“You know you’re writing the first line of your professional obituary, right?”

Our chairman said it the day the board approved the replatforming of our core product. He wasn’t being cruel (he was, if anything, being kind, giving me one last exit). He was doing the math out loud that everyone in the room was doing silently. Rebuild the platform your entire company runs on, the platform your customers depend on every day, and if it goes sideways that sentence writes itself. The alternative was obvious and available: keep the old platform running, patch what hurt, ship features around the edges, and let the thing limp forward for a few more years. Nobody gets fired for keeping the lights on.

That instinct — keep the lights on, don’t touch the load-bearing wall — has a name, and it turns out to be one of the most studied biases in decision science. William Samuelson and Richard Zeckhauser called it status quo bias in their 1988 paper in the Journal of Risk and Uncertainty, and once you see it you cannot stop seeing it in every roadmap review you’ve ever sat through.

The gravity of standing still

Samuelson and Zeckhauser ran a series of experiments where they handed people the same decision with one thing changed: whether an option was framed as the thing they already had. When an option was the incumbent, people chose it far more often than when the identical option was just one choice among several. The pull wasn’t analysis. It was the plain fact of already being there.

They didn’t stop at the lab. They looked at real decisions with real money (faculty members choosing health plans and retirement programs) and found the same thing at scale: people stuck with what they’d picked before, long after better options existed, in choices that mattered enormously for their futures. One of my favorite details in the paper is a small one, a colleague of theirs who ordered the same lunch, at the same place, for 26 years. Standing still isn’t a decision you make once. It’s a decision you quietly re-make every single day you don’t move.

That’s the part that should worry any CTO. Keeping the old platform alive felt like the neutral choice, the no-decision decision. But there is no neutral. Choosing not to replatform was a choice, with compounding costs: technical debt, a widening gap between what customers wanted and what the architecture could bear, good engineers spending their best hours fighting the past. The status quo just doesn’t send you an invoice. It bills you quietly, and you don’t notice until the platform can’t do the one thing the market has started demanding.

We tried the safe version first

Here’s the part I’m less proud of. We didn’t go straight to the big bet. We tried the gradual version first: strangle the old system piece by piece, stand up new services alongside it, migrate quietly in the background, never make anyone nervous. On paper it’s the responsible path (lower risk, reversible, no obituary line required).

It failed. Not dramatically, just slowly. The systems had to talk to each other constantly, the seams multiplied, and every new piece we built had to stay compatible with the old world we were trying to escape. We were spending real money to make the legacy platform more entrenched while telling ourselves we were replacing it. That’s when the honest question showed up: was incrementalism the wise choice here, or was it status quo bias wearing a nicer outfit?

Because incrementalism can be either one. Sometimes gradual really is right, when the risk is lower and the reversibility is worth paying for. And sometimes “let’s do it in phases” is the socially acceptable way to never do it at all. Barry Staw’s research on escalation of commitment describes the close cousin of this problem, the way we keep pouring resources into a course of action precisely because we’ve already poured resources into it. Our incremental attempt had quietly become a thing we were protecting rather than a path to somewhere new.

Telling wisdom from fear

So how do you separate the two in the moment, when both look identical from the inside? The test I landed on is uncomfortable but useful: ask whether the incremental plan has a credible endpoint where the old thing is actually gone. Not smaller. Gone. If you can draw a straight, believable line from “phase one” to “the legacy system is decommissioned,” gradual is probably right. If every phase instead makes the old system a little more necessary, a little more load-bearing, you aren’t migrating. You’re renovating a house you swore you were going to leave.

The second test is about who the plan comforts. If the phased approach exists mostly because it lets everyone (you very much included) avoid a hard conversation, that isn’t risk management. That’s the bias doing what it does best, making the expensive choice feel like the prudent one.

