency through memory, projects, connectors, and agents.

They are segmenting demand. The $20 user subscription is the tease, the top of the pipeline, the slippery slope, not the end game.

The target is the business/enterprise contract, the team tier, the optional premium capability layer that becomes necessary once the base model no longer fully serves the workflow that has been built on top of it.

And they are running live experiments in customer price conditioning. When Anthropic offered the “promotion” doubling off-peak limits, it simultaneously taught users that normal limits are real and a matter of fact. And that peak hours are premium territory.

When it cut OpenClaw from subscriptions while offering discounted “extra usage bundles,” it converted a flat-rate relationship into a variable-cost one, specifically for the users doing the most work.

The OpenClaw move is particularly informative.

Users had discovered they could run what amounted to more than $1,000 in API-equivalent workloads through a $200/month Max subscription, a gap of more than 5x.

Anthropic just slammed that door shut.

I advise on this kind of commercial and price positioning strategy.

It’s part of Product Marketing.

If you need my help, reach out:

Book a meeting

We’ve seen these games with products over and over again. Physical products that suffer ‘shrinkflation’. Software products that have new feature sets that are highly desirable packaged behind locked switches that represent a new (priced) module.

Commercially, this was all so inevitable.

What matters is the method: a subscription that previously had implicit scope now has explicit limits. The shifting boundaries of the $20 and $200 arrangements are just being drawn out in public, in real time.

The migration that caused the surge

There is a backstory to the past three weeks that deserves its own paragraph.

The surge in Claude users that overwhelmed Anthropic’s infrastructure was not random. When OpenAI signed a contract with the Pentagon, Anthropic refused. The Department of Defense labeled Anthropic a supply chain risk. Anthropic fought the designation legally and secured a temporary stay. And a significant number of users who did not want their AI tools contracted to military purposes migrated from OpenAI to Claude.

That principled ethical positioning triggered a direct commercial consequence: a sudden surge in new subscribers that Anthropic’s infrastructure was not prepared to absorb. The throttling, the limit adjustments, the OpenClaw cutoff — all of it followed that surge.

The irony is clean. Anthropic’s stand on military AI generated the very capacity crisis that produced the functionality constraints its users are now experiencing.

OpenAI is not exempt

On April 2, four days ago, OpenAI CFO Sarah Friar told ARK Invest CEO Cathie Wood:

She added that sourcing available compute is where she now spends significant time.

OpenAI has already pulled back from Sora, its video generation product, to redirect resources to core revenue-generating priorities.

Two of the dominant players in the prosumer AI tooling market are openly acknowledging that demand is outrunning infrastructure. The computer shortage will not be resolved quickly. And the prosumer subscription tier (the $20 layer) is the most economically exposed part of the stack when the pressure has to go somewhere.

The Steinberger signal

OpenClaw’s creator, Peter Steinberger, has since announced he is joining OpenAI. Before leaving, he wrote: “Funny how timings match up — first they copy some popular features into their closed harness, then they lock out open source.”

That is the commercial playbook described in one sentence by someone who just lived it.

Incorporate the innovation from the open ecosystem. Replicate the most valuable features inside the proprietary tool. Close the gap. Then close the door.

Steinberger is not the first developer to describe this pattern. He is just the most recent and the most high profile.

What comes next

More of this. And faster!

The April 4 notice states explicitly that the policy applies to all third-party harnesses and will be rolled out to more shortly.

Expect tighter usage limits framed as “a matter of fairness” and “market reasonableness/fair value”. They will frame it as “it’s just business.”

Feature migration to premium tiers is framed as product improvement. More bundling makes the real cost of AI access harder to isolate or compare. Deeper integration of memory, projects, and agents raises the operational cost of switching even when the contractual cost remains nominally the same.

And watch for the quiet moves, the ones that don’t get announced, the ones you discover when suddenly your workflow stops working the way it did last Tuesday.

What to do about it

This is not an argument for abandoning tools that are genuinely useful. It is an argument for not being the last person in the room to notice.

• DO NOT build critical workflows inside a single platform. If it matters to your work, it should be reproducible somewhere else.

• KEEP your assets portable. Prompts, frameworks, outputs, research — archive them outside the platforms. They belong to you. Treat them as intellectual property, not platform content.

• DIVERSIFY across at least two vendors. Not paranoia — risk management. The professional fluent in two environments is half as exposed, not twice as complicated.

• WATCH THE FUNCTION, not just the price. The headline subscription fee will increase slowly, with short or no notice. The quiet reduction in what that number actually buys will move faster.

The streaming wars ended with consolidation and power moves. Everyone is paying more for less on all consumer entertainment platforms. They tried to dress that up as an inflationary response at first. We all saw through that in an instant. However, that outcome was predictable.

Consumers who noticed early adjusted on their own terms rather than getting stiffed by automatic subscription increases.

For AI tooling in the prosumer range? The cage was already being built around you.

As of April 4, they have just started turning the locks.

Prepare for the coming #functionalityconstraints.

SM

SOURCES

Anthropic off-peak promotion — March 13
https://www.pymnts.com/artificial-intelligence-2/2026/ai-usage-limits-are-becoming-the-new-reality-for-consumers/

Anthropic peak-hour session limits tightened — March 26 https://www.theregister.com/2026/03/26/anthropic_tweaks_usage_limits/ https://www.infoworld.com/article/4151196/anthropic-throttles-claude-subscriptions-to-meet-capacity.html

Pentagon user migration driving the surge https://www.pcworld.com/article/3103434/claudes-sudden-surge-in-users-is-slowing-it-down-for-everyone.html

Claude Code usage limits “way faster than expected” — March 31 https://www.theregister.com/2026/03/31/anthropic_claude_code_limits/ https://www.devclass.com/ai-ml/2026/04/01/anthropic-admits-claude-code-users-hitting-usage-limits-way-faster-than-expected/5213575

OpenClaw removed from all subscription tiers — April 4 https://techcrunch.com/2026/04/04/anthropic-says-claude-code-subscribers-will-need-to-pay-extra-for-openclaw-support/ https://www.xda-developers.com/claude-subscribers-just-lost-access-to-openclaw-and-other-third-party-toolsunless-they-pay-more/

Peter Steinberger quote and joining OpenAI https://www.pymnts.com/artificial-intelligence-2/2026/anthropic-says-claude-code-users-will-need-to-pay-more-for-openclaw/

$200 subscription powering $1,000+ API-equivalent workloads https://www.shareuhack.com/en/posts/openclaw-claude-code-oauth-cost

OpenAI CFO Sarah Friar compute shortage statement — April 2 https://www.aol.com/articles/openais-cfo-says-company-passing-112355988.html https://startupfortune.com/openai-says-its-turning-down-opportunities-because-it-doesnt-have-enough-compute/

Original Podcast from ARK Invest

https://www.ark-invest.com/podcast/openai-cfo-sarah-friar

OpenAI pulling back from Sora to redirect compute https://www.digit.in/news/general/openai-admits-it-is-skipping-opportunities-due-to-lack-of-compute.html

Okay, so this article is very much for the nerdy, the technical, and the medically curious. I know there are many people who might be curious about what all these tests are that occupy my two days, or maybe you’re just plain creepy and want to know what my values were! It is very comprehensive (long).

But for those who found this because they are curious to benchmark where they are in their own (or a loved one’s) post-transplant experience, you are welcome to message me, and I will help in any way I can.

Be very clear: I know how blessed I am to have the current status I have, as described here.

I am truly in the elite group of lung transplant recipients at 3 years with these results, and I respect that many others of my Lung transplant buddies are not so lucky and have struggles.

To you, I wish nothing but peace, comfort, and a better way forward.

This also serves as a detailed reference to the companion article, which you can find here. (Or if you received the email of this article, it should arrive momentarily)

This article covers all tests and clinical encounters from my two-day annual assessment at the Mayo Clinic Transplant Center in Rochester, Minnesota, on April 2 and 3, 2026.

For each item, I have used the same five-part format throughout:

• What the test is,

• Why is it used?

• What it involves for the patient,

• What my result was,

• What did that result mean in my specific situation as a 63-year-old male, three years after a double lung transplant for pulmonary fibrosis?

The assessment runs across two days and multiple departments.

Day one covered blood and urine collection, cardiac rhythm, kidney function, lung function, and clinical review.

Day two covered cardiac imaging, chest imaging, exercise testing, and the final physician consultation.

Not every result below is dramatic. Several are routine confirmations of stability. That stability, accumulated test by test, is itself the finding.

They are to use my wife’s word — BORING!

BORING is good!

Day One — April 2, 2026

LABS, LOTS OF LABS

1. Complete Blood Count (CBC) with Differential

What it is: A foundational blood test that counts and examines the types of cells circulating in the blood: red cells, which carry oxygen; white cells, which manage immune response; and platelets, which handle clotting. The differential breaks the white cell count into its component types — neutrophils, lymphocytes, monocytes, eosinophils, and basophils.

Why they use it: For a transplant recipient, the CBC does several things simultaneously. It checks for anaemia, which immunosuppressant drugs can cause over time. It monitors white cell counts, because the same drugs that suppress rejection also broadly suppress immune function. It watches for signs of infection and for the early footprint of certain cancers — including lymphoma, for which immunosuppressed patients carry elevated risk.

What it involves for the patient: A standard blood draw, one of several tubes taken as part of the early morning collection, before medications or food.

My result: Hemoglobin 13.6 g/dL. Platelets 268. White blood cells 8.5. Red blood cells 4.58. Monocytes slightly elevated at 1.23. All other values within normal range.

What that meant in my case: A stable blood picture across the board. The mild monocyte elevation is a commonly observed pattern in transplant recipients and is not, in isolation, clinically concerning. The overall CBC has remained consistent across multiple assessment cycles. Consistency is the target, and this was consistent.

2. Basic Metabolic Panel (BMP)

What it is: A group of blood tests covering electrolytes (sodium, potassium, chloride, bicarbonate), kidney function (creatinine, BUN, eGFR), and blood sugar (glucose).

Why they use it: The anti-rejection medications required after lung transplant — tacrolimus and mycophenolate primarily — are nephrotoxic. That is medical shorthand for “can damage the kidneys over time.” This panel is the primary monitoring tool for that risk. Electrolyte balance also matters because transplant medications can cause shifts in potassium and magnesium that affect heart rhythm and muscle function.

What it involves for the patient: A standard blood draw, part of the morning collection.

My result: Electrolytes within normal limits. Creatinine 1.46 (slightly above the standard upper limit of 1.35). Estimated GFR 54.

What that meant in my case: These kidney numbers look like a flag but are not, in context. Creatinine in the low-to-mid 1.4s is my established baseline — it has been consistent at this level for well over a year. My transplant team tracks it not against the general reference range, but against my own historical line. Stability is the objective, and it is stable. The eGFR of 54 indicates mild reduction in kidney function, a known and accepted consequence of long-term calcineurin inhibitor therapy. The formal iothalamate test later in the day gives a more precise measurement of what that number actually means.

3. Hepatic Function Panel

What it is: A panel of blood tests measuring liver enzymes (AST, ALT, alkaline phosphatase), bilirubin, albumin, and total protein. Together, they provide a picture of liver health and synthetic function.

Why they use it: Several of the long-term medications required after transplant — including tacrolimus and azithromycin — can stress the liver over time. This panel confirms the liver is tolerating the medication burden without accumulating damage.

What it involves for the patient: A standard blood draw.

My result: All markers within normal limits. AST 23, ALT 17, Alkaline Phosphatase 48, Albumin 4.6.

What that meant in my case: A clean liver panel three years into a complex medication regimen. The albumin at 4.6 is particularly reassuring — it is a marker of nutritional status and liver synthetic function, and mine is solidly normal.

4. Lipid Panel

What it is: A fasting blood test measuring total cholesterol, LDL (the type associated with cardiovascular risk), HDL (the protective type), and triglycerides.

Why they use it: Certain immunosuppressant drugs can drive up cholesterol levels, increasing long-term cardiovascular risk. Transplant recipients are already at elevated risk of cardiovascular disease, so lipid management is treated as a priority from the outset.

What it involves for the patient: A blood draw following at least eight hours of fasting.

My result: Total cholesterol 157. LDL 65. HDL 70. Triglycerides 128.

What that meant in my case: An LDL of 65 is below the target threshold used even for high-risk cardiovascular patients. An HDL of 70 is well above the desirable minimum. The rosuvastatin I take daily is doing its job. The overall cardiovascular risk profile from these numbers is low — a meaningful achievement given the challenges the medication regimen typically poses in this area.

5. C-Reactive Protein (CRP)

What it is: A protein produced by the liver in response to inflammation anywhere in the body. The blood test measures how much of it is currently circulating.

Why they use it: CRP is a non-specific marker — it identifies that something is causing inflammation but not what. For transplant patients, elevated CRP can be an early signal of infection, rejection, or immune activation. A low result is a meaningful negative.

What it involves for the patient: A standard blood draw.

My result: Less than 3.0 mg/L.

What that meant in my case: No significant systemic inflammation detected. The clinical threshold of concern is 5.0 mg/L. Being below 3.0 provides reassurance that nothing is silently brewing — no occult infection, no immune activity against the graft.

6. Fasting Glucose

What it is: A measurement of blood sugar after a period of fasting, typically eight hours or more.

Why they use it: Steroids and calcineurin inhibitors — both essential to post-transplant immunosuppression — can impair the body’s ability to regulate blood sugar, a condition called post-transplant diabetes mellitus. Fasting glucose is one of two tools used to monitor this risk.

What it involves for the patient: No food or drink other than water from the previous evening. The draw happens first thing in the morning as part of the batch collection.

My result: 93 mg/dL, within the normal range of 70–100 mg/dL.

What that meant in my case: Normal fasting glucose three years post-transplant, despite being on daily prednisone and tacrolimus. Read alongside the HbA1c below, this provides a consistent picture of well-managed blood sugar.

7. Hemoglobin A1c (HbA1c)

What it is: A blood test reflecting average blood sugar levels over approximately the previous three months. It measures the percentage of hemoglobin — the protein in red blood cells — that has glucose permanently attached to it.

Why they use it: Unlike the fasting glucose test, which captures a single moment, HbA1c shows the longer-term trend. It is the standard tool for monitoring post-transplant diabetes mellitus and for catching patterns before they become clinical problems.

What it involves for the patient: A standard blood draw. No special preparation beyond the fasting already in place for the glucose test.

My result: 5.5%, within the normal range of 4.0–5.6%.

What that meant in my case: Normal long-term blood sugar control, sitting at the top of the normal range but not into pre-diabetic territory. The combination of a normal fasting glucose and a normal HbA1c is a clean result on glucose management — meaningful given the diabetogenic potential of my medication regimen. I am currently managed with weekly Mounjaro injections, which is doing what it needs to do.

8. Tacrolimus Trough Level

What it is: A blood test measuring the concentration of tacrolimus — my primary anti-rejection medication — in the bloodstream at its lowest point in the dosing cycle.

Why they use it: Tacrolimus operates within a narrow therapeutic window. Too low, and the immune system begins moving against the transplanted lungs. Too high, and the drug itself causes kidney damage, neurological side effects, and metabolic disruption. The trough level, drawn at the nadir before the next dose, is the most reliable way to confirm the drug is positioned where it needs to be.

What it involves for the patient: Timing is critical. The blood must be drawn exactly twelve hours after the previous dose, before taking the morning dose. This is not approximate. I set a specific alarm and do not deviate from it.

My result: 7.5 ng/mL, within the goal range of 6–9 ng/mL.

What that meant in my case: Solid immunosuppressive coverage without excessive accumulation. A result that requires consistent, disciplined medication management to produce — and that takes months to establish. It is not a number you arrive at by accident.

9. Thyroid-Stimulating Hormone (TSH)

What it is: A blood test measuring TSH, a hormone produced by the pituitary gland that regulates thyroid function. High TSH suggests an underactive thyroid; low TSH suggests an overactive one.

Why they use it: Long-term medication use and the physiological demands of major surgery can affect thyroid function. Thyroid imbalance causes fatigue, weight changes, and cardiac effects — symptoms that could otherwise be incorrectly attributed to other post-transplant issues.

What it involves for the patient: A standard blood draw.

My result: 1.5 mIU/L, normal range.

What that meant in my case: Normal thyroid function. No contribution from this variable to any fatigue or metabolic concerns.

10. Prothrombin Time (PT) / INR

What it is: A test of blood clotting speed. The INR is a standardised version of the result that allows consistent comparison across different laboratories.

Why they use it: Coagulation status reflects liver synthetic function and provides a safety baseline for ongoing care. It can also detect interactions between medications that affect clotting.

What it involves for the patient: A standard blood draw.

My result: PT 11.1 seconds. INR 1.0. Both normal.

What that meant in my case: Normal clotting function. Normal liver synthetic activity. A clean baseline across the board.

11. PSA (Prostate-Specific Antigen), Total and Free

What it is: A blood test measuring Prostate-Specific Antigen, a protein produced by the prostate gland. The total and free fractions are used together to screen for prostate health and assess the risk of prostate cancer.

Why they use it: Immunosuppressed patients carry modestly elevated risk of certain cancers, including prostate cancer. Annual PSA screening is part of the broader cancer surveillance protocol.

What it involves for the patient: A standard blood draw.

My result: Total PSA 1.2 ng/mL. Free PSA 0.5 ng/mL.

What that meant in my case: A reassuring result. A total PSA well below the 4.0 ng/mL threshold of concern for my age group. No current indication of prostate pathology.

12. Vitamin D (25-Hydroxyvitamin D)

What it is: A blood test measuring total Vitamin D stored in the body.

Why they use it: Long-term steroid use accelerates bone density loss. Vitamin D is essential for calcium absorption and bone maintenance. Monitoring it is part of the broader bone health protocol for transplant patients — relevant in my case given a DEXA scan in April 2025 that identified low bone mineral density.

What it involves for the patient: A standard blood draw.

My result: 25-Hydroxy D3: 50 ng/mL. 25-Hydroxy D2: less than 4.0 ng/mL. Total: 50 ng/mL.

What that meant in my case: Sitting at the top of the optimal range of 20–50 ng/mL. The result is almost entirely D3 — the natural and supplemented form — rather than D2, the plant-derived form. That means the body is converting and storing Vitamin D in the right form and at the right level. Dietary intake and any supplementation are effectively supporting bone metabolism.

13. Testosterone, Total

What it is: A blood test measuring total testosterone, the primary male sex hormone.

Why they use it: Long-term steroid use and the physiological demands of transplant recovery can suppress testosterone levels, contributing to fatigue, muscle loss, and reduced bone density — all concerns for long-term transplant outcomes.

What it involves for the patient: A standard blood draw.

My result: 466 ng/dL, within the normal range of 240–950 ng/dL for a male of my age.

What that meant in my case: Healthy testosterone levels supporting overall vitality and physical function. No deficit from this variable.

14. CMV and EBV PCR (Viral Load Testing)

What it is: Highly sensitive molecular tests that look for the genetic material of two viruses: Cytomegalovirus (CMV) and Epstein-Barr Virus (EBV). PCR — polymerase chain reaction — is a laboratory technique that can detect vanishingly small quantities of viral DNA.

Why they use it: Both viruses are extremely common and normally kept dormant by a healthy immune system. Immunosuppression removes much of that control. Reactivation in a transplant recipient can cause serious illness and directly damage the transplanted organ. Monitoring catches reactivation early, when it can be treated.

A note on my specific profile: I am CMV D-/R-, meaning my donor was also CMV-negative, so no new CMV was introduced by the transplant. I am EBV R+, meaning I carried prior EBV exposure before the transplant. Both designations inform how the team monitors me.

What it involves for the patient: A standard blood draw.

My result: CMV PCR: undetected. EBV PCR: undetected.

What that meant in my case: Neither virus is active. The level of immunosuppression I maintain is achieving the intended balance: suppressing rejection without opening the door to opportunistic viral infection. This is the best possible result on both counts.

15. Urinalysis (Dipstick)

What it is: A rapid chemical screening of a urine sample using a reagent strip that reacts to the presence of specific substances.

Why they use it: To check for early signs of kidney stress, urinary tract infection, glucose spillage (a marker of diabetes), or blood. A fast, low-cost front-line screen that can prompt further investigation if anything appears.

What it involves for the patient: Providing a clean-catch midstream urine sample in a sterile container.

My result: Negative for blood, glucose, ketones, nitrites, and leukocyte esterase.

What that meant in my case: A clean screen. Despite the mildly elevated creatinine in the blood panel, the kidneys are not leaking protein or blood into the urine. No evidence of silent infection.

16. Microscopic Urinalysis

What it is: A laboratory examination of the urine sample under a microscope, looking for cells, crystals, or other microscopic material that a dipstick cannot detect.

Why they use it: To identify microscopic blood or white blood cells that could indicate early kidney inflammation or infection not visible on the dipstick screen.

What it involves for the patient: The same urine sample. No additional collection required.

My result: No red blood cells. No white blood cells. Microscopy normal.

What that meant in my case: Normal at the cellular level. Complementary to the dipstick result. Together, they provide a clean urinary picture.

17. Urine Osmolality and pH

What it is: Two additional measurements from the same urine sample. Osmolality measures the concentration of dissolved particles in the urine, reflecting how well the kidneys are concentrating and diluting. pH measures the acid-base balance.

Why they use it: Both provide supplementary data on kidney function and metabolic balance. Neither is used in isolation but both contribute to the overall picture.

What it involves for the patient: Same urine sample. No additional collection.

My result: Osmolality 455 mOsm/kg (normal range 150–1,150). pH 6.3 (normal range 4.5–8.0). Both normal.

What that meant in my case: Kidneys concentrating urine normally. Acid-base balance is maintained. Consistent with the overall kidney picture: mildly reduced function that is stable and well-compensated.

18. Electrocardiogram (ECG)

What it is A recording of the electrical signals that coordinate the heartbeat. Electrodes placed on the chest and limbs detect the pattern of electrical activity across the heart muscle.

Why they use it: To assess cardiac rhythm, detect conduction abnormalities, and establish a baseline against which future ECGs can be compared. Changes in the electrical pattern can be an early indicator of cardiac stress or medication effects.