We killed the incremental attempt and committed to the full replatform, the one with the obituary line attached to it. It was the higher-variance decision on paper. But the real career risk was never the bet. The real risk was the slow version, the one where nothing ever visibly fails, the platform falls a little further behind every quarter, and one day you look up and the market has moved somewhere your architecture can’t follow. Nobody gets fired for keeping the lights on. They just get passed by.

The chairman was right that I was writing the first line of something. He was just wrong about which decision was the dangerous one. If you want the organizational version of this same gravity, why companies keep deferring the projects they most need, I wrote about that a while back in The Rise of Change Fatigue.

I used to think the fix was more explanation. Slow down, add a glossary slide, spend ten more minutes on what a large language model actually does. That instinct turns out to be almost exactly backwards. The Deloitte Global Boardroom Program’s latest Governance of AI survey of 695 board directors and executives across 56 countries found that two-thirds still say their boards have “limited to no knowledge or experience” with AI. That’s an improvement (it was 79 percent a year earlier), but it means most boards you’re presenting to this year still aren’t fluent. And yet the same survey found only 8 percent of boards don’t engage with management on AI at all. The knowledge gap isn’t shutting the conversation down. It’s just changing what the conversation needs to be about.

The Fluency Myth

Closing a knowledge gap with more education is a sound instinct in almost every other part of the business. It’s the wrong instinct here. Deloitte’s survey found that even as board AI literacy inches up, 31 percent of boards still say AI isn’t on their agenda at all (down sharply from 45 percent a year earlier), and a third of respondents say their boards still aren’t spending enough time on the topic. Boards aren’t asking for a seminar. They’re asking for something they can act on despite not being AI-literate, and that’s a very different assignment than the one most technical leaders prepare for.

This is particularly relevant for a CTO, because it reframes the job. Teaching the board how a model works is optional. Translating what it’s worth to the business is not. Those are two different skills, and most technical leaders default to the one they’re better at (mine, obviously, is the one my board never asked for). The gap between “I can explain this” and “I can make the board comfortable approving this” is where most AI proposals actually stall.

The Ledger Beats the Blueprint

PwC’s Governance Insights Center found something adjacent in its own research on boardroom communication: 74 percent of directors say they want more time spent on strategy, not less, even as total board time stays flat or shrinks relative to everything competing for it. Their advice for executives boils down to a habit most of us break the moment we get nervous: name the headline before the supporting cast arrives, and let every fact in the room work in service of that one line rather than compete with it for attention.

I’ve started building every AI slide backwards from the single number I want the board to remember on the drive home. Not the token count, not the model name, not the roadmap. A dollar figure, a percentage, a date. Everything else in the deck exists to support that number, not to audition for its own applause. It’s a smaller deck than the one I used to build, and it lands better every time.

The Bar Is Higher in Healthcare

If a Deloitte-style knowledge gap is the industry norm, health system boards stack a second layer on top of it. A Black Book Research survey of 202 health system executives and 84 board members found that 94 percent of respondents say their organizations are prioritizing technology investments precisely because of their near-term, easy-to-measure payoff: lighter clinical burden, faster workflows, cleaner documentation. Healthcare boards aren’t skeptical of AI’s long-term promise. They’re skeptical of vague ones.

That tracks with what I see in my own boardroom. Pushback on an AI proposal rarely comes because the technology is unproven; it comes when the value is unproven on the board’s timeline. In a regulated, patient-facing business, “trust me, it pays off eventually” is a much harder sell than it would be in a typical SaaS pitch, and a HIPAA-covered board has every right to ask harder questions than a typical SaaS board would. The ask has to show up with its receipts already attached, not a promise to send them later.

The Translation Layer

The practical version of all this is less a slide template than a discipline: lead with the outcome in the board’s own currency (revenue, cost, risk-adjusted timeline), frame governance as the reason to trust the number rather than as a caveat that undercuts it, and leave exactly one ask on the table instead of three.