What it involves for the patient: Lying still on a table while small adhesive electrodes are attached to the chest, arms, and legs. The recording takes about a minute.

My result: Normal sinus rhythm with a first-degree AV block.

What that meant in my case: A first-degree AV block is a slight delay in the electrical signal passing from the upper to the lower chambers of the heart. It is common, frequently benign, and has been present on my previous ECGs. The clinical read noted “no significant change” from 2025. Stability is the objective. This was stable.

19. Iothalamate Clearance (Formal GFR Measurement)

Okay, so editorial commentary here. This is the full Kidney Function Test. It is my least favorite test of the day, and it took almost 4 hours for reasons. In fact, based on my experience with this actual one, my Physician has clearly recorded that she does not want this test used again. I’ll probably write another article on this from the perspective of a very well-informed former clinician.

Buy me a coffee?

What it is: A precise measurement of kidney filtration rate using an injected tracer substance called iothalamate. Unlike the estimated GFR calculated from a creatinine blood test — which uses a formula — the iothalamate clearance directly measures how much blood the kidneys are actually filtering per minute.

Why they use it: For transplant patients on nephrotoxic medications, knowing the actual filtration rate rather than an estimate matters for clinical decision-making. The iothalamate method is substantially more accurate than the standard creatinine-based eGFR and is used at major transplant centers when precision is needed.

What it involves for the patient: An injection of the iothalamate tracer, followed by a timed waiting period, followed by blood and urine collections that allow the clearance rate to be calculated. This is a longer process than a standard blood draw — it occupies much of the morning and requires multiple timed collection points.

My result: Corrected iothalamate clearance: 53 mL/min. Uncorrected: 59 mL/min.

What that meant in my case: A GFR of 53 is classified as chronic kidney disease stage 3A — mildly to moderately reduced function. In isolation, that sounds alarming. In context, it is my established and stable baseline. The mild reduction is a well-understood consequence of long-term calcineurin inhibitor therapy. My team monitors it closely. The objective is preservation, not recovery, and the trajectory over the past year has been flat — which, in this context, is the goal.

20. Pulmonary Function Tests (PFTs)

What it is: A comprehensive series of breathing tests measuring how much air the lungs can hold, how quickly air can be moved in and out, and how efficiently the lungs transfer oxygen into the blood.

The three key measurements:

• FVC (Forced Vital Capacity): the total volume of air exhaled in a single maximal effort

• FEV1 (Forced Expiratory Volume in one second): the volume exhaled in the first second, a primary marker of airway health

• DLCO (Diffusion Capacity): how effectively oxygen crosses from the air sacs into the bloodstream

Why they use it: PFTs are the primary surveillance tool for the transplanted lungs. A declining FEV1 is the defining marker of CLAD — Chronic Lung Allograft Dysfunction — the most significant long-term threat to graft survival. Stable or improving numbers are the objective at every visit.

What it involves for the patient: Sitting in an airtight glass-walled booth and following precise instructions from a technician: deep breaths, forceful full exhalations, held positions, repeated on demand. The test is physically demanding and requires complete, sustained cooperation to produce valid results. Multiple attempts are made at each manoeuvre to ensure accuracy.

My result:

• FVC: 3.95 litres — 112% of predicted for a healthy man of my age and size

• FEV1: 2.90 litres — 105% of predicted

• DLCO: 26.5 ml/min/mmHg — 112% of predicted

Physician interpretation: “Status post lung transplantation. Normal studies. Compared to the prior study of 10/1/2025, there is no significant interim change.”

Trend over the past two years shows FVC and FEV1 consistently above 100% of predicted — a pattern that has held without a meaningful decline since the first post-transplant assessments.

What that meant in my case: Performing above predicted normal on all three major measures, three years after receiving transplanted lungs, is not a routine result. The FEV1 at 105% means my airways are moving air faster than would be expected of a healthy 63-year-old who had never had a transplant. The DLCO at 112% means the lung tissue is highly efficient at transferring oxygen into the blood. There is no evidence of airway narrowing, no evidence of restriction, and no declining trend.

These are the numbers the team is protecting. They are currently doing well.

21. Clinical Consultation: Adam Eisenbraun, APRN, CNP, MSN

What it is: A comprehensive clinical review with the advanced nurse practitioner on the lung transplant team, conducted mid-afternoon on day one.

Why they use it: By mid-afternoon, most of the morning’s results are available. The APRN consultation allows real-time review of lab values, physical examination, medication reconciliation, and clinical interview — gathering the full picture before the imaging and exercise tests on day two and before the final physician review with Dr. Pennington. It also serves as a checkpoint, catching anything that should be addressed before the assessment concludes.

What it involves for the patient: A full clinical encounter. Medical history update, medication review, physical examination, and discussion of results, concerns, and upcoming elements of the assessment. This is not a checkbox exercise.

My result: No acute concerns identified. Labs overall unremarkable. Tacrolimus level within goal range. Physical examination satisfactory. Several results — including CMV, EBV, and Vitamin D — still pending; follow-up confirmed with Dr. Pennington. Upcoming colonoscopy discussed (due August 2026 for follow-up of a tubular adenoma identified in 2023).

What that meant in my case: A clean day-one checkpoint. The clinical note comments that I have “high health literacy” and am “a wonderful historian and self-advocate.” In post-transplant care, active engagement is not a nice-to-have. Understanding your own regimen, knowing when something has changed, and being able to communicate it clearly — these things have clinical value.

22. Annual Dermatology Consultation — Bruin Pollard, M.D.

What it is: A full skin inspection conducted by a dermatologist, examining the entire body surface for suspicious lesions, changes in existing marks, or early indicators of skin cancer.

Why they use it: Immunosuppression significantly elevates the risk of skin cancers, particularly squamous cell carcinoma. The risk begins accruing from the day immunosuppressive therapy starts and compounds over time. Annual dermatology review is a non-negotiable part of transplant surveillance.

What it involves for the patient; A head-to-toe skin examination. In my case, this has additional context: I underwent a MOHS procedure in 2015 to remove a cancerous lesion from the left temple — prior to the transplant, and before immunosuppression was a factor.

My result: The clinical note from this consultation had not populated in the patient portal at the time of writing. The appointment took place as scheduled on the afternoon of April 2.

What that meant in my case: Unable to report formally at this point. This entry will be updated when the note becomes available.

Day Two — April 3, 2026

23. Echocardiogram (Transthoracic Echocardiogram — TTE)

What it is: An ultrasound of the heart using sound waves to create real-time images of the heart’s chambers, valves, and pumping function. Non-invasive, painless, no radiation.

Why they use it: The relationship between the lungs and the right side of the heart is direct. Diseased or underperforming transplanted lungs place back-pressure on the right heart — a condition called pulmonary hypertension. If right heart pressures rise, it can signal graft complications that the breathing tests have not yet captured, or that the heart itself is beginning to struggle. The echocardiogram monitors this relationship year by year.

What it involves for the patient: Lying on the left side on an examination table while a technician applies gel to the chest and moves an ultrasound probe across the skin. The process takes thirty to forty-five minutes.

My result: Normal left ventricular chamber size. Ejection fraction 56%. Normal filling pressures. Normal right ventricular size and function. Right ventricular systolic pressure 20 mmHg. No significant valvular disease. No pericardial effusion. No significant change compared to the report of April 2025.

What that meant in my case: The heart is under no stress from the transplanted lungs. An ejection fraction of 56% is within the normal range. A right heart pressure of 20 mmHg is excellent — clinicians start to become concerned above 35 mmHg. Year-on-year stability is the most meaningful data point here, and this result was stable.

24. CT Chest (Without Contrast)

What it is: A computed tomography scan of the chest — a series of detailed cross-sectional X-ray images processed by software into three-dimensional representations of the lung tissue, airways, and chest structures.

Why they use it: The CT scan detects structural changes in the transplanted lung tissue that would not appear on a standard X-ray or PFT results. It can identify early fibrosis, air trapping, new nodules, fluid, or problems at the bronchial anastomoses — the surgical junctions where the donor airways were connected to my own. It is the most detailed structural surveillance tool available.

What it involves for the patient: Lying flat on a narrow table that moves through a large ring-shaped scanner. The scan itself takes a few minutes. No contrast dye was required for this study.

My result: Stable postoperative changes consistent with bilateral lung transplantation. No bronchial anastomotic stenosis. Stable mild scattered expiratory air trapping, predominantly in the lower lobes. No new nodules. No significant change from prior imaging.

What that meant in my case: In transplant imaging, “stable” is the word that matters most. The mild air trapping has been present on previous scans and has not changed — it is not progressing, and it is not signalling the onset of chronic rejection. The airways at the surgical junctions remain open. The structural picture is consistent with well-maintained graft function.

A note on my history: my primary post-transplant complication has been recurrent left pleural effusions — fluid accumulation around the left lung in the early post-transplant period. These have resolved, with mild residual scarring visible on imaging. This too is stable.

25. Six-Minute Walk Test (6MWT)

What it is: A functional test measuring how far a person can walk on a flat surface in six minutes, and what happens to their oxygen levels during that effort.

Why they use it: PFTs measure what the lungs can do under controlled laboratory conditions. The six-minute walk tests what they actually deliver during sustained real-world exertion. It assesses exercise tolerance, the efficiency of oxygen delivery under load, and whether oxygen levels drop during physical effort — a phenomenon called desaturation. For transplant recipients, desaturation during a walk that should be manageable can be an early warning sign that the resting test results are not telling the full story.

What it involves for the patient: Walking back and forth along a measured hallway for six minutes at your own chosen pace, while wearing a pulse oximeter on a finger. The technician tracks distance and oxygen saturation throughout.

My result:

• Distance: 1,973 feet (601 metres)

• Baseline oxygen saturation (room air): 94%

• Post-exercise saturation: 94%

• Saturation nadir during the test: 97%

• Peak heart rate: 119 beats per minute (70% of age-adjusted predicted maximum)

• Rated effort at end: 3 out of 10 for both breathlessness and leg fatigue

Physician interpretation: “A satisfactory submaximal six-minute walk study for the indication of post-lung transplant. There is no significant change since prior study of 4/3/2025.”

What that meant in my case: 601 metres is a strong result for a 63-year-old post-transplant patient. More significant than the distance is the oxygen picture.

My resting baseline was 94%. During the walk, my saturation nadir — the lowest point recorded throughout the test — was 97%. In plain English: as my breathing deepened with effort, the transplanted lungs responded and performed more efficiently than at rest, not less. There was no desaturation event of any kind. At a peak heart rate of 119 bpm, representing 70% of my predicted maximum, the lungs were delivering exactly what was asked of them.

A rated effort of 3 out of 10 at the end of six full minutes of walking is also worth noting. I was not at my limit.

26. Clinical Consultation: Kelly Pennington, M.D.

What it is: The closing consultation with the transplant pulmonologist, synthesising all results from both days into a clinical assessment and an updated plan of care.

Why they use it: Individual test results, however reassuring, only carry their full meaning in combination. This is where the clinician integrates the labs, the imaging, the functional tests, the physical examination findings, and the patient’s reported history into a clinical judgment. It is the point at which all the separate questions asked by all the separate tests receive a single considered answer.

What it involves for the patient: A full physician consultation: results review, examination, discussion of any changes to the management plan, and forward planning.

My result: — summary of Dr. Pennington’s assessment:

All results reviewed. Graft is performing well.

On lung function:

FEV1 remains at or near its historical peak. The highest recorded FEV1 to date was 3.01 liters in October 2024. The current 2.90 liters remains at a consistently high plateau. The physician’s interpretation from Dr. Lim (who read the PFTs) used the phrase “normal studies.”

Dr. Pennington’s overall word for the assessment was exemplary.

No ongoing indication for surveillance bronchoscopy. Previous biopsy results have all been A0B0C0 or equivalent — no evidence of rejection on any biopsy taken since the transplant.

Immunosuppression regimen unchanged: tacrolimus 2mg morning / 1.5mg evening, mycophenolate 1000mg twice daily, prednisone 5mg daily.

No new concerns identified. Return in six months. Monthly labs to continue.

Conditions being actively monitored and managed:

• CKD stage 3A: stable at baseline creatinine 1.4–1.5

• Steroid-induced diabetes mellitus: HbA1c 5.5%, well-managed on weekly Mounjaro

• Left pleural effusion history: resolved, mild residual scarring on imaging

• Hypertension: 145/85 on carvedilol; being monitored; no medication change at this time

• Hypomagnesemia: improved with magnesium oxide supplementation

• GERD, OSA, BPH: managed, stable

• Bone health: low density on DEXA (April 2025); repeat recommended 2027; lifestyle measures preferred at this stage

• Cancer screening: colonoscopy due August 2026 (follow-up on tubular adenoma from 2023); annual dermatology complete; PSA normal

What that meant in my case: A comprehensive review across every dimension of post-transplant health, producing a clear and consistent picture. The plan is a continuation — which, after an assessment of this complexity, with results of this quality, is its own form of good news.

DONATE LIFE

The Synthesis

The annual assessment is not a single number or a single test. It is a question asked twenty-six different ways, with twenty-six different instruments, looking for twenty-six different failure modes.

This year, none of them found one.

The transplanted lungs are functioning above predicted normal. The heart is under no stress. The immune surveillance is working. The drug levels are correct. The kidneys are stable. The metabolic picture is well-managed. The structural imaging is unchanged.

Three years post double-lung transplant, at sixty-three years old, with pulmonary fibrosis as the underlying diagnosis that brought me here: the word the team used was exemplary.

I said it was earned.

Both things are true.

SM

This installment is a little different from the previous three: this story is still unfolding. That makes the structural lessons more urgent, not less.

Executive Summary

Delve is a San Francisco-based, Y Combinator-backed compliance automation startup that raised a $32 million Series A at a reported $300 million valuation in July 2025, led by Insight Partners. The company positioned itself as an “agentic layer” for compliance, promising to compress the readiness timelines for SOC 2, HIPAA, ISO 27001, and GDPR through AI-driven workflows and deep integrations.

On March 19, 2026, an anonymous Substack investigation published under the name “DeepDelver” alleged that Delve’s speed was achieved not through clean automation but through manufacturing audit artifacts, generating auditor-facing conclusions before independent review, and routing work through a narrow set of audit firms described as “rubber stamps.”

Delve issued a public rebuttal on March 20, calling the Substack post “misleading” and asserting that final reports are issued solely by independent, licensed auditors, and that templates are standard starting points that customers are responsible for adapting.

TechCrunch reported on the allegations on March 21, 2026, noting both the DeepDelver claims and Delve’s response, along with subsequent reports of possible security vulnerabilities in Delve’s platform and claims of leaked sensitive materials, including background checks and equity vesting schedules.

For healthcare leaders and regulated-industry buyers, the Delve episode matters less as a story about one vendor and more as a structural warning. “Compliance at speed” is attractive precisely because compliance functions as a procurement gate. But assurance frameworks derive their value from independence, operating effectiveness, and evidence integrity.

That tension is sharpening as U.S. regulators push for more prescriptive cybersecurity discipline under HIPAA and emphasize that AI does not exempt firms from truth-in-marketing and fraud standards.

What Is Known vs. What Is Contested

Before analysing the structural implications, it is essential to separate documented facts from contested allegations. This discipline is what distinguishes strategic analysis from outrage, and it matters especially when a story is still developing.

Read more

Before we get to the dinosaurs, we need to talk about the dwarves.

Somewhere in the deep lore of Tolkien’s Middle-earth, there is a cautionary tale that the enterprise software industry has spent the last eighteen months studiously ignoring.

The Dwarves of Moria, the greatest miners and craftsmen of their age, discovered mithril, a material of extraordinary value, light as silk and harder than steel. And because they were very good at what they did, and because the market for mithril was essentially infinite, they kept digging.

Deeper and deeper, faster and faster, with increasingly sophisticated techniques. They broke through geological boundaries that existed for a reason. And in doing so, they woke something up. Things got ugly fast!

Tolkien’s narrator delivers the diagnosis with characteristic understatement:

Not incompetently. Not with any bad intentions.

Just too greedily and too deep.

Their sin was not a lack of capability. It was the absence of restraint in the presence of capability. The dwarves were, by any reasonable measure, operating at peak performance right up until the moment they weren’t.

I’ve been thinking a lot lately about the story of The Mines of Moria, because I feel I am watching that same dynamic play out in real time across the enterprise software development landscape.

And the mithril in this case is Agentic AI coding.

The Coding Swarm Arrives

If you’ve been paying attention to the software development tooling space over the last year, you’ve watched an actual, genuine paradigm shift unfold.

We’ve moved from a world of singular LLM code-writing capability, which was impressive but still essentially reactive. Now, we have a world of multi-agent orchestration: specialised agents coordinating autonomously to complete complex development workflows.

• CrewAI lets developers assign roles, backstories, and goals to autonomous agents, simulating an entire workforce.

• Microsoft’s AutoGen enables multiple agents to “converse” with each other to solve tasks with minimal human intervention.

• LangGraph builds cyclical agentic loops that run, statefully, until a feature is considered complete.

That’s just some of the big noisy names; there are plenty of other options and tools in this ‘Wild West’ Goldrush. The technical infrastructure to achieve 10x production isn’t theoretical. It’s already arrived!

But, ‘Here be dragons’ (Oh…Hi Smaug!)

Here’s what that productivity capability actually produces in practice. A study published last month by DX, the developer intelligence platform, analysed data from 40 companies between November 2024 and February 2026.

Despite a 65% increase in AI tool usage across those organisations, actual PR (Pull Request) throughput, the measure of finished, reviewable, shippable work, improved by only 9.97%.

Not 10x. Not even 1x. Roughly one-tenth of one x.

The study’s conclusion should be tattooed on the forehead of every CTO currently crowing to their board about their agentic deployment:

The bottleneck is, and always has been, the human-centric parts of the product lifecycle.

• Planning.

• Alignment.

• Review.

• Absorption.

• Retirement

These are the parts that don’t speed up just because you’ve pointed an agentic coding swarm at the codebase.

Which brings us, inevitably, to…The Dinosaurs.

‘Thintelligence’

In Spielberg’s Jurassic Park, there’s a scene that has been quoted so frequently that it’s almost lost its diagnostic power. Over lunch, Ian Malcolm (played brilliantly by Jeff Goldblum) delivers what amounts to a one-sentence strategic assessment of the entire park’s operation:

Most people love this quote and treat it as a pithy anti-technology soundbite.

It isn’t. You need to look deeper.

Crichton’s novel actually gives Malcolm two separate monologues that, taken together, form the most precise diagnosis I’ve encountered of what is currently happening in enterprise software development (and in the Agentic AI space in general).

In the first, Malcolm is talking about the park’s engineers. He doesn’t question their competence. He questions their field of vision.

They have what he calls “thintelligence,” the ability to see the immediate situation with clarity while remaining completely blind to the surroundings. They think narrowly and call it being focused. But they don’t see the consequences.

That, Malcolm says, is how you get an ‘island like this’.

Later in the novel, as he deteriorates from his injuries, Malcolm delivers a longer, angrier argument about what he calls inherited wealth.

Most kinds of power, he argues, require a substantial sacrifice. An apprenticeship. A discipline lasting many years. And by the time someone has acquired the ability to act with that kind of force, the process of acquiring it has changed them enough that they won’t use it recklessly. The discipline of getting the power is what prevents you from abusing it.

Scientific (technological) power, Malcolm argues, doesn’t work that way.

It’s inherited wealth. You read what others have done. You take the next step. You can do it very young. You can move very fast. There is no long apprenticeship. There is no mastery required.

And then Crichton, writing 36 years ago, lands the line that should be pinned to the wall of every enterprise that purchased an agentic framework this year.

The buyer of that power, Malcolm says, will have even less discipline than the person who created it. The buyer simply purchases the power like any commodity, and doesn’t even conceive that any discipline might be necessary.

That is not a metaphor. That is a societal warning flare. Michael Crichton called it 36 years ago.

Swap “scientific power” for “agentic AI capability”, and you have a precise description of the current market dynamic.

The new paradigm frameworks are extraordinary. The teams building them are, in many cases, genuinely brilliant. And the organisations purchasing those frameworks are deploying them at speed, without the institutional discipline to manage the output, because they don’t even conceive that such discipline might be necessary.

WORKSLOP!

Researchers from Stanford University and BetterUp Labs, writing in the Harvard Business Review, recently identified a phenomenon they call “workslop,” AI-generated output that masquerades as good work but lacks substance.

Their longitudinal study of 4,000 workers found that 41% of U.S. desk workers had encountered workslop in the preceding month.

Each instance costs the recipient an average of one hour and 56 minutes in what the researchers term a “rework tax,” the time spent diagnosing, interpreting, correcting, or simply discarding AI output that arrived looking like finished work but wasn’t.

One hour and 56 minutes. Per instance.

In an environment where the production rate of AI-generated output is accelerating every quarter.

Learn more about Workslop from in the following article:

Decision Intelligence

Have You Been Workslopped Yet?

(Here for AI news? Scroll to the very bottom for 7 recent AI headlines you should know about.)This is a reader-supported publication. To encourage my writing, consider becoming a paid subscriber…

Read more

7 months ago · 34 likes · Cassie Kozyrkov

That is the arithmetic that should terrify product leaders.

You have NOT 10x’d productivity. You have 10x’d code production, which is fundamentally different. But in doing so, you have loaded a rework and disruption tax onto every downstream function that has to absorb, evaluate, and operationalise that production.

• The marketing team receives 100 AI-generated landing pages.

• The engineering team ships ten new autonomous workflows per quarter.

• And the compliance function, the support team, the documentation writers, the end users (or clients), all operating at their very human 1x absorption rate, are drowned in the delta overload.

Let’s call it an Absorption Limit. Every organisation has a finite capacity to intake, validate, operationalise, and sustain new capabilities in a given period. That capacity has a ceiling. And that ceiling exists to protect the organisation's ability to meet its core mission.

It is bounded by real, human constraints: change management bandwidth, compliance review cycles, user training infrastructure, support team ramp-up, and the simple cognitive load of the people who have to actually use the damned thing (feature, agent, app) that you just shipped.

Yes, Agentic AI has blown the software production ceiling out of the water. BUT…it has done absolutely nothing to raise the organizational absorption floor.

The Balrog is getting grumpy!

I’d love it if you’d consider
buying me a coffee… The Balrog knocked mine over!

Buy me a coffee

Earned Features vs. Unearned Features

This is where I want to plant a flag.

I think there’s a framework the industry needs but doesn’t yet have the language for.

An Earned feature is one that has passed through the full discipline of product management before it was built.

The problem was identified. The user’s need was validated. The solution was constrained to what the organisation can actually support. The deployment was scoped to what the user base can actually absorb. Earned features are slower to ship and harder to celebrate in all-hands meetings, and they are the only features that compound into durable value.

Meanwhile, an Unearned feature is one that exists because the technical capability to build it arrived before the organisational wisdom to evaluate, plan, launch, and adopt it.

If you want to know what an unearned feature looks like when it meets the legal system, consider Moffatt v. Air Canada, decided by the British Columbia Civil Resolution Tribunal in 2024.

Air Canada deployed an autonomous chatbot that, operating without adequate constraint, generated a bereavement fare policy that did not exist. When a customer relied on that fabricated policy and was subsequently denied the fare, Air Canada’s defence was remarkable: the airline argued that its chatbot was a “separate legal entity” and that the company therefore bore no liability for its output.

The Tribunal disagreed. Forcefully!

The ruling established that an enterprise is responsible for all commitments made by its agents, whether those agents are human or automated. The chatbot was not a separate entity. It was an Unearned feature, deployed because the capability existed, without the organisational infrastructure to manage, constrain, or even monitor what it was saying to customers.

Air Canada didn’t have a technology problem. It had an absorption problem. It shipped an autonomous agent without building the compliance, review, and governance muscle to sustain it.

The reckoning between what it could produce and what it could manage arrived not as a Balrog, but as a Tribunal ruling and a very public embarrassment.

This distinction between earned and unearned matters because the current discourse treats all agentic deployment as innovation. But they aren’t. Some of it is innovation.

Much of it is simply production without product, and the fact that it was produced quickly doesn’t make it more valuable. It makes it more dangerous because it was deployed before anyone developed the institutional reflexes to manage it.

The Henry Ford Defence (And Its Limits)

Let’s pause for a moment because we have a third guest at the bar and I want to be fair to the counterargument, because it is a real one.

The counterargument is Henry Ford.

Nobody asked for a car. If Ford had listened to his customers, he’d have bred faster, stronger horses. True paradigm shifts, by definition, outpace the market’s ability to articulate demand. You build the future, and the market catches up. This is a defensible position, and when it’s right, it’s the difference between Apple and BlackBerry.

But here is where the analogy breaks. Ford didn’t ship the Model T and then immediately ship 100 variants, each with a different steering mechanism, to see which the market preferred.

He built one car. He constrained it ruthlessly. He earned his market’s trust through radical simplicity before he earned the right to complexity.

The phrase “You just don’t understand the paradigm” is doing an enormous amount of heavy lifting in boardrooms and PE firms right now. And much of what it’s lifting is unvalidated, unconstrained, unearned output that has been dressed up as vision.

The Henry Ford defence is legitimate when it describes genuine paradigm-shifting matched with restraint. It’s bullshit when it’s used as a shield against the hard, slow, unsexy work of product validation.

The Balrog in the Backlog

So…let me bring the metaphors home.

The Dwarves of Moria did not fail because they lacked skill. They failed because their extraction capability outpaced their risk management capability, and the gap between those two things contained a Balrog.

InGen did not fail because genetic engineering was inherently wrong. They failed because their production capability outpaced their containment capability, and the gap was filled by Velociraptors.

Enterprise software organisations currently deploying agentic AI at 10x production velocity are not, according to the best available evidence, achieving 10x outcomes.

They are achieving roughly a 10% improvement in finished output, per the DX data, while imposing a cumulative rework tax on every downstream human function. The gap between what they can produce and what their organisations can absorb and sustain is widening every quarter.

Something lives in that gap.

So what will be the signal that throttles this mad dash?

It might be Regulatory exposure, as Air Canada discovered.

It might be the slow, grinding churn of customers who quietly leave because your product became bloated and unusable while you were busy celebrating your throughput.

It might be the catastrophic system failure that arrives when autonomous agents, operating without adequate human oversight, interact in ways nobody anticipated because nobody had the bandwidth to anticipate (far less test) them.

Traditional product management, the slow, unglamorous discipline of asking “what problem are we actually solving?” before anyone asks “how many agents do we need for this swarm?”, is not a bottleneck.

It is the only thing standing between your organisation and whatever is waiting in the dark and the deep.

Michael Crichton called it “Thintelligence”.

JRR Tolkien called it “delving too greedily and too deep”.

The DX data calls it a 65% increase in tool usage for a 10% improvement in output.

Me? I’m calling it the Absorption Limit, and I think it’s the most underrated risk in enterprise software right now.

The dwarves had their word for what they found down there. What’s it called in your organization?

What’s the most “unearned” feature you’ve seen shipped recently? Tell me in the comments. Comment by clicking below.

Leave a comment

Stuart Miller is Managing Director of Haverin Consulting, a healthcare IT strategy consultancy. He has spent 20+ years watching organisations confuse production with progress, and he remains stubbornly optimistic that product discipline will eventually be fashionable again. #IamPragmatic

References:

Crichton, M. (1990). Jurassic Park. Ballantine Books.

Grewal, D., Satornino, C.B., Davenport, T.H., et al. (2025/2026). “AI-Generated ‘Workslop’ Is Destroying Productivity.” Harvard Business Review.

Moffatt v. Air Canada, 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal). Available at: https://canlii.ca/t/k2v9z

Reock, J. (2026, March 11). “AI productivity gains are 10%, not 10x.” DX (Developer Intelligence Platform). https://getdx.com/blog/ai-productivity-gains-are-10-percent-not-10x/

Tolkien, J.R.R. (1954). The Lord of the Rings: The Fellowship of the Ring. George Allen & Unwin.

The story of how and why I created the above video from a single image and an idea continues after this brief introduction.

AI hasn’t replaced the creative process, but the tools have compressed the distance between “idea” and “visualization.” Just as the shift from pen and ink to digital drawing and animation accelerated and democratized execution, the AI technologies are doing the same.

It allowed me to be the screenwriter, director, and editor of a story that, just a year ago, to realize would have required a whole team of animators, and a budget to match.

This article isn’t really a tutorial on using Google Flow. There are plenty of great YouTube videos and examples out there. The key takeaway here is that you can learn to use these tools with you acting as the screenwriter, Director and editor.

But first! Don’t over-index on artistic perfection; this is not a Hollywood movie.

That’s not Stephen Spielberg calling.

As a storytelling and an inspiration tool, these tools are incredibly easy to use, but difficult to tame.

Like everything with AI, it’s evolving incredibly rapidly, so what’s true at the time of writing could change in the next week.

I deliberately chose to stay away from photorealistic person storytelling, as that’s where things get very ‘uncanny valley-y’ and weird, but used it like this with an animation/comic book style…Wow!

However, if you WOULD like to see how I went through this process, just leave a comment below, and I might consider creating a video on how I made this.

Leave a comment

In the world of Product Management, we often talk about NIHITO (Nothing Important Happens Inside The Office). It’s a core pillar of the Pragmatic Framework,

But dryly explaining the “what and why” is one thing—stimulating “wow” and visualizing the “how “ is a whole other.

For the Alignment Files: Part 3 — Intellectual Hitchhiking article, I had already expressed a little creative vision by creating my Product Manager noir detective character.

But I wanted to take my readers beyond the static dullness of a dry written explanation.

I wanted them to follow our PM Detective’s ‘Hero’s Journey’ as he navigates the fog of synthetic data to find the “Ground Truth.”

Here is how I turned a single image, which I was already delighted with (thanks, Nano Banana), into a cinematic journey using Google Flow, Adobe Premiere, and a bit of creative experimental grit.

Disclaimer. I have absolutely minimal to zero design or artistic capabilities, but I do know how to verbally visualize and tell a great story. :-)

The Challenge: Maintaining Aesthetic Continuity

One of the biggest hurdles in AI content creation is “style drift.”

I had a specific seinen manga style for the illustrations in the five articles. In part 3, the original image shows our Product Manager hero as a noir-style detective seeking the “Ground Truth,” surrounded by Synthetic personas.

That fits the lessons in the article about the risks and dangers of overly relying on AI-generated synthetic user personas, instead of getting out of the office, away from the computer, and actually talking to users and potential users.

In this video story, I wanted our Product Manager detective to look the same, feel the same, and inhabit the same world in every frame and flow from scene to scene.

The Workflow: From Brainstorm to Premiere

I’ve always believed that the best tools are the ones that get out of your way. This was my first time using Google Flow, and after a bit of bumbling around and a few key questions to Gemini, the learning curve was surprisingly intuitive.

The Anchor Image: I started with the high-fidelity reference image that established the lighting, character design (our fedora-clad detective), and the “Synthetic Persona” visual language.

Prompt Engineering in Flow: Using the anchor image as a stylistic North Star, I crafted a series of custom prompts for each scene.

The first prompt is super important for describing the story’s overall style and approach, so it should be more structured and detailed.

The First Prompt

Visual Style: 16:9 aspect ratio, high-contrast Japanese Seinen Manga aesthetic. Deep blacks, heavy purposeful ink lines, and professional screentone textures for atmospheric shading.

Scene & Character: A crowded, dimly lit corridor outside “NIHITO Clinic”. The protagonist—a plucky Japanese male product manager with a rugged, action-hero face and short, neat, dark hair—is wearing a noir trench coat and fedora.

Action: He holds a “ghostly” magnifying glass up to his eye, scanning the somber crowd. As the lens passes over seemingly normal people, it reveals “Synthetic Personas”—shimmering, glitching wireframe apparitions that flicker in and out of existence. The protagonist’s expression shifts from intense, detective-like focus to a sudden, sharp look of shock as he realizes how many in the corridor are artificial.

Camera Movement: A slow, dramatic dolly-in toward the protagonist’s face, ending in a tight close-up of his widening eye reflected in the magnifying glass.

I used the video clip generated from this prompt as my anchor scene 1 in Google Flow. Everything from there on was pure storytelling using the ‘Text to Flow’ feature.

You MUST let each scene generate before then repeating the process to proceed with the scene sequence in your story. I simply used my imagination to describe each scene, then generated the next one after the other.

Capturing the “Ground Truth”:

In my case, the storyboard follows a clear arc:

The detective realizes his current data is synthetic, physically breaks through those digital barriers, and finally enters the NIHITO Clinic to sit face-to-face with real customers.

I described the action—the detective peering through his “Ground Truth Lens,” the physical fight with the digital personas, the warm interior of getting to real people inside the clinic, then his slow, satisfied walk back to the office, job well done—while ensuring Flow referenced the original style and kept a consistent…no pun intended…flow!

Buy me a coffee

Post-Production in Adobe Premiere:

Now, let’s be clear: what gets generated is great, but it’s not push-the-button, release-ready.

• The generated clips had some weird overlaps that needed trimming.

• Once I had generated them, I found I didn’t like the sequence, and I wanted to change it up to adjust the story slightly.

• The audio it generated was interesting and fun, but wasn’t a script… lesson for next time: give it a clear script in the prompts.

• Also, the voice dialog didn’t come from the same characters consistently. That kind of fits okay with my aesthetic of English-dubbed Japanese manga, but probably wouldn’t work with other styles as well.

• So, being a good YouTube editor with basic skills, once the clips were generated, I moved to Premiere to stitch the narrative together.

• I added the following:

  Text Overlays: Black lower bar with comic-style fonts to drive the “Manga comic” rather than a more anime storytelling feel.

  Music Track: Because I didn’t want to do a full narrated overdub with foley sounds at this time, I decided to use the comic book captions and instead found a good royalty-free music track with an Asian feel to provide that driving, investigative energy soundtrack.

Lessons Learned: The High-Velocity Creative Workflow

Moving from a concept to a finished narrative video in a single afternoon taught me several things about the future of professional communications.

For product leaders and business consultants, the “Lesson” isn’t just about the AI video technology—it’s about the shift in how we tell stories.

1. The “Anchor” is Your Insurance Policy

The biggest risk in AI content is “hallucination” or stylistic drift. By starting with a high-fidelity anchor image, I wasn’t just giving the AI a suggestion; I was giving it a fixing it in place with a constraint.

• The Takeaway: In any AI project, define your “Non-Negotiables” (style, character, tone—and script is applicable) first. It’s much faster to refine a consistent style than to try to fix a disjointed one later in the edit.

2. Constraints Breed Speed

I didn’t try to make a 10-minute ‘Oscar-winning’ short movie. I focused on a 35-second “Visual Metaphor.” By limiting the scope to five key storyboard beats, I avoided the “infinite refinement” trap. I’m a business professional, not a Tisch School of Film MFA grad.

• The Takeaway: Use AI to solve a specific communication problem, not to replace an entire film studio. A compelling 30-second high-impact video is often more effective for a professional audience than a long-form presentation.

3. Human “Direction” is the New “AI Creation.”

Even with tools as powerful as Google Flow, the AI didn’t know the meaning of the “Ground Truth Lens” or why the detective needed to tackle the synthetic persona. The narrative intent came from years of product/market experience; the AI simply provided the “labor” to visualize it.

• The Takeaway: Your value as a leader isn’t in your ability to draw or edit—it’s in your ability to direct the tools to deliver a deep, proprietary insight and tell a compelling story. In my case, leave them wanting more.

4. The “Good Enough” vs. “Perfect” Threshold

Was every frame of the manga perfect? Perhaps not to a professional animator. But for an article explaining a product management framework principle, it was extraordinary.

• The Takeaway: AI allows you to hit the “80% quality” mark at 5% of the traditional cost and time.
  For business communication, that 80%—when paired with a strong message—is your competitive advantage.

5. Post-Production is the “Glue.”

Raw AI video can feel disjointed. The real “magic” happened in Adobe Premiere when I edited, added the music, and the manga-style captions. That “human touch” at the end of the process is what turned a collection of clips into a cohesive brand asset.

• The Takeaway: Never post “raw” AI output. It’s fun, it might be kooky, or it might just be incredibly dull. The final 10% of effort—editing, branding, and pacing—is what provides the professional polish that will hold your audience’s attention.

Why This Matters for Product Leaders and Consultants

As a product coach, I’m constantly looking for ways to communicate complex ideas, models, and frameworks (like the Herringbone or Alignment Files) in ways that stick.

Use these visualization tools to reach people with your story in a way that words (spoken or written) just won’t do.

SM

This is Part 5 and the finale of The Alignment Files, a series that continues from The Delivery Delusion.

In Part 1, we named the three tribes. In Part 2, we exposed the Factory’s gravity well. In Part 3 , we confronted Epistemic Freeloading. In Part 4, we introduced the Validity Buffer to domesticate the HIPPO.

Now, we bring it home.

The Trap of “Picking a Winner”

Throughout this series, we have diagnosed what goes wrong when one tribe dominates:

• When the Factory wins, you get the Build Trap — performance theater, Zombie Products, and a roadmap that feeds the machine instead of the market (Part 2)

• When Market Grounding surrenders to the Factory’s demand for faster, you get Epistemic Freeloading — synthetic evidence masquerading as truth (Part 3)

• When the HIPPO overrides both tribes, you get unvalidated initiatives consuming factory capacity while evidence-backed work is perpetually deferred (Part 4)

The instinct, at this point, is to pick the right tribe and give it control. Organizations do this all the time:

This is an “Either/Or/Or” trap. And it is the final, deepest failure mode this series needs to address.

The three-way tension between Strategic Leadership, Market Grounding, and Scaled Technical Execution is not a war to be won. It is not a problem to be solved.

It is a set of Polarities to be Managed.

What Is a Polarity?

The concept comes from Barry Johnson’s Polarity Management framework, and it is deceptively simple.

A polarity is a pair of interdependent (but often competing) values that need each other over time for a system to be effective. They are not “problems” with a right answer.

They are ongoing tensions that require “Both/And” thinking.

Breathing is the simplest example of polarity management you can imagine.

Inhaling is good. Exhaling is good.

But if you decide that inhaling is “the answer” and commit to it exclusively…you die.

The value is not in either pole. It is in the rhythm and balance between them. In practice, healthcare (and indeed life) is chock full of polarities.

One well-known healthcare polarity is the tension between the Mission of why a hospital or health system exists, and the Money of running a healthcare system.

Product management in a health system works the same way. Strategy is essential. Market evidence is essential. Execution is essential.

The value is not in choosing which one matters most. It is in the organizational rhythm that holds all three in dynamic balance.

Over-focus on any one pole, and you experience the downsides of all three.

Why Pairs? Why Not a Three-Way Map?

A natural instinct at this point is to attempt to map all three tribes onto a single model and manage the tension across all axes at once. While that sounds elegant, it is also unworkable.

Polarities are managed in pairs. Always. This is not a simplification; it is the fundamental architecture of the Polarity Management methodology.

Each polarity is a specific energy system: two interdependent poles in constant, predictable tension. Think of the infinity loop in the cartoon diagram. Energy flows from one pole to the other and back again, endlessly.

Over-focus on one side, and the system naturally generates pressure to swing toward the other. The skill is not in stopping the swing. It is in managing the rhythm so the system captures the upsides of both poles while minimizing the downsides of each.

Negotiation is the weapon of choice here.

You cannot do this across three axes simultaneously. The tensions are too different, the failure modes too distinct, the early warning signals too specific.

Instead, you stack them. Each pairing at the tribal borders is its own polarity, with its own infinity loop, its own upsides and downsides, its own rhythm.

If you can successfully manage each pair discretely, the aggregate effect is organizational alignment across all three tribes.

No single polarity achieves the greater purpose alone. The rhythm across all three sets of polarities is what gets you there.​​​​​​​​​

The Three Core Polarity Pairs

When we map the trilateral conflict from Part 1 onto Johnson’s Polarity Management framework, three specific polarities emerge.

Each operates at a border between two of the tribes. If one dominates the other, then strains and fault lines arise.

Each of these pairs has a healthy state and a failure state. And each has early warning signals that tell you when the organization has tipped too far.

1. Strategic Vision vs. Execution Throughput

(The border between Kellogg and SAFe — the 3–5 year horizon meets the 3–6 month cycle)

The tension: Long-term business architecture and capital allocation vs. immediate delivery commitments and burn-down charts.

Over-focus on Vision: Visionary Paralysis. The organization strategizes beautifully and endlessly. Decks multiply. Visionary roadmaps are pristine. Nothing ships. The Factory starves while leadership debates and shifts the priorities for the five-year horizon for the third consecutive quarter.

Over-focus on Throughput: Soul-Crushing Agile. Velocity becomes a hollow metric. Sprints are full. PI objectives are met. But nobody can explain how any of it connects to where the organization needs to be in three years. The North Star is in the executive presentation and stays there. The agile work is somewhere else entirely.

The healthy rhythm: Every sprint is a measurable step toward the strategic horizon. The Factory does not just execute; it executes in a direction. Strategy does not just envision; it is expressed in deliverable increments that the Factory can plan, focus on, and act on.

Early warning sign: When was the last time anyone in a sprint review referenced the three-year strategic plan? If the answer is “never,” you have tipped into throughput.

2. Market Agility vs. Process Predictability

(The border between Pragmatic and SAFe — the 6–18 month horizon meets the PI cadence)

The tension: Responding to real-time market truth (NIHITO) vs. the stability of a committed Program Increment plan.

Over-focus on Agility: Constant pivoting. The roadmap keeps shifting whenever a new interview round surfaces a new complaint. No feature achieves depth. The Factory never finishes anything because the target keeps moving. Discovery becomes an excuse for indecision.

Over-focus on Predictability: The Build Trap, in its purest form. PI plans are locked. Commitments are sacred. The market moved six weeks ago, and the Factory is still building yesterday’s features because they were “committed” in the PI Planning ceremony. Evidence arrives and has nowhere to go.

The healthy rhythm: PI Planning includes dedicated, protected capacity for discovery sprints. New market evidence has a defined entry point into the cadence — not disruptive, not ignored, but structurally accommodated. The Factory can pivot within the increment, not just between them.

Early warning: Is 100% of engineering capacity locked in PI objectives with zero room for investigation? If yes, the organization has formally decided to stop learning. Most have. Few have admitted it.

3. Strategic Vision vs. Market Evidence

(The border between Kellogg and Pragmatic — the 3–5 year transformative horizon meets the 6–18 month evidence base)

The tension: Leading the market toward a future the user hasn’t articulated yet vs. anchoring every investment in validated, pervasive problems that real people have described in real interviews.

This is the most intellectually interesting and challenging polarity — and the one organizations are typically worst at managing.

The strategist says,

The market investigator says:

They are both correct. That is what makes it a polarity.

Over-focus on Vision without Evidence: Elegant Irrelevance — the phrase we used in Part 1. The strategic deck is beautiful. The five-year vision is compelling. The CEO is energized. And none of it maps to a problem that real clinicians experience as urgent or pervasive.

This is also where the HIPPO lives. The HIPPO is almost always a Strategy tribe actor — imposing vision without evidence, powered by pattern recognition and authority rather than ground truth. The Validity Buffer from Part 4 sits directly on this border.

Over-focus on Evidence without Vision: Incrementalism. The product team has exquisite awareness of every friction point in the current workflow. Every complaint is catalogued, validated, and saturation-tested.

And the roadmap becomes a sophisticated maintenance operation — solving today’s problems with precision, but with no line of sight to a fundamentally different model of care. You become the best-informed team in the building, and you never transform anything.

The healthy rhythm: Strategy proposes the direction. Market validates whether the direction maps to real, pervasive problems. Vision enters the system — but it earns its evidence before consuming factory resources.

The Validity Buffer is the mechanism that makes this rhythm operational: strategic initiatives enter at a low Truth Score, and discovery work either validates them upward or exposes them as elegant guesses. Meanwhile, market evidence is not just catalogued — it is interpreted through a strategic lens that asks: “Does this friction point tell us something about where care delivery is heading, or is it just a complaint about today?”