Deloitte’s survey also found that boards already pull CFOs into AI conversations 27 percent of the time, more than double the rate for CISOs or CROs. That’s a tell worth acting on. Your board already wants the finance lens in the room before you show up. Beat them to it, and the technical story becomes the appendix it should have been all along.

The next AI deck I build starts with the number, not the architecture. Everything else has to earn its slide by supporting that number, or it doesn’t make the cut.

Alcázar-Peral and colleagues at the Quirónsalud Healthcare Network in Madrid ran a multicenter observational study of their ambient documentation tool, Scribe, across outpatient care from September 2024 to December 2025. Of 11.6 million outpatient visits in that window, 2.3 million were completed using the tool — an overall usage rate above 20 percent, with monthly adoption climbing from 2.7 percent at launch to a peak above 31 percent by the study’s end.

The Adoption Curve

That climb is the number most people will skim past. Digital health tools that plug into the electronic health record typically take years to reach a third of an organization’s visits, if they get there at all. Going from single digits to nearly a third of every outpatient visit in a little over a year puts ambient documentation in different company than almost any other EHR-adjacent technology of the last decade.

Fast adoption alone doesn’t prove a technology works — it proves clinicians didn’t hate it enough to stop using it, which is a lower bar than most vendor decks imply. What makes this study genuinely useful is that it didn’t stop at the adoption curve. It kept measuring transcription accuracy, consultation length, and clinician experience across the entire sixteen months, which is where the honest numbers show up.

Twenty-Two Seconds

Here’s the honest number. Average consultation length barely moved: 15.01 minutes without the tool versus 14.65 with it, with several months favoring Scribe before converging toward parity as workflows matured. Call it twenty seconds, and even that faded as clinicians settled into the tool. Transcription accuracy, measured as semantic agreement between the AI-generated note and the actual encounter, stayed remarkably stable the entire time (87 to 89 percent).

That’s a smaller time win than UCLA’s own trial found for its best-performing scribe, and this one comes from a real deployment ten times the size. The study’s authors offer a more interesting explanation than “the tool doesn’t save time”: their working hypothesis is that whatever minutes got reclaimed weren’t banked as shorter visits — they were reinvested into the visit itself, in patient interaction and the kind of care that doesn’t show up on a stopwatch. Clinician-experience surveys, run in two waves (184 clinicians, then 163), support that read: five of the seven measured domains improved meaningfully as the tool matured, including administrative workload and stress and quality of life, with gains holding steady across heavy and light users alike.

The Oversight Clause

That pattern lines up with the most rigorous controlled evidence available. UCLA Health ran the first randomized trial of ambient scribes in U.S. outpatient practice — 238 physicians, three arms, two months — and found a real but modest time savings for one tool (about 9.5 percent versus control), no significant change for the other, and improvements in burnout and workload scores for both. Different country, different design, same basic shape: genuine gains, smaller than the pitch, uneven across tools.

The UCLA trial also surfaced something a usage log can’t: occasional clinically significant inaccuracies in the AI-generated notes. Senior author John Mafi’s response to that finding is the line worth keeping: “This technology requires active physician oversight, not passive acceptance.”

That’s a governance requirement dressed up as a safety caveat, and it applies regardless of how smoothly a rollout scales. A tool that’s improving on average, at 2.3 million encounters, can still be wrong in one specific chart on one specific day — and the fix for that isn’t a better model. It’s a review step nobody’s tempted to skip.

From Capture to Insight

Zoom out further and the field’s own literature agrees on where this goes next. A recent narrative review of ambient scribe research — Razaghi and colleagues, published this past January — found the same recurring limitations across eighteen studies: omission errors, inconsistent performance, and what they call “note bloat,” notes that get longer without getting more useful. Their conclusion is that today’s scribes, however well adopted, are still fundamentally passive. They wait for a clinician-led encounter, transcribe it, and stop.