Early warning: Are strategic initiatives entering the roadmap without passing through any market validation gate? You have tipped into vision. Conversely, has the roadmap become entirely reactive, driven by current complaints with no connection to a transformative outcome? You have tipped into incrementalism.

The sharpest diagnostic question: Can anyone on the product team articulate how the current quarter’s validated market problems connect to the three-year strategic vision? If the answer is a blank stare, the two tribes are not talking to each other.

The Role Nobody Has Named

If these polarities are real (and if you have been nodding along through this series, you know they are), then someone in the organization needs to own the rhythm.

Not just Strategic Leadership. Not just Market Grounding. Not just Scaled Technical Execution.

The rhythm between the three.

That is not typically the Product Manager’s job as typically defined. The Product Manager, as we discussed in Part 1, is caught in the crossfire — serving three masters with no structural mechanism to reconcile them, and no authority to enforce balance.

It is not the Chief Technology Officer’s job. The CTO owns the Factory and its performance.

It is not the Chief Strategy Officer’s job. The CSO owns the North Star.

And it’s not the Chief Digital Officer’s job. The CDO owns the Market and its users' and buyers' needs.

The role that is missing is the Polarity Navigator — the person (or function) whose explicit purpose is to detect when the organization has over-rotated toward one pole and to initiate the correction before the failure mode sets in.

In practice, this is instinctive for the very best executive leaders. They sit at the intersection of vertical strategy and horizontal throughput, quietly resolving the contradictions that arise at tribal boundaries. Negotiating the fragile peace of organizational politics and culture.

But “instinctively” and “quietly” are not organizational disciplines. They are individual heroics. And individual heroics do not scale.

Organizations that embrace polarity management will formalize it. Not as another process layer — God knows we don’t need more process — but as an explicit accountability with early warning metrics, intervention authority, and a seat at the table where roadmap decisions are made.

Bringing It Home

This series began with a simple observation: the Delivery Delusion persists not because organizations lack talent or technology, but because they have never resolved a fundamental philosophical conflict about what product management is and what it is there to do.

Five articles later, here is where we landed:

The three-way war between Strategic Leadership, Market Grounding, and Scaled Technical Execution is real, structural, and present in every large health system IT organization. It is not going away quietly. (Part 1)

The Factory’s gravity well is the most common failure mode — not because the Factory is malicious, but because it is loud, visible, and self-reinforcing. (Part 2)

Epistemic Freeloading is the newest and most seductive threat — the promise that AI can replace the hard, human work of market discovery. It cannot. (Part 3)

The Validity Buffer is a practical mechanism for making evidence quality transparent and protecting the roadmap from the HIPPO effect. (Part 4)

And Polarity Management is the discipline that holds it all together — the organizational change regulator that prevents any single tribe from dominating at the expense of the others. (This article)

The Factory will always be busy and loud. Strategy will always be urgent and imperative. The Market will always be grumpy and inconvenient.

Nobody’s baby is ugly. But the babies have to learn to share the toys in the nursery.

The organizations that thrive will not be the ones that pick the right tribe. They will be the ones who master the rhythm between all three — breathing in and breathing out, repeatedly, deliberately, with the discipline to notice when they have been holding their breath too long.

Choose to breathe!

Thank you for joining me for The Alignment Files. If you are ready to move beyond Feature Factories, Zombie Products, and Synthetic Stagnation, I would love to help you build your own alignment architecture.

If you would like to explore what we can do together, please use my Calendly link below to book a short discovery call.

You can find my availability by clicking this link to book some time.

Book 15 minutes

Stuart Miller is Managing Director of Haverin Consulting, a healthcare IT strategy consultancy. Subscribe to Haverin About… for more deep dives into the intersection of AI, Product Management, and the human experience of healthcare.

This is Part 4 of The Alignment Files, a series that continues from The Delivery Delusion. In Part 1, we named the three tribes. In Part 2, we examined the Factory’s gravity well. In Part 3, we exposed Epistemic Freeloading — the dangerous substitution of synthetic validation for real market truth.

In this part, we introduce a practical mechanism for making evidence quality visible on your roadmap — and for surviving the most politically dangerous animal in any health system: the HIPPO.

The Most Dangerous Animal in the Room

Every organization has them. Often Several!

The HIPPO — the HIghest Paid Person’s Opinion.

The HIPPO is not necessarily wrong. That’s what makes it so dangerous. The CEO, the Chief Medical Officer, the influential Principal Investigator, or the opinionated board member who walks into the room with a “great idea” is often a smart, experienced leader with genuine strategic instinct.

They have pattern recognition. They have authority. They have budget.

What they frequently do not have, is evidence.

And in most health system product organizations, there is no structural mechanism for distinguishing between an initiative backed by validated market evidence and one backed by the enthusiasm of someone with a senior title and a firmly held opinion.

On the roadmap, they look identical.

A feature born from fifty clinician interviews and a feature born from a single executive’s conference takeaway occupy the same row in Jira.

Same priority field. Same projected cost. Same promised benefit. The absence of evidence is invisible — until the product fails.

This is the gap that the Validity Buffer is designed to close.

The Problem of Invisible Evidence Quality

In a typical PI Planning session or quarterly roadmap review, every initiative looks the same. They all have a name, an estimated cost, a projected benefit, and an executive sponsor.

Some have been through months of rigorous Market Grounding work — NIHITO interviews, win/loss analysis, pervasiveness validation. Others arrived because someone important said, “We should really look at this” during a leadership meeting last Tuesday.

On paper? Indistinguishable.

This is not a minor formatting issue. It is a fundamental failure of transparency. To the Strategic Leadership tribe, every item on the roadmap appears equally justified. To the Factory, every item is equally “ready” for the backlog. Nobody is lying. The information simply isn’t there to tell the difference.

And when the HIPPO walks in and drops a new initiative on the table — with energy, conviction, and the implicit weight of their authority (and often additional budget!) — what happens?

The roadmap shuffles. Discovery work is compressed or skipped entirely. The Factory absorbs the new item because the Factory is built to absorb. And the Product Manager, who was six weeks into validating a different problem that three hundred nurses actually need solved, quietly reprioritizes.

We have all watched this happen. Most of us have been complicit in it.

To celebrate the launch of Haverin Consulting, I am providing 50% off an annual paid subscription. Just $25.

Get 50% off for 1 year

Enter the Validity Buffer

The Validity Buffer is a simple concept: a Truth Score for your roadmap.

Instead of evaluating initiatives only by their projected value, we also evaluate the validity of the evidence supporting them. How was this idea researched? What is the quality of the data behind it? Has the problem been validated for pervasiveness, or is it a single anecdote that made it to the right meeting?

Think of it as a confidence coefficient. A discount rate on certainty.

Low Validity (Truth Score: 10–30%):

• The idea originated from a single internal meeting, an executive’s conference experience, or an AI persona simulation (Epistemic Freeloading, as we discussed in Part 3)

• No NIHITO interviews have been conducted

• No win/loss data exists

• The “buffer” is thin — the organization is acknowledging, transparently, that it is placing a bet based on assumption

Medium Validity (Truth Score: 40–70%):

• Some discovery has been conducted, but has not yet reached saturation

• Initial interviews suggest the problem is real, but pervasiveness is not yet confirmed

• The initiative has earned conditional entry to the roadmap — worth further investigation, not yet worth full factory commitment

High Validity (Truth Score: 80–95%):

• The team has followed NIHITO rigorously

• Multiple interviews have been conducted and the team has reached saturation — the point where they stop hearing new complaints and start hearing patterns repeat

• Win/loss data or equivalent internal evidence validates the problem

• The initiative has earned its place on the roadmap through evidence, not enthusiasm

The Validity Buffer does not stop work. It changes the conversation. It forces the Factory to ask: “Do we have enough truth to build this yet?” And it gives the Product Manager a structural mechanism to say “not yet” — even to the HIPPO.

Speaking the Language of Risk

Here is why this matters strategically, not just methodologically.

The HIPPO does not respond well to “we need more time for research.” That sounds like delay. That sounds like resistance. That sounds like the Market Grounding tribe being precious about process while the initiative driver gets more urgent.

BUT…the HIPPO does understand risk.

Reframe the conversation. You are not saying “I think we need more time.” You are saying: “This initiative currently has a Truth Score of 15%. We haven’t validated it enough to bet the quarterly budget on it. Here’s what it would take to move it to 70%.”

That is the language of risk management. That is the language the Strategic Leadership tribe — those Kellogg-trained, P&L-owning, five-year-horizon tribe — actually speaks.

The Validity Buffer is a translation device. It takes the Market Grounding tribe’s insistence on evidence and expresses it in the Strategic Leadership tribe’s language of investment risk. And it gives both tribes a shared, visible mechanism for evaluating whether the Factory should be activated.

Remember the Tower of Babel from Part 1? The tribes all using the same words but meaning different things? The Truth Score is a shared vocabulary that cuts through the noise. “Priority” can mean ten different things. A Truth Score of 15% means exactly one thing: we don’t know enough yet.

The Sweet Spot: Saturation, Not Perfection

A common objection: “We can’t validate everything to 95%. We’d never build anything.”

Correct. And that is not what the Validity Buffer demands.

The goal is not perfection. The goal is saturation — the point at which additional research stops yielding new information. In qualitative research, this is well understood:

• 5–8 interviews will typically uncover approximately 85% of usability issues (the Nielsen Norman “5-user guideline”)

• 9–12 interviews will usually achieve theme saturation — you know what the problem is

• 15–24 interviews will achieve meaning saturation — you understand why the problem exists and what drives the behavior around it

You do not need a thousand interviews to move the needle. Often, five to eight deep, meaningful conversations with real humans will shift a Truth Score from 15% to 65%. That is the difference between “expensive guess” and “informed bet.”

And that is when the Factory should be given the green light.

Saturation is not a finish line. It is a threshold — the point at which you have earned the right to commit organizational resources to a solution. Below it, you are speculating. Above it, you are investing.

What the HIPPO Looks Like on a Truth-Scored Roadmap

Here’s the practical shift.

The board member walks in with a compelling idea. Instead of the roadmap shuffling silently to accommodate it, the organization has a visible, transparent response:

“Great idea. It enters the roadmap at a Truth Score of 10%. Here is the discovery plan to validate it. If it reaches 60% within six weeks, it earns factory commitment in the next PI. If it doesn’t, it remains in the discovery pipeline — visible, tracked, but not consuming factory capacity.”

The HIPPO’s idea is not rejected. It is respected enough to be validated. And the three hundred nurses waiting for their workflow redesign — the initiative sitting at a Truth Score of 85% — do not get bumped because someone important had a good meeting.

The Validity Buffer does not fight the HIPPO. It domesticates it. The opinion still enters the system. It just has to earn its way to the factory floor.

The Organizational Discipline Nobody Wants to Talk About

Okay (deep breath)…Let’s be real.

The reason most organizations do not implement something like the Validity Buffer is not that they lack the methodology. It is that transparent evidence quality makes powerful people uncomfortable.

A Truth Score of 15% next to a board member’s pet initiative is a politically dangerous artifact. It says, in plain language: this idea has not been validated. That is not a message that every organizational culture is equipped to hear.

And yet, it is exactly the discipline that separates organizations that build products of consequence from organizations that generate expensive Zombie Products.

The Delivery Delusion persists because organizations lack the structural courage to say “not yet” to ideas that haven’t earned their place. The Build Trap persists because the Factory lacks a mechanism to distinguish between validated and speculative work. Epistemic Freeloading persists because there is no visible penalty for synthetic evidence.

The Truth Score addresses all three. Not by adding process.

By adding transparency.

In Part 5: Breathing In and Out, we bring the series home.

The three-way war between the tribes is not a problem to be solved or a battle to be won. It is a set of polarities to be managed — and the organizations that master this discipline will build the products that others only talk about.

Stuart Miller is Managing Director of Haverin Consulting, a healthcare IT strategy consultancy. This is Part 4 of The Alignment Files. Subscribe for Part 5.

This is Part 3 of The Alignment Files.

In Part 1, we named the three philosophical tribes at war for the product manager’s soul. In Part 2, we examined how the Factory’s gravity well turns excellent execution into disconnected performance theater.

Today, we confront a very new and more insidious problem: the promise that AI can close the Market Grounding gap without anyone having to leave their desk!

The Temptation

In the Pragmatic Framework, we talk about the NIHITO rule: Nothing Important Happens Inside The Office. The Market Grounding tribe lives by it.

Get outside. Talk to real clinicians. Sit in the workflow. Watch the workarounds. Validate the problem before the factory is activated.

But let’s get real with each other — NIHITO is HARD.

It is time-consuming to recruit clinicians for interviews. It is expensive to travel to sites. It is exhausting to conduct twenty conversations to find that one insight that persists and reframes the entire problem. And in a health system where the factory is hungry, and the next PI cycle waits for no one, the pressure to skip the fieldwork and just get something into the backlog is relentless.

So along comes the Synthetic Persona. A few prompts, and AI generates a “Digital Twin” of your ‘customer’ that answers questions instantly, never pushes back on your scheduling, never cancels at the last minute, and costs nothing.

It feels like a breakthrough.

It is not a breakthrough. It is Epistemic Freeloading — intellectual hitchhiking. And it is quietly becoming the most dangerous shortcut in modern product management.

In celebration of the launch of Haverin Consulting, I’d like to offer 50% of a paid subscription. Thank you for supporting me.

Get 50% off for 1 year

What Is Epistemic Freeloading?

Simply put, it is wanting the authority of evidence without the labor of collecting it.

When a Product Manager uses AI to simulate a customer instead of actually talking to one, they are borrowing from the statistical average of the internet’s past record to predict their product’s future. They get something that “looks like” research. It has themes. It has quotes. It has a nice summary document.

It just doesn’t have grounded truth.

This is not the same as using AI to synthesize what you have already learned from real people. That is acceleration, and it is genuinely powerful.

But simulation is not synthesis. Simulation is substitution.

And the difference, while subtle in the workflow, is catastrophic in the outcome.

Here be Dragons!

Three invisible dangers compound over time:

1. Synthetic Stagnation. LLMs are trained on historical data. They know what happened yesterday; they cannot invent tomorrow. Testing your ideas exclusively against AI-generated personas will only ever get you “average” results.The statistical center of what has already been written, said, and thought. You lose the capacity for genuine discovery.
   

   In healthcare, the next breakthrough often comes from an outlier observation in a clinical workflow nobody thought to study. A synthetic persona will never take you there. It will confidently take you to the middle of the road. Every. Single. Time.

2. Heterogeneity Collapse. Real humans are messy, irrational, and wonderfully contradictory. A night-shift ICU nurse and a Monday-morning outpatient clinic nurse occupy the same role title and inhabit entirely different clinical realities. AI personas are logical and flat. They represent the “statistically average” clinician — a person who does not actually exist. Build for that synthetic average, and you miss the outliers.
   

   In healthcare product design, outliers are often where the most critical unmet needs reside.

3. The Echo Chamber. If the factory is building based on what an AI persona “said,” and that persona was constructed from the assumptions of the person who prompted it, congratulations — you have created a closed loop. You are asking an algorithm what it thinks about a product designed by an algorithm, validated by an algorithm.
   

   The market — the actual, inconvenient, grumpy, contradictory market — is nowhere in the circuit.

Why AI Interviews Aren’t Interviews

There is a deeper methodological problem here that product teams don’t often consider.

In behavioral science, the Observer Effect describes how participants alter their behavior when they know they are being studied. In AI-driven research, this takes on a new and problematic dimension. Human users interacting with AI moderators experience “social desirability bias”. The AI provide the responses they think are expected rather than what they actually experience.

The result is what I’ll call Synthetic Emotion: affective expressions performed for the machine rather than genuinely felt. If your discovery process captures these adapted responses and treats them as ground truth, you are not studying your users. You are studying their performance of being studied.

In healthcare, this matters enormously. Trust is the currency of adoption.

A clinician will tell an AI chatbot that a new documentation workflow “seems fine.” That same clinician, in a face-to-face conversation — after the third coffee, when the formal interview is over, and the real talking starts — will tell you that the workflow adds 7 clicks to every patient encounter. And that they’ve already built a workaround on a Post-it note stuck to their monitor.

NIHITO is not just a research methodology. It is a trust-building exercise. The insights that matter most are the ones people only share when they believe you genuinely care about the answer.

The AI doesn’t ever see the Post-it on the monitor. It never will.

“But It’s Faster!”

The counterargument is always speed. A L W A Y S.

“We can generate insights in hours instead of weeks. We can test ten personas before lunch. We can move faster.”

And this is true.

AI-accelerated discovery is faster. The question nobody asks is: faster toward what?

Speed without ground truth does not close the gap between the factory and the market. It closes the gap between the factory and a simulation of the market. The factory builds faster. The product ships faster.

And that Zombie Product — live, yet not alive — arrives faster too.

Think about it through the lens of what we discussed in Part 2:

• If the gravity well of the factory is the problem of building the wrong things efficiently…

• …then Epistemic Freeloading is the problem of validating the wrong things efficiently

It is the Build Trap with an added veneer of research.

The roadmap looks evidence-based…But the evidence is synthetic. And the organization has no way to tell the difference until the product fails downstream — quietly, expensively, and with no one accountable for the gap.

In the language of Part 1’s tribes, Epistemic Freeloading is what happens when the Market Grounding tribe surrenders its methodology to the factory’s demand for speed.

The tribe still exists. It still has a seat at the table. But it has stopped doing the work that gave it authority.

It has become a hitchhiker, thumbing a ride on the factory’s momentum, borrowing the language of evidence without bearing the cost of collection.

So What Should AI Actually Do in Discovery?

None of this is an argument against AI in the discovery process. I use these tools daily. They are genuinely transformative…when pointed at the right problem.

The distinction is simple: AI is brilliant at synthesis. It is a hot mess to manage as a substitute.

The Workflow:

Stage 1 — Raw Collection (NIHITO. No shortcut.): Conduct real human interviews. In healthcare, 12–15 deep conversations with clinicians, administrators, and operational staff will surface the messy, emotional, contradictory reality that no model can simulate. This is the non-negotiable foundation.

HOT TIP: Ambient Scribing is one of the fastest adopted technologies in healthcare. USE IT!!!

After securing permission to record, I use a Plaud personal recording device to capture interviews. Within 10 minutes of finishing, I have a complete transcript of everything that was discussed, AND the AI has generated a targeted summarized output that analyzes against a specific product field interview template.

Stage 2 — AI-Powered Synthesis: Use AI to transcribe, cluster, and identify thematic patterns across those real conversations. This is where the speed advantage is legitimate and enormous. What used to take a research team two weeks can now be done in an afternoon.

Stage 3 — Dialogic Interrogation: Engage with the AI about the data. Ask it to surface contradictions. Flag outlier responses. Identify gaps in the interview set. The AI becomes a thinking partner, not a thinking substitute.

Stage 4 — Calibration: Validate the AI’s thematic conclusions against a subset of real human responses. If the machine’s patterns do not match the ground truth, you have caught the signal problem before the factory was activated.

AI is the accelerant. Humans are the fuel. Light the accelerant without the fuel and you don’t get a controlled burn. You get a flash and nothing left.

The Uncomfortable Truth

The real reason Epistemic Freeloading is spreading is not that Product Managers are lazy. Most of them are anything but.

It is that the organizational structure rewards the shortcut.

The factory demands a populated backlog. The PI cycle demands committed objectives. The executive team demands a roadmap by Thursday. In this environment, a Product Manager who says, “I need six weeks and access to fifteen clinicians before I can validate this,” is not supported and rewarded for rigor.

They are penalized for the delay.

And so the shortcut becomes the norm. AI personas replace field research. Synthetic validation replaces NIHITO. The backlog fills. The factory runs. And nobody asks whether the evidence underneath the roadmap is real — until the Zombie Products start piling up.

The fix is not to ban AI from discovery. The fix is to stop treating the factory’s appetite as the discovery team’s deadline.

BUT…that requires a mechanism. A way to distinguish between validated evidence and expensive guessing. A way to put a “truth score” on the roadmap so that leadership can see, at a glance, which initiatives have earned their place through evidence and which are riding on assumption and enthusiasm.

That mechanism is the subject of Part 4.

In Part 4: The Truth Score, we introduce the Validity Buffer — a practical tool for surviving the HIPPO (the HIghest Paid Person’s Opinion) and ensuring your roadmap is grounded in evidence, not enthusiasm.

Stuart Miller is Managing Director of Haverin Consulting, a healthcare IT strategy consultancy. This is Part 3 of The Alignment Files. Subscribe for Parts 4–5.

This is the third installment in a series examining high-profile collapses in healthcare technology. Previously: After Olive: Why Healthcare Leaders Don’t Trust AI Promises and IBM Watson Health: Yet Another 10 Year Tale of Woe.

Together, these analyses form a body of work tracing a pattern that healthcare leaders and investors ignore at their peril.

Executive Summary

Pandemic-Fueled Rise: Carbon Health rode a COVID-19 windfall to rapid prominence. In 2020–21, it expanded from a handful of clinics to over 80 locations, leveraging pop-up testing sites and telehealth surges. The company raised $128 million in 2020, then a $350 million Series D in mid-2021 at a $3.3 billion valuation. Leadership boldly aimed to reach 1,500 clinics by 2025 to become “the largest primary care provider in the U.S.” By 2023, CVS Health had invested $100 million in Carbon, which had 125 clinics across 13 states and served over 1 million patients. In total, over $600 million in venture funding flowed into Carbon’s vision of a tech-enabled, nationwide primary care network.

Read more

This is Part 2 of The Alignment Files, a series that continues from The Delivery Delusion.

In Part 1 of this series, we named the three philosophical tribes competing for control of the product function — Strategic Leadership, Market Grounding, and Scaled Technical Execution — and mapped the structural conflict between them.

Now, in Part 2, we go inside the belly of the beast. What happens when one tribe, the Factory (Scaled Technical Execution), wins by default?

The Trap Nobody Sees

In most large health system IT organizations, the delivery engine has become the dominant force in product decision-making. Not because anyone chose it. But because the factory has gravity.

Scaled execution frameworks, like SAFe, promise predictability and coordination. And executed well, they deliver both. But in doing so, they create an organizational pull that is difficult to resist:

• throughput over judgment

• commitment over discovery

• the measurable over the meaningful

The result is not incompetent work. It is often excellent work — technically rigorous, delivered on schedule, passing every quality gate. The problem is that when applied dogmatically, the work is often disconnected from the market problems it was supposed to solve and too short-term to keep the strategic north star in sight.

This is The Build Trap. Not a failure of execution, but a failure of connection. Performance theater, with excellent production values.

The Siren Song of “On Time”

There is a natural pull in large organizations toward certainty.

We love schedules. We love “Green” status reports. We love the quarterly review where the Release Train Engineer can report that 94% of PI objectives were met. It makes the CTO’s heart sing. We are a smooth humming machine!

This is the factory's gravity well. And it is seductive precisely because it is real.

These are not fabricated metrics. Real engineers wrote real code. Real testers validated it. Real deployments went live. The factory is not lying about its output.

It is simply silent about whether that output matters.

When the factory’s cadence becomes the organizing principle of the product function, the questions change. Not consciously — nobody announces “we have decided to stop caring about outcomes.” The shift is gravitational. Slowly, but imperceptibly, the dominant questions become:

• Did we hit the PI commitment?

• Did we clear the backlog?

• Did we ship by Q3?

And the questions that should be dominant?

• Did this actually change clinical behavior?

• Did the patient experience improve?

• Did the operational cost move?

Those are deferred. Not rejected. Just never quite asked with the same urgency, rigor, or accountability.

The Enabler Trap: How the Factory Feeds Itself

Every SAFe organization runs on a dual currency:

Features (user-facing value) and Enablers (technical infrastructure, architecture, debt reduction). Both are absolutely legitimate and necessary. A factory that never invests in its own foundations will eventually collapse.

The problem is not that these are the focus. The problem is what happens when the factory’s prioritization framework — Weighted Shortest Job First (WSJF) — is applied without rigorous market grounding.

WSJF is elegant in theory: prioritize by the cost of delay divided by job duration. In practice, Enablers are easier to estimate, scope, and defend internally. Their cost of delay is framed in the language the factory understands — system stability, security risk, and scalability.

Meanwhile, the cost of delay for a market-driven feature requires evidence that the organization has often not provided, or even gathered, and that this is often the case.

Remember from Part 1:

• The Strategic tribe operates in the kingdom of the North Star on a 3-5 year horizon

• The Market Grounding tribe operates on a 6–18 month horizon.

• The Factory operates on a 3–6 month cycle.

Guess which horizon’s priorities are easier to articulate in a PI Planning ceremony?

The result is a roadmap that gradually tilts toward technical self-maintenance. The factory feeds itself. In the absence of a strong external signal, internal needs expand to fill available capacity.

A CTO reviewing the quarterly portfolio sees infrastructure modernization, security hardening, API refactoring, and platform upgrades. All defensible. All necessary. And somewhere near the bottom, perpetually deferred: the workflow redesign that three hundred nurses have been requesting for two years.

The Enabler Trap is the Build Trap’s quiet cousin.

You’re not building the wrong product; you’re just building the right plumbing for a house nobody wants to live in.

Output Is Not Outcome

The most insidious quality of the Build Trap is that it produces real artifacts of success. Dashboards show velocity. Burndown charts trend correctly. Release notes are published on schedule. The organization’s delivery metrics are genuinely impressive.

This is where the performance theater lies. Not because the performers are faking (they absolutely are not), but because the audience is watching the wrong show. The factory is performing execution.

Leadership loves to applaud execution. The factory is producing ‘better’ than ever. Every launch is a organization wide celebration email. Every drama-free quarterly update is a cause for quiet satisfaction

And yet, no one has the structural authority, or the political incentive, to ask whether the executed work is producing the outcomes that justified the investment.

In healthcare, this disconnect has a specific, recognizable pattern. A clinical decision support tool ships on time and passes every compliance check. But the alert logic fires so frequently that physicians develop “alert fatigue” and dismiss it reflexively. A patient portal launches with full feature parity. But the registration workflow requires fourteen steps, so actual enrollment plateaus at 8%.

The factory did its job. The product failed.

And because no one owns the space between delivery and value realization — as I argued in The Delivery Delusion — the failure is attributed to “change resistance” or “user training gaps.”

The product quietly joins the portfolio of Zombie Products — live yet not alive.

Maintained, supported, budgeted for, and producing nothing.

The Short-Horizon Problem

The factory operates in increments. A Program Increment typically spans 8–12 weeks. Sprints these days run in 2-week cycles. This cadence is useful for coordination. It is catastrophic for strategy.

A meaningful clinical workflow transformation does not resolve in a single PI. It requires sustained discovery, iterative prototyping and testing with real clinicians, and the organizational patience to let behavioral adoption develop over months, not in 2-week sprints.

But the gravity well pulls everything toward the boundary of the increment. Features are scoped to fit the PI. Value is promised at the PI boundary. Success is evaluated at the PI boundary. Anything that cannot demonstrate tangible output within the increment is deprioritized. Not because it lacks value, but because it lacks visibility in the cadence the organization has chosen to worship.

BUT…that strategic north star: the three-year vision; the system-wide care model redesign; the TRANSFORMATION; does not simply disappear. It remains in the executive presentation deck.

It simply loses its connection to the work being done, one PI at a time, as the factory optimizes for what it can show in the next twelve weeks.

The 3–5-year strategic horizon and the 3–6-month execution horizon are not just different timeframes. They are different languages, different tribes, different definitions of what “progress” means. And as we established in Part 1, when the tribes can’t align between those definitions, the loudest cadence wins.

Breaking the Orbit

Escaping the gravity well does not mean dismantling the factory. The factory is necessary. The engineering discipline, the coordination mechanisms, and the deployment infrastructure are all real, all valuable. The factory is the means of production. The question is whether what is being produced matters.

The Prescription:

Upstream: The factory must not be activated until the problem has earned its place on the roadmap through validated evidence — not executive enthusiasm, not vendor promises, not a single anecdotal complaint from a politically connected department head. This is the upstream judgment that The Delivery Delusion identified as missing, and it is the subject of Part 4 of this series.

Downstream: Delivery cannot be the terminal event of product responsibility. Shipping is a milestone, not an achievement. If no one owns the question “Is this product producing value?” after go-live, the factory will celebrate the launch and move on to the next PI, never looking back. Zombie Products are born in this gap.

Laterally: The factory’s cadence must include protected capacity for discovery. Not as a concession or an “innovation spike” bolted on at quarter-end, but as a structural commitment. If 100% of engineering capacity is locked in PI objectives with zero room for investigation, the organization has formally decided to stop learning.

The Build Trap is not a failure of quality. It is a failure of connection. The factory is not broken. It is simply unmoored — running hard, running well, running in the wrong direction.

In Part 3: Intellectual Hitchhiking, we will examine the seductive promise that AI can close the discovery gap — and why speed without ground truth doesn’t get you there quicker.

It gets you somewhere else entirely.

Stuart Miller is Managing Director of Haverin Consulting, a healthcare IT strategy consultancy. This is Part 2 of The Alignment Files. Subscribe for Parts 3–5.

This is Part 1 of The Alignment Files, a series that continues from The Delivery Delusion. Whereas that paper diagnosed the structural failure of delivery-as-success, this series examines the deeper philosophical fractures that make that failure inevitable.

And maybe some prescriptions to gain alignment.

An Existential Question

Health systems are increasingly investing in digital product management. They fund the teams, create the titles, buy the badges, and hire for the role. And yet the function consistently underperforms. Not because the people are weak, but because the organization has never resolved a fundamental question:

What does product management actually do?

Typically, three distinct (and legitimate) philosophies compete for control of the product function. Each one carries its own definition of objectives, KPI’s, success, and operating logic. And its own gravitational pull on resources and authority to make decisions (often times on behalf of the other).

When these philosophies are unreconciled or misaligned, product teams do not fail because of incompetence. They fail from a lack of coherence.

Let’s define the three combatants, map the structural conflict, and define why the failure to manage the tension between them is the root cause of so much of the product dysfunction and dissatisfaction from within health system.

The Disease Beneath the Symptom

In The Delivery Delusion, I argued that many Internal Product Organization routinely deliver technically sound solutions. BUT they fail to produce the outcomes that justify their existence. The factory runs, and yet…outcomes and value never arrive. User customers remain disappointed, jaded.

The Delivery Delusion came out of observing the structural misalignment between three philosophical traditions that now shape organizations' definitions of Digital Product Management.

Each methodology is internally consistent. Each produces genuine insight. And each, when left unchecked and to its own devices, will attempt to dominate the conversation.

The ‘Players’

Strategic Leadership — the product manager as business leader, responsible for vision, P&L, and long-term positioning aligned with organizational imperatives. Exemplified by the Kellogg School of Management’s “CEO of the Product” model.

Market Grounding — the product manager as investigator, responsible for discovering and validating real-world problems before anything is built. Exemplified by the Pragmatic Institute’s NIHITO principle: Nothing Important Happens Inside The Office.

Scaled Technical Execution — the product manager as delivery coordinator, responsible for translating intent into shippable increments through a disciplined process. Exemplified by SAFe and similar adjacent Agile frameworks.

In theory, there is immense value in each of these methodologies' perspectives.

In practice, these traditions do not politely coexist. They compete for the Product Manager’s time and soul, organizational legitimacy, and, most importantly, the prioritization and control of the roadmap and resources.

What Each Gets Right—Where Each Goes Wrong.

Strategic Leadership:

Horizon perspective — 3-5 years; “The North Star”

Principled Stand — insists that product decisions are business decisions. It demands financial literacy. It refuses to let product management retreat into a purely tactical role. Essential questions: “Does this align with our five-year plan? What is the total cost of ownership? What are we not doing to do this?”

Failure mode — altitude sickness. It operates at 30,000 feet, where the oxygen of clinical reality is thin. A product vision can be strategically coherent, financially modeled, beautifully presented — and yet completely disconnected from the nurse's workflow who will be expected to use it. When it dominates unchecked: Visionary Paralysis. Extraordinary clarity about the future BUT…Almost nothing shipped moves toward it.

Market Grounding

Horizon perspective — 6-18 months; “The needs of the user”

Principled Stand — provides a structural defense against the pathologies that plague health system IT: executive intuition masquerading as strategy, vendor promises substituted for user evidence, the “vocal minority” problem where one politically powerful stakeholder dictates direction for an entire population. Its insistence on equal win/loss analysis yields market intelligence that internal product teams rarely possess.

Failure mode — incrementalism. By anchoring so firmly in observed problems, it can become overly responsive to current complaints at the expense of a transformative vision. It excels at identifying friction; it is less equipped to imagine a fundamentally different model of care. When unchecked, exquisite awareness of today’s problems, but a structural inability to lead toward tomorrow’s solutions.

Scaled Technical Execution

Horizon perspective — 3-6 months; “The release schedule”

Principled Stand — addresses a real need: coordination at scale. When a health system has hundreds of developers, dozens of product lines, and regulatory constraints touching everything, disciplined delivery cadence is a rational response to genuine complexity. It makes the invisible visible and the chaotic manageable.

Failure mode — Gravity Well (Black Hole). What I described in The Delivery Delusion: the factory becomes the center of gravity. The roadmap ceases to be a strategic instrument and becomes a feeding schedule. Capacity seeks utilization. The question shifts from “Should we build this?” to “Can we fit this into the next PI?” When it dominates unchecked — and in most large health system IT organizations, it does — the result is a high-performing factory that generates expensive waste.

Why the War Is Structural

It would be convenient to blame individual failure. A strategy leader remote from front-line operations; a Product Manager who is too connected to the voice of the user; a Product Owner who can’t prioritize; a CTO who over-indexes on process. But the conflict is architectural. The limitation is the inability of each ‘tribe’ to be able to horizontally switch

The Chief Strategy Officer thinks in five-year horizons and in terms of competitive positioning. The Chief Digital Officer is supposed to reflect the user's voice. The Chief Technology Officer or Chief Information Officer owns the factory and is measured on throughput and efficiency.

Each is incentivized according to a different definition of product success.

The product manager (if you are lucky!) sits at the intersection — serving all three masters, typically with no power or structural mechanism to resolve the contradictions and conflict. This is not a coordination problem that better meetings can solve. It is a governance and alignment failure. The organization has never defined how the tribes will resolve their differing horizons and priorities, far less which tribe has primacy, under what conditions, and with what accountability.

In the absence of that decision, the loudest voice or the most politically influential wins.

Many times, the factory is loudest — because it has the most visible metrics, the most tangible artifacts, and the most immediate political consequences for failure. The factory wakes up every day with one objective: keep the factory running…survive…grow.

The SAFe methodology compounds this by bifurcating the product role into Product Manager (strategic) and Product Owner (tactical). In theory, this creates focus; in practice, more barriers to understanding.

The overall effect of these different viewpoints and methodological approaches is it creates a Tower of Babel. Everybody is talking. The words sound similar. But “roadmap” means something entirely different to the strategist than it does to the factory. “Priority” means one thing to the Chief Strategy Officer and another to the Scrum Master. “Value” is a financial concept in one room and a velocity metric in the next.

They think they are communicating. They are not.

And the confusion itself generates conflict and offense.

Meanwhile, the people writing code never hear the frustrated cardiologist describe why the referral workflow adds fifteen minutes to every encounter. They receive requirements: sanitized, decomposed, stripped of the context that would let them exercise judgment. And the PM — who should be in the field gathering that context — is pulled into PI Planning ceremonies and stakeholder alignment, translating between tribes who have not yet realized they are speaking different languages.

Why does Healthcare Get Punished Hardest?

There is a difference between Outside Product Organizations (most commercial companies) and Inside Product Organizations (Healthcare and many government, education, and similar organizations).

In a commercial SaaS company, philosophical misalignment eventually surfaces as lost revenue. Customers churn. The users provide the market with very direct, existential feedback through their wallets. The Factory is driven by the singular metric—Revenue.

Health systems lack this corrective. Internal products do not churn. They persist. A clinical tool that fails to gain adoption is not decommissioned; it is maintained. An analytics platform nobody trusts is not sunset; it is “enhanced.”

These are the “Zombie Products” — Live, and yet, not alive.

The cost of misalignment is not visible as lost revenue. It compounds silently as organizational drag — systems that consume resources without producing value.

The three-way war is not an abstraction. It is the structural explanation for why a health system can employ talented product managers who understand product lifecycle, invest in sophisticated frameworks, and still maintain a portfolio of digital products that clinicians do not trust, administrators do not use, and patients never see.

The Path Forward: Alignment, Not Victory

The resolution to the conundrum is not to pick a winner.

• Strategy without market grounding produces visionary irrelevance.

• Market evidence without strategic framing produces reactive incrementalism.

• Both—useless without a factory capable of turning intent into elegant working software.

The failure is not in the traditions of each discipline. It is in the refusal to hold all three simultaneously, and in negotiating and managing the tension as ongoing discipline.

This is the work the remaining articles in this series will take on:

Part 2: The Factory’s Gravity Well examines how the Build Trap operates as performance theater.

Part 3: Intellectual Hitchhiking confronts the question of why AI-accelerated discovery may simply speed the organization over a cliff.

Part 4: The Truth Score introduces a mechanism for surviving the HIPPO — the Highest Paid Person’s Opinion.

Part 5: Breathing in and Out argues the case that this tension is not a problem to be solved or a war to be won, but a series of polarities to be managed.

The factory will always be loud. Strategy will always be seductive. The market will always be grumpy and inconvenient.

The question is, does your organization have the discipline to balance all three, or will it keep letting the loudest voice win by default?

Stuart Miller is Managing Director of Haverin Consulting, a healthcare IT strategy consultancy. Subscribe for Parts 2–5.

Below is a practical chronology (1990 → today) of personal knowledge management (PKM) software on personal computers—covering classic “PIM” organizers, free-form databases, outliners, mind maps (including Tony Buzan), hypertext/graph “brain” tools, wikis, research libraries, and the modern backlink/AI wave.

I’m going to start with a quick “pre-1990 DNA” note, because a lot of the weird indexing ideas and stuff were already proven before Windows 3.x. and macOS were a thing!

Pre-1990 DNA (the ideas that later PKM tools keep reinventing)

• Hypertext as personal knowledge base: systems like HyperCard (released 08/11/1987) made “cards + links” a mainstream concept on Macs and eventually everywhere.

• Research-grade hypertext: Xerox PARC’s NoteCards (1984) modeled notes + links + “fileboxes” as a personal knowledge base.

• Free-form text databases: early Lotus Agenda (DOS) and askSam showed you could dump unstructured text in first, then classify/search later. Lotus Agenda is explicitly described as “free-form.”

These three foundations—hyperlinks, free-form capture, and post-hoc indexing—drive almost everything from 1990 onward.

1990–1996: The “PIM” era + free-form/outliner experiments

This is the age of contacts/calendar/todos plus power-user indexing.

• Lotus Organizer (Windows era begins): becomes a flagship “organizer” style PIM in the 1990s. Evidence of Organizer for Windows exists by 1993.

• Ecco Pro (1993): outliner-based PIM with database-like “fields” that let you filter/sort your notes like a spreadsheet—one of the best examples of “cross-indexing” in consumer PKM.

• Info Select (early 1990s): positioned as a do-it-all data manager / tickler / project tracker—part organizer, part database.

What changed in this era:

• PKM splits into two camps:

  1. Planner PIMs (calendar/contact/task centric)

  2. Free-form information managers (notes/outlines + user-defined fields + search)

1997–2003: Visual mapping + “your brain in software” + early capture tools

• TheBrain / PersonalBrain (1.0 released 1998): one of the earliest mainstream tools that explicitly models knowledge as a navigable graph, not folders.

• MindManager (released 1998): commercial mind mapping for business/project thinking; becomes a “serious work” mind map tool (not just brainstorming).

• OneNote (announced 11/17/2002; shipped as OneNote 2003): free-form page canvas + fast capture becomes Microsoft’s answer to “digital notebook.”

Tony Buzan & mind mapping in the same timeline arc

• Buzan popularized mind mapping (not “invented visual notes,” but popularized/formalized his technique). Wikipedia summarizes him as popularizing mind mapping and “radiant thinking.”

• His “formal introduction” of Mind Maps is often tied to the BBC series Use Your Head* (Spring 1974)—an important connection because later software (MindManager, iMindMap) explicitly references the Buzan method.

What changed in this era:

• The “folder tree” is starting to lose its monopoly. Visual/graph navigation becomes credible for individual knowledge work.

2004–2011: Wikis, research libraries, and the “capture everything” cloud

This is where modern PKM really takes off.

• TiddlyWiki (first released 09/30/2004): a single-file personal wiki—nonlinear notes, easy linking, intensely customizable.

• DEVONthink (DEVONthink 1.0 released April 2002; grows through the 2000s): Mac-heavy “research vault” approach—OCR, email archiving, search/AI-ish classification features for its time.

• Zotero (first released 2006): makes “personal research library + web capture + citations” accessible, especially for academics and evidence-heavy professionals.

• Evernote (web service open beta 06/24/2008): the “external brain” promise goes mainstream: fast capture + sync + later retrieval. (This is the one you explicitly asked to include as part of your personal history.)

What changed in this era:

• Two big shifts:

  1. Ubiquitous capture (web clipper + phone + sync)

  2. Search beats filing (tags + full-text search become the center)

2012–2018: Workspaces, databases, and “docs that act like apps”

• Notion: founded 2013; Notion 1.0 released Aug 2016; Notion 2.0 released March 2018—a big move toward “notes + databases + lightweight internal wiki” in one place.

What changed in this era:

• PKM becomes workspace software: your notes can behave like structured systems (tables, relations), not just pages.

2019–2021: Networked thought goes mainstream (backlinks + local-first)

This is the “Roam → Obsidian/Logseq” inflection.

• Roam Research (released 2019): popularizes backlinks as the default way of thinking in notes for a broad audience.

• Logseq (released 10/01/2020): open-source, local storage, graph-first, outline-first workflows.

• Obsidian: becomes the flagship “local markdown vault + backlinks + graph view” model. (Wikipedia summarizes it as a markdown-file PKB.)

What changed in this era:

• A lot of serious users stop trusting “forever SaaS” and move to local-first plain text + link structure.

2021–today: Visual canvases + AI-assisted recall + “notes that think”

• Heptabase (founded 09/2021): pushes the visual canvas metaphor—cards on a board for sense-making over time.

• Evernote’s AI-era reboot (recent major updates): Evernote continues evolving under new ownership and has shipped significant AI-focused changes in recent versions (e.g., v11 highlights an AI assistant and a semantic-ish search direction) I’ll wait and see how this develops!.

• TheBrain continues (latest releases through 2024): still one of the purest “associative network” tools and is actively maintained.

What changed in this era:

• “Search” is no longer enough; products compete on:

  • retrieval (semantic search / “ask my notes”)

  • synthesis (summaries, structured extraction)

  • workspace cognition (canvas + graph + timeline views)

A simple way to read this history (so it’s useful, not trivia)

Most tools fall into 5 recurring “design philosophies” that cycle every ~8–12 years:

1. Planner-first PIM (calendar/contacts/tasks)

2. Free-form database / outliner (capture first, structure later: Ecco Pro)

3. Wiki/hypertext (TiddlyWiki lineage)

4. Graph/associative (“brain software”: TheBrain, Roam, Obsidian, Logseq)

5. Workspace + databases (Notion)

Evernote sits in a sixth “mass-market capture + cloud sync” category that influenced everyone—even people who later left it for local-first systems.

Why this exists

If you’ve used ChatGPT heavily for strategy, writing, planning, or problem‑solving, your chat history is no longer “just conversations.” It’s a record of how you think: ideas tested, decisions made, frameworks refined.

The problem is that ChatGPT is optimized for in‑the‑moment interaction, not long‑term reuse. Even with search, most past thinking is effectively trapped behind scrolling and guesswork.

This page describes a practical, non‑technical method for turning your full ChatGPT history into a searchable, analyzable knowledge archive — a true second brain.

What this method does

This is not about exporting data for the sake of it.

Using a simple, one‑time local process, you will:

• Safely export your entire ChatGPT history

• Convert it into document‑style sources that NotebookLM can ingest reliably

• Analyze years of your own thinking as a single corpus

Once complete, you can ask higher‑order questions like:

• What themes recur most often in my work?

• What decisions have I already reasoned through?

• Which ideas are worth turning into content, talks, or products?

  ---

The User guide describes how you can use the method with both Windows and macOS devices.

The included Python script will work on both macOS or Linux, provided you follow the instructions in the guide relevant to your system.

What’s included

This method is delivered as a small, self‑contained package:

• One‑page Quick Start (PDF)— the entire process at a glance

• Full guide / mini‑course (PDF) — explains the why, the constraints, and the exact steps. Includes a “Starter” prompt pack for NotebookLM. Questions you can use immediately to explore your new “Second Brain”.

• Python conversion script — the same “easy button” script I used to prepare my own archive

No coding experience is required. The script is run once, locally, and your original data remains unchanged.

Companion videos (optional)

If you prefer seeing the process demonstrated you can find walkthroughson my YouTube channel:

• Video 1: How I found out that I was a TOP 1% ChatGPT user

• Video 2: Why your ChatGPT history is TRAPPED — and the method to unlock it

• Video 2: Using a ChatGPT “SECOND BRAIN” in practice — Usage in NotebookLM

These videos are optional. Everything required to execute is documented in the guide.

Who this is for (and who it isn’t)

This is for:

• Heavy ChatGPT users

• Consultants, executives, writers, product leaders

• People who think in drafts, frameworks, and iterations

This is NOT for:

• Casual or novelty users

• People looking for prompt lists or AI hacks

• Anyone unwilling to run a simple one‑time local script

How to get it

Option 1 — One‑time purchase (lowest friction)
Buy the complete package via Ko‑fi for $15 $10.

Buy on Ko-Fi

Option 2 — Included with Substack Annual
Annual paid subscribers receive this download, along with future updates to the method and the entire range of other related tools or premium assets that will be published over time.

Read more

Okay, disclaimer before I get into this.

I am an enthusiastic user and consumer of LLM and GPT technologies. I enjoy using them.

I’ve had the good fortune to work alongside teams that are embracing these technologies at fundamental and applied levels in healthcare right now. Their aim is to improve healthcare delivery and reduce the administrative burden on clinicians, all to improve the patient experience and outcomes.

BUT…

I am what I would describe as an enthusiastic skeptic and a sceptical enthusiast.

So what follows is my take, based on a solid understanding of the underlying technologies and the challenges of using what is still relatively immature.

And the five constraints I describe are the same ones we have been struggling with in clinical and applied healthcare information technology for as long as I have been in healthcare … and that’s a very long time.

So what’s the deal?

In January 2026, OpenAI and Anthropic each announced healthcare-specific AI platforms. The headlines were immediate and enthusiastic:

My first reaction wasn’t so much excitement.

It was a question, a BIG one!:

They do understand that the chart, the EMR, is not the patient? Right?

I’ve spent over 20 years in healthcare IT sales, working with hospital systems and health technology vendors across the United States. I’ve pitched, sold, and implemented: EMR’s; clinical decision support tools; billing and coding solutions; and analytics platforms.

I know the demos and the workflows. I know the promises. I know what these systems are supposed to do (and why).

But I also know what they actually do—because I live on the other side of the transaction.

I’m a double lung transplant recipient, 2.5 years post-transplant, under continuous rejection surveillance. That means monthly blood draws for tacrolimus levels and metabolic panels, biannual comprehensive evaluations, and constant coordination across pulmonary, infectious disease, nephrology, and cardiology teams.

My medical history spans two Epic installations—Vanderbilt University Medical Center (where I was diagnosed and transplanted) and Mayo Clinic Rochester (where I receive ongoing care). They’re both Epic. They use Community Connect.

And (SPOILER!) neither system (despite Community Connect) has the complete truth about me.

I’m also the father and primary caregiver for my 39-year-old son, Jamie, who has, over the past 18 years, suffered from complex pancreatic disease with significant congenital and genetic components that has deteriorated over time.

In the past two years alone, he’s undergone pancreatectomy, spleen removal, partial liver resection, upper bowel resection, and partial stomach resection. His care involves 12 specialists at Mayo Clinic—gastroenterology, pancreatobiliary surgery, endocrinology, hepatology, dietetics, neurology, pain specialists, geneticists, and others.

Every one of them has access to “his chart.”

None of them sees the same patient. The WHOLE Patient.

What OpenAI and Anthropic Actually Announced

Let’s be precise about what was launched in January 2026, because the gap between what vendors said and what the market heard is already widening.

OpenAI’s ChatGPT Health introduces a dedicated health experience within ChatGPT. Users can link personal health data from medical record exports or wellness apps like Apple Health. The system then uses that data as context for personalized responses about health trends, lab results, and appointment preparation.

OpenAI explicitly states that ChatGPT Health is not designed for diagnosis or treatment recommendations. It’s intended for interpretation, navigation, and preparatory uses. Connected health data is encrypted, compartmentalized, and not used to train their models.

Anthropic’s Claude for Healthcare was unveiled at the J.P. Morgan Healthcare Conference. The announcement emphasizes HIPAA-ready infrastructure and connectors to authoritative data sources, including CMS coverage databases, ICD-10, and NPI registries. The focus is on administrative tasks: prior authorization, claims support, and billing workflows.

Like OpenAI, Anthropic positions this as a tool to augment workflows, not a direct clinical decision engine. Neither vendor claims diagnostic authority. Neither claims to replace clinical judgment. Both are explicit about limitations.

And yet.

The commentary I’m seeing across healthcare and tech media frequently frames these launches as if AI will now be able to understand the patient and their health/disease.

As if connecting to medical records equals reasoning about clinical reality. As if natural language interfaces suddenly solve the underlying data quality, fragmentation, and context problems that have plagued healthcare informatics for decades.

They don’t.

And now, Here’s the Beef!

The Five Structural Constraints That Won’t Go Away

Over the next few weeks, I’ll be publishing a detailed analysis of why even the most sophisticated language models struggle with healthcare data. Not because the models aren’t good—they are (for what they are)—but because the data substrate itself is fundamentally incompatible/or inaccessible for the kind of reasoning these platforms promise.

Here are the five constraints we’ll explore, drawn from both clinical informatics research and my lived experience navigating two complex disease processes across multiple health systems:

1. Electronic Medical Records Are Not Ground Truth

EMRs were designed primarily for documentation, billing, and care coordination—not epistemic certainty. The clinical aspecte came later and have predominantly design ed to support the primary mission…get a bill out!

They are socio-technical artifacts, not clinical truth repositories.

I know this truth as I have tried to introduce some of the EMR technologies into countries that do not support the US Payer/Provider split, with its endless administrative dance of billing and record reconciliation to play the bill/deny/clarifypay arbitrage game. That made a number of established workflows in those countries irrelevant!

In addition, errors persist across encounters. Copy-forward documentation propagates inaccuracies. Problem lists accumulate historical mistakes with limited governance for correction. Provenance is ambiguous—you often can’t tell who entered data, under what context, whether it was observed or inferred, or whether it remains valid.

It took me 6 months and 8 requests with different providers before an erroneous tracheostomy procedure code was removed from my record. I swear the exterior of my neck is intact and unscarred!

When a language model uses Retrieval-Augmented Generation (RAG) to “ground” its responses in medical records, it retrieves what is documented rather than what is true.

And because LLMs render information fluently and confidently, incorrect or outdated information becomes more trusted, not less—especially by non-clinicians or time-pressured physicians.

The constraint: Fluency is not accuracy. Access to the record is not access to truth.

2. “Normal” Is Contextual, Not Absolute

Most laboratory reference ranges are derived from population statistics—the central 95% of values in a reference cohort. They are not personalized. They don’t reflect what’s normal for a specific patient with chronic disease, ongoing therapy, or individual biological baselines.

Every month, I have a lab visit. Tacrolimus levels, BMP, and CBC. My white cellcount is always abnormal high, my creatinine is always elevated, and my EGFR is always abnormal low. But they are normal for me on my anti-rejection regimen and where I am in my lung transplant journey. In fact, compared to many of my cohort, I’m one of the very normal and boring ones. That is true of many patients on active long-term or surveillance care.

An AI summarizing my labs would see repeated abnormal flags and might suggest concern—when in reality, this is my stable baseline.

Diagnostic interpretations are similar. Automated ECG systems over-call abnormalities. Radiology reports contain templated language that may overstate certainty. What clinicians learn to filter as routine noise, AI systems re-elevate as authoritative findings.

The constraint: Population norms are not patient norms. Flagged abnormalities are not clinical emergencies. Context-free summarization is clinically dangerous.

3. Historical Records Are Fragmented, Lossy, and Often Irretrievable

Longitudinal patient understanding requires more than a sequence of encounters. It requires knowing what happened, when it happened, why, and what’s missing.

My transplant history is in Vanderbilt’s Epic. My ongoing surveillance is in Mayo’s Epic. Community Connect shares some data between them, but not all.

Genetic test results are often PDFs. Imaging from outside facilities is sometimes uploaded, sometimes not. Medication histories are incomplete when care crosses institutional boundaries.

For Jamie, the situation is worse. Two decades of chronic pancreatitis across seven health systems, subspecialty consults that never made it into the primary record, genetic variants that were interpreted years ago and never updated as new evidence emerged.

What’s missing is invisible to AI. Retrieval-augmented systems retrieve what’s available, not what’s absent. They cannot say, “This record is incomplete in these specific ways.” They present partial histories with high narrative confidence, masking the fact that critical episodes, interpretations, or corrections were never captured.

The constraint: Absence is silent and invisible to the AI’s. Fragmentation is structural. AI amplifies false completeness because it confidently believes it sees all!

4. Cross-Specialty Diagnostic Synthesis Is Not Automatic

Modern medicine is organized around specialization (and even subspecialization) for good reason: biological complexity demands deep domain expertise. But electronic medical records aggregate data across specialties without integrating meaning across those domains.

Jamie has 12 specialists. Each clinician’s notes are optimized for their local decision-making. Gastroenterology documents GI function. Endocrinology tracks metabolic markers. Hepatology monitors liver enzymes. Pain management records medication adjustments. Dietetics tracks nutritional intake.

Every note is in the same Epic system. But nowhere does the system synthesize: How do these findings interact? Which is primary? Which is downstream? What’s the unified explanation?

For that, you need a really smart, diligent, competent clinician who can lift their eyes from their own little body system silo and see the bigger picture. Guess what. The majority have no time for that. It's a volume business, baby!

Language models excel at describing the relationship between words and language concepts. They struggle to adjudicate them. When an LLM summarizes a multi-specialty record, it typically produces what is call narrative stitching—a fluent story that connects disparate findings linguistically without committing to a causal hierarchy.

That’s not synthesis. That’s summarization.

The constraint: Data co-location is not conceptual integration. Fluent narrative is not diagnostic reasoning.

5. Genomics and Genotype-Phenotype Complexity Break Record-Centric Models

Genomic data does not behave like traditional clinical observations. Variant interpretation is probabilistic, evolving, and deeply context-dependent. A genomic result stored in an EMR is not a fact—it’s a snapshot of scientific consensus at a moment in time.

Jamie’s genetic markers (A1AT, CFTR, others still being evaluated) inform his disease process, but their clinical meaning changes as research evolves. Variant classifications get updated. New interactions are discovered. Population data refines penetrance estimates.

In addition, we have come to understand our genetic changes in response to the body and other environmental factors. My own pulmonary fibrosis wasn't a baked-in time bomb. It was probably some of that predisposition, but with lifestyle and environmental factors that accelerated the cellular changes that led to the demise of my lung tissue!

EMRs are not designed to store or reason over this kind of data. Genomic reports are typically attached as PDFs. Interpretation is still as much an art as a science, and it is also contained in narrative reports (PDFs). Variant-level data is not normalized because it is highly person-specific, not disease-cohort-specific. Links to evidence bases are lost. Updates in interpretation don’t propagate.

When an LLM summarizes a genomic report, it inherits whatever was documented at the time—potentially outdated, potentially incomplete, definitely static. Without active reinterpretation pipelines, AI systems simply re-surface old conclusions with greater fluency and confidence.

The constraint: Static records cannot hold dynamic biology. Genomics exposes the limits of document-centric reasoning.

What’s Coming in This Series

Over the next few weeks, I’ll unpack each of these constraints in depth. You’ll see:

• Published research from clinical informatics shows these aren’t edge cases—they’re structural realities

• My personal experience as both a transplant recipient and caregiver navigating these gaps

• What healthcare leaders should ask vendors before deploying these platforms

• And what should providers and their patients expect from these tools as they go mainstream?

Summary

This isn’t a critique of OpenAI, Anthropic, or large language models. These platforms are meaningful enablers. They improve access, reduce friction, and open new forms of interaction with health data.

But they are not, by themselves, engines of clinical understanding.

The hardest problems in health AI are not solved by better language models alone. They’re solved by acknowledging the limits of the data substrate, designing governance that surfaces uncertainty rather than obscuring it, and resisting the temptation to mistake fluent summarization for genuine insight.

Because the chart is not the patient. It never has been.

What You Can Do Now

If you’re a healthcare executive, before even evaluating these AI vendors, start asking different questions:

• How does your system handle error persistence in medical records?

• How do you distinguish population norms from patient-specific baselines in care assessment and aggregate analysis?

• How do you represent missing data or temporal gaps?

• How do you adjudicate conflicting findings across specialties (believe me it happens)?

• How do you handle evolving genomic interpretations and their fluid nature for that individual patient?

If you’re a clinician being asked to trust AI-generated summaries, demand transparency:

• What data sources were used?

• What was excluded or unavailable?

• What assumptions were made?

• Where is uncertainty represented?

If you’re a patient or caregiver, recognize that “chatting with your chart” is not the same as understanding your health. The chart is a starting point. It’s not the destination.

Stuart Miller is a healthcare technology executive and advisor with over 20 years of experience in enterprise strategy and client transformation. He is also a double lung transplant recipient and caregiver for his son, who has complex pancreatic disease. He writes at the intersection of healthcare systems and lived experience.

If you enjoy this article, you can find the follow-up series, The Alignment Files, starting with the link to Part 1 of a 5-part series below.

In this series, I dig deeper into the root causes and whythis is one presentation of a broader problem for Product Managers in the Healthcare Digital domain of hospitals and health systems.

Why This Matters Now

Health systems are investing heavily in digital platforms, clinical tools, analytics, and workflow automation. Many of these initiatives are delivered on time, technically sound, and compliant—yet still fail to achieve sustained clinician adoption or meaningful operational impact.

This is not primarily a technology problem. It is a product judgment problem.

The Structural Problem in Health Systems

Health system IT and digital organizations have become highly effective delivery engines. Agile, SAFe, DevOps, and vendor implementation playbooks have professionalized execution. However, this success has created an imbalance: Delivery excellence has replaced product judgment as the primary success signal.

As a result:

• Products, both vendor- and in-house-developed, are approved without rigorous validation of their clinical or operational value.

• Go-live is treated as success.

• Adoption struggles are blamed on “change resistance”.

• Systems are maintained and persist long after their value has eroded.

Two Predictable Failure Modes in Healthcare Digital

1. Upstream Failure: Initiating Without Clinical or Operational Legitimacy

Initiatives are often approved based on executive sponsorship, strategic narratives, vendor positioning, or isolated complaints.

What is often missing is evidence of scale, clarity on the clinical or operational decision being improved, explicit tradeoffs, and a credible answer to: “What happens if we do nothing?”

Without this discipline, health systems activate expensive digital initiatives that were never positioned to deliver meaningful value.

2. Downstream Failure: Implementation Without Clinical Success

In healthcare, implementation is not success.

A system can be live, integrated, trained, and supported, and still fail because it disrupts clinical workflow, adds cognitive load, conflicts with professional judgment, or lacks trust.

• Adoption becomes partial.

• Advanced features go unused.

Failure is not technical; it is behavioral

The Bottom Line for Health Systems

Digital transformation fails when the organization mistakes activity (shipping features) for achievement (changing outcomes).

The Factory will always be loud; it demands to be fed.

Leadership’s role is not to speed up the line, but to assert the Upstream Judgment to validate the work and the Downstream Accountability to measure the result.

Internal product organizations do not fail because they lack the capacity to build; they fail because they build the wrong things efficiently.

When delivery excellence substitutes for clinical and operational impact, the result is a high-performing factory that generates expensive waste.

The true advantage lies in the organizational discipline to decide

A.) what not to build upstream, and;

B.) the rigor to ensure value is realized downstream.

The Persistent Paradox of “Successful” Failure

Across internal IT, digital, and informatics organizations, particularly within large healthcare and hospital systems, there exists a persistent and deeply frustrating paradox.

Product teams routinely deliver solutions that are technically robust, compliant with specifications, validated through testing, and deployed into production environments on schedule. From a delivery standpoint, these initiatives are frequently labeled as successes to be celebrated.

And yet, when examined over time, many of these same products fail to produce the outcomes that originally justified their existence. Users adopt them unevenly or reluctantly.

Advanced features remain unused. Workarounds persist. Stakeholders express dissatisfaction that the product is “not doing what we expected,” even though it is doing precisely what it was designed and built to do.

Eventually, the product becomes quietly sidelined—maintained, supported, and paid for, but no longer central to the work it was meant to improve.

This paper argues that these failures are neither accidental nor primarily technical. They are structural.

They arise from a deeply ingrained misplacement of emphasis within product organizations, where execution excellence has become the dominant—and often exclusive—measure of success.

The Factory as the De Facto Center of Gravity

Over the past two decades, internal product organizations have professionalized delivery at an unprecedented level.

Agile methodologies, scaled frameworks, DevOps practices, and modern engineering platforms have dramatically improved teams' ability to build and deploy software.

However, these same advances have also produced an unintended consequence: the delivery engine—the factory—has become the moral and operational center of the product system.

In practice, this manifests in subtle but powerful ways. Organizational language shifts toward throughput, velocity, and capacity. Funding models align around project delivery milestones. The factory becomes not just a mechanism for execution, but a justification for its own existence.

This paper does not argue that factory excellence is undesirable. On the contrary, without a capable factory, even the best product judgment cannot be realized.

The problem arises when factory operations crowd out two equally essential responsibilities: upstream judgment and downstream value realization.

Internal Products and the Illusion of “Non-Commercial” Work

A further complication in internal IT and digital organizations is the widespread belief that internal products are fundamentally different from commercial ones because they do not directly generate revenue.

In commercial environments, poor product decisions are punished quickly and visibly. Customers do not buy products they do not value.

Internal products lack this immediate feedback. They do not fail loudly. Instead, they fail quietly.

The absence of revenue does not mean the absence of economics. Internal products consume scarce and valuable resources: skilled labor, leadership attention, infrastructure, compliance effort, and opportunity cost.

The difference is not that internal products are non-economic, but that their economics are obscured. This obscurity makes disciplined product judgment harder, not less necessary.

Framing the Core Question

The central question this paper seeks to answer is not:

How can we deliver faster?
or,
How can we improve adoption through better training?

Those questions are downstream of a more fundamental one:

How do product organizations ensure that they are solving the right problems, and that the solutions they deliver actually produce sustained, meaningful value in the environments where they are deployed?

Why the Pragmatic Framework is the Right Reference—Even When It Is Resisted

The Pragmatic Institute Product Management Framework (Pragmatic) occupies a distinctive position in the contemporary product management discipline.

Unlike delivery-centric methodologies, Pragmatic does not define product management by ceremonies, artifacts, or roles. Instead, it defines product management by accountability across the full lifecycle of value creation, from problem discovery (left) through sustained market or organizational impact (right).

The Persistent Misinterpretation: “Too Commercial for Internal IT”

Pragmatic is frequently rejected by leaders in internal IT and digital organizations because the framework is perceived as “too commercial”. This objection conflates language with intent.

Pragmatic’s underlying philosophy is not revenue-centric; it is economics-centric.

Internal products do not escape economics simply because revenue is indirect or absent. They still require investment decisions, tradeoffs, prioritization, and lifecycle accountability.

What differs is not whether economics apply, but how value must be demonstrated.

Reframing Pragmatic for Internal Product Contexts

When translated appropriately, the Pragmatic framework aligns naturally with internal product environments. Upstream Pragmatic activities map cleanly to internal contexts when language is adjusted:

• “Market” becomes population served or clinical domain.

• “Customer” becomes user, clinician, operator, or stakeholder.

• “Revenue potential” becomes patient outcome impact, cost avoidance, risk reduction, throughput improvement, or quality gain.

Crucially, Pragmatic insists that these questions be answered before the factory is activated.

Failure Mode 1: Upstream Initiation Without Legitimacy

We now turn to the first primary failure mode: upstream initiation without legitimacy.

This section examines how internal organizations routinely activate the factory without sufficiently validating whether a problem is real, material, and worth solving.

Upstream work is not primarily creative. It is discriminatory. Its purpose is to separate problems that are real, material, and worth solving from those that are merely plausible, fashionable, or politically convenient.

The Illusion of Demand and Political Sponsorship

In internal environments, demand is often inferred from requests, escalations, or executive interest, rather than validated through systematic inquiry. A common upstream error is to treat problem articulation as proof of problem legitimacy.

One of the most pernicious upstream failure patterns in internal IT is the substitution of political sponsorship for product judgment.

When a senior leader champions an initiative, questioning the underlying problem can be perceived as resistance rather than responsibility.

In such environments, upstream product management becomes performative, conducted to justify a predetermined decision rather than to interrogate its necessity.

The “Mandate Lane” and High-Risk Bypasses

The political reality of large health systems often means that some initiatives will inevitably bypass the full scrutiny of the validation gate due to a “strategic imperative” or the sheer political capital of a senior stakeholder.

The danger is not that these mandates exist, but that the risk is unacknowledged. If an initiative forces its way into the factory without upstream validation, the organization must explicitly log this as a “High-Risk Mandate.”

The logic is simple: By skipping the Upstream Safety Filter, the stakeholder is trading certain execution cost for uncertain value. The Sponsor, not the Product Team, must therefore retain accountability for the resulting adoption gaps.

The “Default to Build” Trap: Escaping the “You Hum It, We Code It” Mindset

A critical component of Upstream Judgment is the rigorous evaluation of how a validated problem should be solved. In mature product organizations, this is the Build/Buy/Partner decision matrix.

However, in many internal IT and digital organizations, this matrix has collapsed into a single default reflex: Build!

This phenomenon—often colloquially described as “You hum it, we code it”—stems from the Factory mindset. When an organization measures itself by the utilization of its engineering capacity, it subconsciously rejects “Buying” or “Partnering”, because those paths do not feed the Factory. They are viewed as “lesser” IT work, even when they will often offer faster time-to-value, lower maintenance burdens, and higher reliability.

The Product Manager’s Role as Solution Architect
Upstream responsibility requires decoupling the problem from the code. A Product Manager’s job is not to fill the engineering backlog; it is to solve the business problem.

• Buy (SaaS/COTS): When the problem is standard (e.g., HR, ERP, CRM) and speed is paramount.

• Partner: When a specialized third party can deliver the outcome with shared risk.

• Build: Reserved only for problems that provide a unique competitive advantage or where no market solution exists.

When this discipline is missing, internal organizations build expensive, custom versions of commodities, creating massive long-term technical debt for problems that could have been solved with a subscription.

How Vibe Coding Exacerbates the “Build” Default

The danger of the “Default to Build” mindset is currently being amplified by the rise of AI-powered “Vibe Coding.”

In the past, the barrier to “building” was high: it required budget, architects, and developers. This friction sometimes forced a “Buy” decision. Today, AI tools have lowered the cost of initial code generation to near zero.

• The Accelerator: A domain expert (doctor, researcher) who might have previously requested a vendor solution now feels empowered to “just build it myself” using AI.

• The Consequence: This democratizes and distributes widely the “You hum it, we code it” fallacy. The Vibe Coder skips the market scan entirely. They assume that because they can generate the code, they should.

This results in a proliferation of “Micro-Builds”—custom, single-use applications that ignore established market platforms. The organization moves from “Default to Build” (at the enterprise level) to “Default to Code” (at the individual level), bypassing the Build/Buy/Partner analysis entirely and accelerating the accumulation of unmanaged shadow assets.

While often framed as innovation, this dynamic creates a Shadow Factory. Because the cost of “building” has dropped to near zero for these individuals, the natural economic friction that usually stops bad ideas vanishes.

• The Trap: Organizations mistake a functioning script for a production-ready product.

• The Reality: These tools escape the Upstream Safety Filter and the Factory’s architectural controls. They arrive downstream as technical debt disguised as innovation—unsecured, unintegrated, and unscalable.

To correct this, organizations must reframe Citizen Development not as “Product Delivery,” but as “Upstream Discovery and Validation.”

A user-built tool is not a product; it is a high-fidelity requirement. It serves as robust evidence of a problem, but it must still pass the Validation Gate before the central organization accepts the liability of supporting it.

The Missing Question…“What Happens If We Do Nothing?”

Practical upstream judgment requires explicitly asking: What happens if we do nothing?

If the answer is “inconvenience” or “frustration,” further scrutiny is required. Not all inefficiencies warrant product intervention⁷⁴. Internal product teams often fail to articulate this counterfactual because they are structurally incentivized to build. Capacity seeks utilization.

Failure Mode 2: Downstream Delivery Without Value Realization

The False Equivalence of Delivery and Success

In internal product organizations, delivery is often treated as the terminal event of product responsibility. Once a solution is deployed, documented, and operationally supported, the organization shifts its attention elsewhere.

This behavior reflects a deeply ingrained but flawed assumption: that implementation is synonymous with success. From a technical standpoint, delivery confirms only that a system can function as designed. It says nothing about whether the system is used, whether it changes behavior, or whether it produces the outcomes and value that initially justified its creation.

Implementation is a Technical Event; Customer Success is a Social One

A central argument of this paper is that customer success, whether internal or external, is fundamentally a human and organizational phenomenon, not a technical one.

Implementation answers questions such as: Does the system work? Is it secure? Is it compliant?

Customer success answers very different questions: Do people trust it? Does it fit their workflow? Does it reduce friction or add to it? Is value, however measured, achieved?

The Absence of Ownership for Value Realization

Perhaps the most consequential downstream failure is the absence of clear ownership for value realization. Delivery teams own build. Operations teams own uptime. Training teams own onboarding. But no single role or function owns the question: Is this product delivering the value we expected, and if not, what will we do about it?.

This accountability gap allows downstream failure to persist indefinitely. Products remain in production because no one has both the authority and the mandate to declare them unsuccessful.

Why Customer Success Exists — and Why Healthcare Is Not Exempt

Why Commercial Organizations Created Customer Success

Customer Success did not emerge as a customer-experience enhancement or a rebranding of support. It emerged as a control function in organizations that discovered a structural truth: Delivery does not guarantee value.

In modern commercial models, organizations have learned, often painfully, that:

• Selling a product does not ensure adoption.

• Shipping a product does not ensure outcomes.

• Training users does not ensure behavior change.

As revenue shifted from one-time transactions to recurring, outcome-dependent models, organizations faced a new risk: silent failure after delivery. Customer Success was created to close this gap by owning adoption beyond launch, proving value, and protecting the economic return on product investment by eliminating attrition and churn.

Why the Same Logic Applies to Health Systems

Health systems often reject Customer Success because they lack subscription revenue or explicit churn. But this difference is superficial. Health systems also invest heavily before value is proven and depend on sustained adoption to realize the benefit.

The absence of revenue does not remove the risk. It just removes the signal. In healthcare, value erosion does not appear as revenue churn. It appears as workarounds, partial feature adoption, parallel systems, and user abandonment.

The Core Insight

Customer Success exists because value is realized over time, not at delivery. Health systems need it because they routinely invest in products whose success depends on sustained human behavior change, and currently have no function that owns that risk.

Defining Customer Success in Healthcare

Customer Success in healthcare does not operate in a conventional commercial environment. Instead, it exists at the intersection of clinical behavior, operational workflows, institutional risk, and cultural trust.

Healthcare Customer Success can be defined as:

The ongoing responsibility to ensure that a delivered digital or clinical product is adopted in practice, integrated into real workflows, trusted by its users, and demonstrably contributing to the clinical or operational outcomes that justified its creation.

Outcome Realization and “Worth the Squeeze”

Healthcare Customer Success should exist to answer a single, uncomfortable question:

Was this product worth the total cost of discovering, building, deploying, operating, and sustaining it?.

In the absence of revenue, internal products require a different—but no less rigorous—test of success. This value may be indirect or qualitative, but it cannot be purely aspirational.

Conclusion: Reclaiming Product Judgment Beyond the Factory

This paper has argued that the most persistent failures of internal product organizations are not failures of execution. They are failures of judgment.

The factory runs efficiently—products ship.

And yet value either fails to arrive or quietly dissipates as user attrition occurs.

Reframing Product Management’s Highest-Order Responsibility

The fundamental responsibility of product management is to decide if and whether the factory should run, and to remain accountable until value is derived or outcomes are measurably demonstrated. This responsibility exists regardless of whether products are commercial or internal.

What Changes If This Thesis Is Taken Seriously

If the arguments of this paper are accepted, several practical shifts follow:

1. Initiation becomes a decision gate, not a formality: Product teams are empowered—and expected—to stop weak initiatives before incurring costs.

2. Delivery is re-centered as a means, not an end: Factory excellence remains critical, but no longer substitutes for product life cycle judgment.

3. Adoption and behavior change become product work: Not training artifacts or support afterthoughts, but integral design concerns.

4. Value realization must be owned, measured, and revisited: Even when imperfect, even if qualitative, and especially if uncomfortable.

The factory will always be loud. Judgment must be made louder.

SM

Previously I published an article as an introducer to the concept of Medical Digital Twin technology and it’s application of contemporary AI technologies. You can find that article HERE.

This Strategic Insight article of current state Medical Digital Twin technology argues to assess Digital Twin use cases using a maturity scale grounded in contemporary Digital Twin standards.

It discusses the topic from a particular viewpoint. For further critique and analysis of this article and discussion of further work needed, please check out this Complimentary article published HERE.

I’ll also be creating a ‘Dummy’s Guide to Medical Digital Twin’ series of shorter articles for those looking to understand the application of the technology.

Introduction

In modern health care, we are overwhelmed with data—but too often lack the insights to act on it. Electronic health records, medical images, genetic profiles, and continuous vital signs from wearables all contribute to an overwhelming volume of patient information. The concept of the AI “medical digital twin” has emerged as a compelling answer to this data-action gap.

A medical digital twin is generally understood as a virtual representation of an individual patient, continuously fed by that patient’s real-world health data and capable of modeling their physiology to inform clinical decisions.

The idea of a digital replica of a physical asset originated in engineering: NASA pioneered early digital twin concepts during the Apollo era to mirror spacecraft systems, and Dr. Michael Grieves formally described a “digital twin” framework for industry in 2002. The term itself was coined by NASA in 2010 in the context of aerospace engineering, but has since taken on new life in medicine.

Healthcare digital twins remain a nascent but rapidly evolving technology. As a field, medical digital twin research is still emerging – yet its growth is remarkable. By the mid-2020s, researchers had already produced multiple literature reviews on the topic, signaling a “growing maturity” of the field despite its relatively recent origins. In practice, however, true digital twins of human patients are still more promise than commonplace.

According to the U.S. National Academies of Sciences, Engineering, and Medicine (NASEM), to qualify as a ‘genuine’ medical digital twin, a system must be personalized to the individual, dynamically updated with real-time data, and possess predictive capabilities to inform clinical decision-making.

A link to their report and assets is included in the callout below.

The NASEM Definition of a Digital Twin

NASEM Definition:
"A digital twin is a set of virtual information constructs that mimics the structure, context, and behavior of a natural, engineered, or social system (or system-of-systems), is dynamically updated with data from its physical twin, has a predictive capability, and informs decisions that realize value. The bidirectional interaction between the virtual and the physical is central to the digital twin."

The NASEM report however presents several key challenges for digital twins:

1) “Fit for purpose” problems to address,

2) Verification, Validation and Uncertainty Quantification (VVUQ) of dynamic, interacting virtual and physical assets,

3) Bi-directional, real-time assimilation of measurable data,

4) Interoperability of physical and virtual assets that can be sustained long-term through rapidly changing technologies,

5) Democratizing access of digital twins that are ethical and appropriately used.

By that strict definition, very few current projects qualify. A 2025 scoping review found that only 12% of surveyed studies claiming to use “digital twins” in health care actually met all of these criteria, with many efforts falling short on personalization, real-time data integration, or predictive power. In fact, many so-called digital twins today might better be described as static digital models or one-way data “shadows” rather than fully realized, interactive twins.

All of this underscores that while the ambition is high, the maturity level of digital twin technology in medicine remains relatively low overall – primarily due to the sheer complexity of living systems. As a U.S. Government Accountability Office (GAO) assessment noted, digital twin technology is “less mature in the biological sciences” (such as health care) than in other fields, such as engineering, mainly because human biology and systems are extraordinarily complex and variable.

Worth the chase

Yet the potential rewards are enormous. Proponents envision that mature medical digital twins could revolutionize clinical decision-making by enabling doctors – or even the AI twins themselves – to simulate and predict outcomes for individual patients with unprecedented precision.

A fully realized digital twin might forecast how a specific patient’s disease will progress, how they would respond to various medications or surgeries, and what complications could arise – all before any treatment is actually applied to the patient’s body. In essence, these systems promise to transform healthcare from reactive to proactive.

Instead of solely relying on trial-and-error or population-based guidelines, clinicians could explore treatment options in a risk-free virtual model of the patient, then bring data-driven confidence to real-world interventions. The journey to that level of capability, however, involves a stepwise climb in sophistication.

This article proposes a 4-stage ladder of medical digital twin maturity from basic data dashboards to fully autonomous agent-like systems. By taking this classification approach, we can better understand and be discerning about what is being described when the term ‘Medical Digital Twin’ is thrown around.

Buy me a coffee?

Level 1: Dashboard Stage

From Static Data to Living Model

In its simplest form, a medical digital twin starts as a digital doppelgänger – a faithful virtual copy of a patient’s data. Think of it as an interactive, 4D health record: not just a stack of static reports, but a coordinated, evolving representation of the person’s state. At this foundational stage, the twin is essentially a descriptive or informative model. It integrates data from various sources (lab results, imaging, sensor readings, genomics, etc.) into one cohesive picture, often visualized as a dashboard or 3D avatar of the patient.

This provides a “single source of truth” about the patient’s condition at any given moment. For clinicians struggling with fragmented information across charts and systems, such a unified digital mirror can be immensely valuable. It’s essentially an upgrade of the medical record into a living model of the patient. However, at this stage, the twins’ role is mostly passive – data-rich but insight-poor. It mirrors the patient, but does not yet deeply analyze or predict much beyond what an astute human clinician could infer from the data.

Many early examples of “digital twins” in medicine are in fact closer to this descriptive level. They might be better termed ‘digital shadows’ or static models rather than fully functional twins. The distinction matters. A true twin continuously syncs with the patient and can influence real-world decisions. In contrast, a ‘shadow’ only updates one-way (from patient to model), and a static model might be a one-off virtual construct that isn’t updated in real time.

For instance, consider a hospital that creates a detailed 3D model of a patient’s organ for surgical planning. If that model is based on the patient’s imaging data but does not update or interact after it’s created, it’s essentially a sophisticated digital model – a snapshot in silico – but not a dynamic twin.

Such tools are helpful for visualization and rehearsal (a surgeon can practice on a patient-specific model), yet they lack ongoing feedback or predictive simulation. Similarly, some “digital twin” pilot projects amount to aggregating a patient’s electronic health record with maybe a couple of monitoring streams into a unified interface. This is a step forward for data integration, but again, unless it leverages that data for modeling and prediction, it sits at the lower rungs of the maturity ladder.

A major 2025 analysis highlighted how far most implementations are from the ideal. Of 149 medical studies surveyed, only 18 systems (12%) met all NASEM criteria for a true digital twin. The majority were essentially data harmonization efforts or computational models missing one or more key elements (personalization, real-time updates, or predictive analytics). Tellingly, only two studies in that review mentioned having performed verification and validation of their twins’ predictions. In other words, not only are most current “twins” relatively simple, but the critical step of checking that a twin’s outputs match reality is often overlooked.

This gap between aspiration and reality defines the early stage of maturity: building a faithful data mirror is challenging enough, and ensuring its fidelity via rigorous validation is rarer still. Nevertheless, establishing this foundation is essential. You cannot have a smart, prescient twin without first aggregating an accurate, comprehensive dataset on the patient. The aphorism “garbage in, garbage out” holds: a digital twin is only as good as the data streams feeding it. Thus, the initial phase on the path from data to agency is about getting the data right and creating a robust virtual analog of the patient – a platform upon which higher-level intelligence can be built.

Level 2: Model Stage

Adding Dynamics, Simulation and Scenario Testing

The next leap in maturity is when a digital twin becomes predictive of how a patient will behave under various conditions, not just descriptive of their current behavior. Here we move into the realm of dynamic simulation. The twin stops being a passive reflection and starts serving as an active testbed for hypothesis and insight. In practical terms, this means incorporating computational models of physiology and disease progression that can run forward in time or under altered parameters. A twin with these capabilities can answer “what if?” questions.

What if we titrate this patient’s blood pressure medication upward? What if the patient’s tumor continues to grow at the current rate for six months? What if we virtually “knock out” a particular gene to see its effect on metabolism? Such simulations transform the twin into a kind of virtual patient that clinicians and researchers can safely experiment on. It is in this stage that the oft-cited vision of an “in silico clinical trial” comes to life: instead of exposing a real patient to potential harm, trial a therapy on their digital twin first.

Some of the most exciting medical digital twin projects to date live in this predictive simulation space. For example, the FDA has embraced the idea of using high-fidelity virtual models to evaluate medical devices and drugs before they ever touch a human patient. The FDA’s 2021 strategic roadmap explicitly highlighted digital twins’ cost-saving potential for virtual testing of medical devices, enabling issues to be identified in silico rather than in vivo.

Likewise, in Europe, ambitious initiatives are underway to leverage simulation-heavy twins for personalized medicine. One European Union-funded project is developing digital twins of patient brains to simulate the brain’s electromagnetic activity and physiology, helping treat neurological diseases such as Alzheimer’s and epilepsy.4 The vision is that a clinician could virtually model how a specific patient’s brain might respond to a new therapy or an invasive procedure, tailoring treatment to the individual’s neural profile.

Early efforts in cardiology follow a similar script: researchers have created patient-specific heart models (so-called “virtual hearts”) that can run through different intervention scenarios – such as how placing a stent or administering a specific drug would affect cardiac function – thereby guiding cardiologists in decision-making. In oncology, projects such as the “virtual tumor board” concept use mathematical models to predict a patient's tumor response to various chemo regimens, helping oncologists choose the most promising option.

Crucially, these predictive twins draw on advances in computational modeling and machine learning. The twin becomes a living simulation environment. For instance, a digital twin of a diabetes patient could incorporate a mathematical model of glucose-insulin dynamics, enabling it to predict the patient’s blood sugar levels hours or days ahead under different insulin dosing plans or dietary inputs. Similarly, a twin of a COVID-19 patient might simulate the progression of infection in their lungs and the likely impact of early intervention vs. watchful waiting.

These use cases are not purely hypothetical; they are being actively explored in research settings. The National Institutes of Health and academic consortia have funded digital twin models for critical illnesses, like sepsis and traumatic injury, to forecast deterioration and test interventions virtually. Even in surgical contexts, so-called “digital twin surgical simulators” have been piloted, where a patient’s anatomy and physiology are modeled so that surgeons can practice and refine an operation on the twin – adjusting techniques and seeing projected outcomes – before performing the real procedure.

Despite these promising developments, it’s important to recognize the challenges that accompany this stage of twin maturity. High-quality, predictive simulation requires accurate underlying models of biology, which are difficult to build and personalize. Human physiology is enormously complex; simulating it involves everything from fluid dynamics (for blood flow) to electrical conduction (for neural and cardiac rhythms) to biochemical networks (for metabolism and drug interactions).

Many early digital twins make simplifying assumptions or focus on one organ system in isolation, which can limit predictive accuracy. There’s also the issue of computational cost – detailed simulations can be computationally intensive, sometimes requiring supercomputer-level resources or cloud computing power that may not be readily available in clinical environments. Interpreting simulation results adds another layer of complexity: predictions always carry uncertainty, and clinicians must understand the confidence and error margins of a twin’s forecasts.

Validation, again, looms large. A predictive twin must be validated against real outcomes to be trusted. For example, if a twin predicts a patient’s tumor will shrink by 50% on a specific drug, and in reality, a comparable patient’s tumor shrinks by only 10%, the model needs recalibration. Ensuring that digital twins reliably mimic real physiology over time is an active area of research. Encouragingly, when done well, the payoff can be significant. Cedars-Sinai Medical Center recently demonstrated this by creating a “Molecular Twin” platform for patients with pancreatic cancer.

By integrating multi-omics data (genomics, proteomics, etc.) and using machine-learning models, their twin accurately predicted disease survival for individual patients with pancreatic cancer – outperforming standard clinical risk markers.5 In their study published in Nature Cancer, the Cedars-Sinai team showed that the twins’ predictions of patient outcomes were more accurate than the existing FDA-approved biomarker (CA 19-9) used in pancreatic cancer prognostics.

This is a striking example of a predictive digital twin adding real value. By crunching through thousands of biological variables and learning an optimal prognostic model, the twin could identify which patients were likely to have better or worse outcomes after surgery, enabling doctors to personalize follow-up treatments accordingly. It’s a glimpse of the near future where for certain conditions, trusting an AI-driven twin’s prediction could become standard practice in making clinical decisions.

Level 3: True Digital Twin Stage

Toward Prescriptive and Personalized Care

As digital twins become better at predicting the future, the logical next step is to have them recommend actions – or even take action. This stage moves beyond forecasting “what will likely happen” into suggesting “what should be done about it.” In analytics maturity terms, we shift from predictive to prescriptive.

A prescriptive medical digital twin doesn’t just passively project outcomes; it actively searches for the best interventions to improve those outcomes. Essentially, the twin can function as a clinical decision support system tailored to the individual patient. For example, a cancer patient’s twin might run dozens of virtual treatment protocols (varying drug combinations, dosages, timing, etc.) and identify which approach is statistically most likely to shrink the tumor with minimal side effects.

A twin for a patient with congestive heart failure might continuously optimize the patient’s diuretic dosage or beta-blocker therapy by analyzing their real-time vitals and lab trends against a learned model of what dosage keeps them most stable. In critical care, a digital twin of an ICU patient could suggest ventilator setting adjustments by simulating lung function at different oxygen or pressure levels, helping respiratory therapists fine-tune support to improve oxygenation while minimizing injury.

At this prescriptive tier, the digital twin acts like a tireless, hyper-competent medical consigliere. It combs through the myriad of possible decisions – medication choices, lifestyle modifications, procedural options – and provides data-driven guidance on which choices could yield the best patient-specific results. The power of this is personalization: two patients with the same diagnosis might get entirely different recommendations from their respective twins because the system recognizes their prognoses and risk profiles differ.

In the traditional one-size-fits-all paradigm, both patients might have been treated the same way; with prescriptive twin support, care can be finely tuned to each individual’s predicted response. This is the fruition of the precision medicine promise, operationalized by AI. Rather than relying solely on physicians’ general knowledge and experience, the twin provides an evidence-based roadmap for that unique patient.

It’s important to stress, however, that at this stage the twins’ role is advisory. The agency – the authority to act – still firmly rests with human clinicians and, in many cases, the patients themselves. A prescriptive digital twin might tell us, “Patient X’s twin suggests that increasing the dose of Drug Y by 20% and adding moderate exercise would likely reduce their 5-year cardiac risk from 15% to 5%,” but it cannot enforce those changes on its own. The care team must interpret and implement the advice. In this sense, the twin becomes an intelligent assistant, enhancing human decision-making rather than replacing it.

Many in the medical AI community advocate for this “human-in-the-loop” approach as the safest and most acceptable use of AI in healthcare. Doctors, nurses, and patients are more likely to trust and adopt AI recommendations when they retain control and can use clinical judgment on whether to follow the twins’ suggestions. Indeed, early trials of AI decision support show that clinicians appreciate AI input as a second opinion or a safety net for things they might miss. Still, they bristle at the idea of ceding final decision power to an algorithm.

During this prescriptive phase, transparency becomes a key issue. It’s not enough for a digital twin to spit out a recommendation; ideally, it should also explain or justify it in understandable terms. If a diabetes management twin advises cutting the nightly insulin dose in half, the care team will want to know why – perhaps the twin detected a pattern of overnight hypoglycemia in the virtual simulations and found that a lower dose would maintain control without those dangerous dips.

Explainability builds trust. This is one reason many digital twin systems incorporate physiological models or rule-based components alongside machine learning: a purely black-box neural network twin might make great predictions but be unable to explain them, whereas a model grounded partly in known physiology can at least be interpreted within a familiar framework. As the twins’ suggestions start influencing genuine care, such considerations of accountability and clarity become ever more critical.

Before moving on, it’s worth noting that the progression from predictive to prescriptive is not a clean jump but a gradual transition. In some domains, digital twins are already blurring the line—for instance, oncology twins that not only predict outcomes but also rank possible therapies, or critical care twins that not only forecast patient instability but also alert clinicians to “try intervention A before B.” We are already seeing glimmers of prescriptive guidance in specialized systems. However, broad deployment of entirely prescriptive twins across medicine is still in its infancy.

Clinical workflows and regulatory frameworks are just beginning to adapt to this new source of guidance. In the coming years, we can expect more pilot programs in which an AI twin’s recommendations are presented in multidisciplinary team meetings (e.g., tumor boards for cancer or heart teams for cardiac patients) to support complex decision-making. With careful validation and demonstrated success, acceptance of prescriptive twins will grow, setting the stage for the final, most transformative leap: granting digital twins a degree of autonomy in carrying out actions.

Level 4: Agentic Twin stage

The Autonomous Twin (and the Centaur’s Dilemma)

At the top of the maturity ladder is the autonomous digital twin – a system that not only recommends, but can act on its own. This is the “agency” in our journey from data to agency: the point at which the digital twin transitions from tool to actor. An autonomous medical digital twin would be empowered to make certain decisions or adjustments in real time without always needing human approval. In theory, this could be the ultimate realization of personalized medicine and automated care.

Imagine a digital twin managing a chronic condition around the clock: a twin for a person with type 1 diabetes that continuously monitors glucose (via sensors), predicts upcoming blood sugar trends (via its personalized model), and autonomously directs an insulin pump to deliver precisely calibrated doses to keep the patient in range – all while the person sleeps. In fact, this example is not far-fetched; closed-loop “artificial pancreas” systems are already an early form of AI-driven autonomy in healthcare. We can view them as a rudimentary digital twin of the patient’s glucose metabolism, with limited agency to act (insulin dosing).

Another example: a hospital could employ a digital twin of its emergency department – an AI agent monitoring patient flow, resource utilization, and predicting surges. Such a twin might automatically reroute ambulances to other hospitals when it predicts overload, or reallocate staffing and beds in anticipation of a spike in patient volume. If implemented, the twin isn’t just modeling or advising on operations; it is directly steering certain decisions in real time to optimize outcomes.

These scenarios highlight the enticing benefits of autonomy: speed, efficiency, and responsiveness that surpass human capabilities. An autonomous twin can react within microseconds to a change (far faster than a human), operate 24/7 without fatigue, and integrate vast data streams to inform its actions. In complex, high-velocity environments like critical care or emergency management, such agility could literally save lives.

Proponents of agentic AI in healthcare point out that there are many areas where humans can’t process information or react quickly enough — but a well-designed AI agent can. For instance, detecting the earliest signs of sepsis and initiating protocol-driven treatment is something an autonomous twin might do minutes or hours before a busy clinical team would typically recognize the pattern, thereby improving patient outcomes.

Indeed, the Digital Twin Consortium (an industry group) has begun developing frameworks to standardize and accelerate “agentic AI” capabilities, laying the groundwork for intelligent digital systems that can operate with a degree of self-direction. Their recent work on an AI Agent Capabilities toolkit underscores that industries from manufacturing to healthcare are pushing toward digital twins that are not just simulations, but interactive agents in their own right.

For all its promise, this final stage of maturity comes with profound challenges – technical, ethical, and legal. Handing a portion of control to an AI twin raises the stakes on reliability and safety. If a predictive twin is occasionally wrong, it might mislead a doctor; if an autonomous twin is incorrect, it might directly harm a patient by making a poor decision. Therefore, the bar for trust in autonomous medical twins has to be extraordinarily high.

This means rigorous validation, fail-safes, and likely a constrained scope of autonomy (at least initially). We may be comfortable letting a twin autonomously manage insulin dosing within set limits. Still, we’re not about to let a twin decide to start chemotherapy or discharge a patient without human sign-off. There’s also the question of accountability: if an AI twin makes a medical error, who is responsible? The physician overseeing it? The hospital? The software developer? Regulations have yet to keep pace with these possibilities fully.

The FDA has begun drafting guidelines for “adaptive” and “autonomous” AI in medical devices, but clear policies governing autonomous clinical decision systems are still works in progress.

Then there are the ethical dimensions and what might be called the philosophy of care. Medicine is not just an engineering problem; it’s a human endeavor built on trust, empathy, and moral responsibility.

Even if an autonomous twin could make faster, more accurate calls 99% of the time, many patients (and clinicians) might balk at being “treated by a machine” with minimal human involvement. The fear of a cold, algorithmic healthcare devoid of human touch is real and must be addressed as we tiptoe into autonomous territory.

It’s one thing to use AI as a co-pilot, and quite another to hand it the controls. We must also consider biases – any bias in data or algorithms could be perpetuated or amplified by an autonomous system, with potentially serious consequences if not checked. GAO’s assessment of digital twin challenges points out that data quality issues or biased training data could lead a twin to make misleading or even discriminatory decisions. Public trust in these systems will depend on demonstrable fairness, transparency in decision-making, and robust oversight mechanisms.

The Centaur’s Dilemma

The tension between AI autonomy and human control in domains like healthcare has sometimes been dubbed “the Centaur’s Dilemma.” The term, drawing on the mythological half-human, half-horse creature, encapsulates the challenge of combining human judgment with artificial intelligence in an optimal way. Do we strive for independent AI? Or do we always yoke AI to human oversight?

See my article ‘Centaur v’s Cyborg’

In chess, for instance, the best performance has often come from human-AI teams – nicknamed “centaurs” – rather than AI alone, because human intuition and creative insight complement the brute-force analysis of algorithms. In medicine, a similar principle likely holds. An AI, no matter how advanced, lacks the lived experience, empathy, and holistic context that human clinicians bring. Conversely, the best human doctors cannot match a sophisticated AI in data processing, pattern recognition across millions of cases, or unwavering consistency. The Centaur’s Dilemma asks: how do we allocate tasks and decision-making between human and machine to get the best of both?

Legal scholar James E. Baker popularized the term in discussing national security AI, noting that as AI systems become more capable, there is a temptation to take humans “out of the loop” for speed or efficiency . But doing so forfeits the human values and accountability that are crucial, especially in life-and-death decisions.

In healthcare, the dilemma manifests in questions like: Should an AI be allowed to autonomously adjust life support settings, or must a human clinician review each change?

If an AI flags a likely cancer on an image, should it just quietly mark it for a radiologist’s attention, or could it eventually authorize a biopsy on its own?

Finding the centaur-like balance – AI and human working in concert – may yield the safest and most effective outcomes in the foreseeable future.

Many medical ethicists argue for a “human in command” approach: AI can drive or fly on autopilot, but a human pilot is always ready to take control if needed. That way, autonomy is granted conditionally, under watchful human supervision, until we gain near-unshakeable confidence in the AI’s judgment.

The Centaur’s Dilemma is essentially about control and trust: how much to give, how much to retain, and how to know the difference. Solving it will be key to realizing the full benefits of medical digital twins without undermining the human-centric nature of healthcare.

Conclusion: Charting a Responsible Path to Agency

From data to agency – that is the evolutionary arc that medical digital twins are traversing. We began with isolated islands of health data and a dream that connecting and computing on that data could yield actionable knowledge for each patient. Today, the earliest forms of that dream are in motion as integrated digital models (proto-twins) that serve as detailed references for the patient's state.

With advancing analytics, some of these models have grown into narrowly focused predictive simulators that preview what might come next in a patient’s health journey. A few are starting to close the loop by recommending personalized interventions. And on the horizon glimmers the possibility of AI-driven twins that might themselves enact changes in real time, ushering in a new paradigm of autonomous healthcare agents.

It’s a progression not just of technology, but of trust and collaboration between humans and machines. Each rung of the ladder raises new requirements: better data quality and interoperability at the descriptive stage, robust computational models and verification at the predictive stage, decision logic and explainability at the prescriptive stage, and ironclad safety and ethics at the autonomous stage.

Where are we now on this ladder?

Broadly speaking, the healthcare industry is straddling the line between predictive and prescriptive. In specific niches – say, personalized oncology or critical care monitoring – we’re already seeing digital twin prototypes flex their predictive muscles. In others, like chronic disease management, early prescriptive systems are being trialed (often with a human in the loop) to suggest treatment tweaks.

But fully autonomous agency remains more of a thought experiment than a reality in 2026. The consensus among experts is that we should proceed cautiously. The technical barriers are significant but surmountable with time: more comprehensive data collection (including social, behavioral, and environmental determinants of health) will enrich twins; improvements in AI algorithms and computing power will make simulations faster and more accurate; and techniques like federated learning and privacy-preserving analytics will help address data sharing concerns.

The more challenging part may be building the social license for these tools – convincing practitioners, patients, and regulators that digital twins are reliable partners in care. This will require continued validation in clinical trials and real-world settings to demonstrate tangible benefits, whether it’s improved outcomes, reduced costs, or ideally both.

Early signs are encouraging: even a brief GAO survey noted that digital twins “could help facilitate predictive and personalized medicine to improve patient outcomes and reduce some health care costs.” Those twin goals of better care and lower cost are the holy grail of healthcare innovation, and they are driving significant investment and interest in digital twin technology.

The harder work?

The road to mature medical digital twins will also demand updated frameworks in medical education, ethics, and law. Tomorrow’s clinicians may need training to work effectively with AI twin systems, just as pilots had to learn to work with autopilot systems. Guidelines will need to clarify how responsibility is shared when an AI is involved in care decisions. We will likely see new industry standards and perhaps certifications for clinical AI systems, including digital twins, to ensure a baseline of safety and efficacy. Interdisciplinary collaboration – among clinicians, data scientists, engineers, and ethicists – will be essential to navigate the remaining hurdles.

Ultimately, the pursuit of giving data “agency” in medicine is about amplifying human capabilities, not diminishing them. A mature medical digital twin should function as a guardian angel of sorts: ever-vigilant, knowledgeable, and empowered to act (or assist in action) in the patient’s best interest. Achieving that in reality will take longer than early hype cycles might have suggested, but step by step, the gap is closing. As we stand today, we see the outlines of what’s possible through pioneering projects and pilot studies.

Closing

Over the next decade, as more data becomes harnessed and more trust is built, medical digital twins are poised to move from lab curiosities and isolated successes to mainstream clinical tools. When they do, healthcare can transform – shifting from reactive care to a model of continuous, proactive health partnership powered by our digital counterparts.

The journey from data to agency is far from over, but its destination holds the promise of safer, more innovative, and more personalized care for all.

Reference Notes

1. Mickaël Ringeval et al., “Advancing Health Care With Digital Twins: Meta-Review of Applications and Implementation Challenges,” Journal of Medical Internet Research 27 (2025): e69544. The authors note that despite its emergent status, the field has produced multiple literature reviews, indicating a growing maturity in this research area.

2. U.S. Government Accountability Office (GAO), Science & Tech Spotlight: Digital Twins—Virtual Models of People and Objects, GAO-23-106453 (February 14, 2023), 1. The GAO report observes that digital twin technology is “less mature in the biological sciences, primarily due to the complexities of living systems” in health care and related fields.

3. GAO, Science & Tech Spotlight: Digital Twins, 1. The FDA’s strategic plan highlights the cost-saving potential of using digital twins to evaluate medical devices virtually prior to patient use.

4. GAO, Science & Tech Spotlight: Digital Twins, 1. The same GAO report notes an EU-funded personalized medicine project aiming to use digital twins to simulate patients’ brains for treating Alzheimer’s and epilepsy.

5. Cedars-Sinai Newsroom, “Cedars-Sinai Develops New Tools to Improve Pancreatic Cancer Patient Care,” press release, January 22, 2024. The press release summarizes a Nature Cancer study in which Cedars-Sinai’s “Molecular Twin” platform accurately predicted pancreatic cancer patient survival and outperformed the standard CA 19-9 biomarker.

6. Digital Twin Consortium, “Digital Twin Consortium Announces Next Phase of AI Agent Capabilities Periodic Table Framework,” press release, June 12, 2025. This announcement describes the development of an AI Agent Capabilities framework to accelerate “agentic AI” across industries, enabling implementation of intelligent, adaptive digital twin systems.

7. GAO, Science & Tech Spotlight: Digital Twins, 2. The report cautions that data quality issues or bias in digital twin models could mislead decision-making, noting that inaccurate training data might result in misleading or discriminatory analyses by the twin.

8. GAO, Science & Tech Spotlight: Digital Twins, 2. Under “Opportunities,” the report states that digital twins could help facilitate predictive and personalized medicine to improve patient outcomes and reduce some healthcare costs.

This article is a personal commentary following the publication of our Podcast episode featuring brilliant work on ‘The Abominable Creature’.

You can find the YouTube episode here

Haverin About Podcast YouTube

and the Apple Podcast episode here

Haverin About Apple Podcasts

TL:DR

The U.S. does not have socialized medicine.

It has socialized financing routed through the most inefficient, complex administrative machine ever built around healthcare.

Understanding that distinction—especially for people working in healthcare technology, health policy, and care delivery—is essential.

Because no matter which sector we operate in, this is the economic terrain on which every product, every policy, every innovation, and every clinical decision sits.

For decades, the debate about “socialized medicine” in the United States has been framed as a fight over whether government should play a bigger role in our healthcare system. But this framing leaves out a more uncomfortable reality: Americans already pay for the system—almost all of it—no matter which political label we attach to it.

A more honest question is NOT whether we want to pay for healthcare collectively.

It’s how much friction, inefficiency, and financial arbitrage we want in between the dollars we contribute and the care we receive.

And when you examine how money actually moves through the U.S. healthcare economy, one fact becomes unavoidable:

We have already socialized the cost.
We have privatized the inefficiency.
BUT, we have increasingly eliminated the consumer.

Let’s unpack this.

Defining “Socialized Medicine”

(and why the US political debate frames the term incorrectly)

In political and economic policy terms, socialized medicine describes a system where:

• The government owns the hospitals

• The government employs the clinicians

• Healthcare is paid for through taxes

• Patients receive care at little or no direct cost

And shocker!…by this definition, we already have two major U.S. systems that qualify as socialized medicine:

• The Veterans Health Administration (VA)

• The Military Health System / DoD hospitals

These operate much like the UK’s NHS, or the Nordic national systems. They are unquestionably socialized.

Yet in US political rhetoric, the term “socialized medicine” gets flung at almost any policy involving public dollars, from expanding Medicaid to implementing a public option to proposing Medicare for All.

This rhetorical overload obscures the system's actual structure and hides the real inefficiencies we all pay for.

The Funding Reality: Americans Already Pay for Most of the System Through Taxes

When economists analyze national health expenditures, a surprising pattern emerges:

A majority of U.S. healthcare spending already originates from taxes—federal, state, and local.

This includes:

• Medicare payroll taxes

• Medicaid and CHIP funding

• ACA subsidies

• VA and DoD healthcare budgets

• Public health departments

• IHS and federal rural health programs

• Tax expenditures for employer-sponsored insurance (a massive federal subsidy)

• General revenue that supports Medicare Parts B and D

Add all this together, and you discover something that disrupts the entire political narrative:

A huge share of U.S. healthcare spending is already publicly financed—even though most care is privately delivered.

Leave a comment

Premiums, Deductibles, Employer Contributions—It’s all just public money in disguise

Economists consider employer-sponsored insurance to be quasi-public for three reasons:

A. Premiums replace wages

Employers don’t pay premiums out of generosity.
They pay them instead of paying more to you!

B. Employer premiums are tax-exempt

SHOCKER! This tax exclusion is a federal subsidy that costs more than the mortgage-interest deduction.

C. Deductibles and Out-of-pocket spending is the most regressive tax in America

Deductibles and copays disproportionately burden low- and middle-income Americans and those with chronic illness. Oftentimes the plans that lower-income Americans have available and choose have some of the highest deductibles and OOP limits in the country.

HOT TAKE! The ACA Marketplace exacerbated that inequity

Support our publishing by buying some cool merch from the Haverin About store.

Haverin About STORE

When you combine:

• Taxes

• Wage offsets through employee benefits premiums

• Employer benefit contributions (offset by) Employer tax subsidies

• Deductibles and out-of-pocket spending

…you arrive at this reality:

The U.S. healthcare system is already collectively financed. We just finance it in the least efficient way possible.

The Insurance Revenue Cycle Adds Massive Friction, Waste, and Arbitrage

Once the dollars enter the system, they pass through a labyrinth of revenue-cycle intermediaries:

• Private and public insurers

• Claims adjudication systems

• Prior authorization

• PBMs and rebate structures

• Coding and billing departments

• Clearinghouses

• Utilization managers

• Third-party administrators

• Hospital revenue cycle vendors

• Compliance and credentialing workflows

This all creates enormous economic friction, made 10 times worse by the regulations, government schemes, and subsidies introduced to try to adjust for the mess.

Thanks to the legislative and regulatory tinkering by our state and federal politicians, the Road to economic hell truly is paved with good intentions.

If you tried to design the most administratively expensive healthcare system in the world, you couldn’t do a better job than the United States of America.

Administrative waste estimates

• $265 billion to $500+ billion annually

• 2–5× the administrative burden of other high-income nations

This isn’t bureaucracy by accident.
It is bureaucracy by economic design.

Insurers, PBMs, hospitals, and vendors profit by managing—with increasing complexity—the gaps between what is charged, what is allowed, and what is actually paid.

This is not a care-delivery ecosystem.
It is a financial ecosystem built around delivering care.

The Sharpest Irony: Americans Pay for the System Yet Have Almost Zero Agency

Despite being the ultimate funders, individuals have very little control over:

• Which clinicians they can see

• Which hospitals are in network

• When or whether appointments are available

• Whether specialists will be authorized

• What anything will cost

• What their out-of-pocket burden will be

• How long claims take to adjudicate

• Whether they can get needed medications filled

• Whether bills are accurate

In a functional market:

• buyers have choice

• pricing is transparent

• competition improves service

• consumers have agency

In U.S. healthcare, none of these conditions exist.

Americans finance the system like taxpayers, experience it like disempowered participants, and pay for it like customers—yet enjoy the rights of none of those roles.

So What Do We Call This?

Not socialized medicine.

Not a free-market healthcare.

Not a public system.

Not a private system.

What we actually have is this:

A publicly funded…privately delivered…administratively bloated…economically opaque healthcare economy…

where the costs are socialized…the profits are privatized…and consumer agency is almost nonexistent.

This is the Gordian Knot our political system refuses to address.

It is far easier to shout “socialism!” and defend the indefensible than to understand or explain how the system truly works and why it is so inefficient.

But if we want the future of U.S. healthcare to be more rational (economically, clinically, and operationally) we have to confront the underlying truth:

• We are already paying for the whole thing.

• The only question is how much we want to waste in the process.

Please leave your comments to share your thoughts and opinion. Do you have a view of the abominable creature?

SM

Leaders aren’t short on intelligence; they’re short on uninterrupted cognition. Back-to-back video meetings create constant context switching and shallow participation. The result: less coaching, less time for deep work, and a culture of “update theater” rather than leadership. Worse, many leaders try to “multitask” their way through it—dividing their attention during calls and degrading both outcomes.

Below is:

• a summary of the Science on why this is an issue;

• a 12 point Action Audit you can use to scope your organizations problem;

• 5 Operating Rules to enable your leaders to lead and your teams to effectively get sh*t done!

Buy me a coffee

Why this is an issue (the science, briefly)

For each one of these the link takes you to the deep science.

• Task switching is costly. Switching between tasks imposes measurable time and error penalties; doing it repeatedly degrades performance. Executive control of cognitive processes in task switching

• Attention “residue” lingers. After moving from one task/meeting to another, part of your attention remains on the prior work, suppressing your current performance. Challenge of attention residue

• Heavy media multitaskers show poorer filtering and working memory control. This undermines critical thinking in meetings. Cognitive control in media multitaskers

• Videoconference fatigue is real and mechanistic. “Nonverbal overload” (constant close-up faces, reduced mobility, sustained eye contact) increases fatigue; validated scales exist to measure it. Causes of Zoom Fatigue

12 Metrics that matter: How do I tell if I really have this problem?

Below is a 12-point, structured, data-backed action plan to audit the reality within your org and diagnose the problem:

1. Total meeting load per leader

What to measure: Total meeting load per leader (hours/week; % of workweek)
How to capture it: Viva Insights “Meeting habits” + calendar export; segment by role level.
Why it matters: Establish baseline capacity for high value work; >60% in meetings kills time for coaching and strategy.
Owner & cadence: PMO or People Analytics, monthly
Guardrails / targets: ≤50% of workweek in meetings for VP−; ≤60% for C-suite
Fast remediation fixes: Reduce default meeting lengths to 25/50 min; consolidate/eliminate recurring series quarterly.

Continued below for paid subscribers:

Read more