That’s the distinction I keep coming back to in my own work. A scribe that transcribes reliably at scale is a genuinely useful tool. A system that turns the same encounter into a structured, reasoned signal (flagging what needs follow-up, catching the pattern a tired physician might miss at 4 p.m. on a Friday) is a different category of thing. The first is documentation done well. The second is documentation that starts pulling weight the note never could.

Two million consultations is real proof that clinicians will adopt this technology, and keep using it. It’s not proof that the technology has moved past capturing what happened toward understanding what it means. That gap is still the frontier, and it’s the one worth watching.

A study out of PLOS Digital Health this March says that assumption doesn’t hold up. Chen, Li, and Lin systematically reviewed 171 studies of medical AI models that had already cleared the bar most of us treat as the finish line — approved for, or actively deployed in, real clinical settings, not research prototypes. Across more than 200,000 real patients, the advantage these “mature” models hold over human clinicians turns out to depend almost entirely on whether the patient in front of the model looks like the patients it was built on.

Home Field, By The Numbers

Across all 171 studies, AI models beat human practitioners on accuracy, 81.7 percent to 77.8 percent. That’s the number every vendor deck leads with, and it’s real. But Chen and colleagues split the studies by geography: evaluations that happened in the same country or city where the model was developed (”in-distribution”) versus everywhere else (”out-of-distribution”). In-distribution, the AI advantage actually widens, to 82.9 percent versus 77.3 percent for clinicians. Out-of-distribution (cross-border or cross-city validation), it evaporates. AI scored 74.1 percent against 76.3 percent for humans, a gap so small the p-value came in at .45. Statistically, no gap at all.

This is particularly relevant because these weren’t lab prototypes with an asterisk next to their claims. Every model in this dataset had already cleared regulatory approval or active clinical deployment, the checkpoint we generally assume catches exactly this kind of problem. It didn’t.

The honest way to describe the headline number, then, isn’t “AI beats doctors.” It’s “AI beats doctors on its home field.” Move the game somewhere else and the scoreboard resets to even.

Who Gets the Home Field

The geography of that home field is not evenly distributed. Geographic concentration across the 171 studies was extreme (a Dagum-Gini coefficient of 0.97), with 95.1 percent of patient cohorts coming from high-income (62.2 percent) or upper-middle-income (32.9 percent) countries — overwhelmingly China (28.7 percent) and the United States (18.9 percent). Not one study in the dataset involved a low-income country. Racial composition followed the same pattern: White (49.1 percent) and Asian (42.6 percent) patients made up most of the cohort, and nearly two-thirds of the studies showed substantial sex imbalance in who the model was tested on.

None of this started at the deployment stage. A 2024 review in npj Digital Medicine, looking one step upstream at the clinical-trial pipeline itself, found the identical skew: of 159 AI-enabled clinical studies, Yang and colleagues found 74 percent were conducted in high-income countries, with exactly one (in Mozambique) coming from a low-income country. Only 6.3 percent were international multicenter studies, the kind of design that would actually test a model against a population it wasn’t born into.

Put the two studies side by side and you get a full pipeline: the trials that build these models skip the populations that need them most, and the deployment data confirms the skip never gets corrected on the way out the door.

The Checkpoint That Isn’t

Here’s the part that should bother anyone buying AI for a health system rather than building it. “Regulatory-approved” and “already in clinical use” sound like assurances that the population question has been handled. Chen and colleagues’ data says those labels tell you almost nothing about how the model performs on a patient who doesn’t resemble its training set (the aggregate accuracy number vendors publish blends in-distribution and out-of-distribution performance into one reassuring figure).

I’ve sat through more vendor pitches than I can count where the accuracy slide is a single number with no population breakdown attached. After this study, that single number is the wrong artifact to evaluate a vendor on. The right question is closer to: show me your performance sliced by the population it was validated against, and show me what happens outside it.

That’s a different conversation than the one most procurement processes are built for. A security questionnaire tells you whether a vendor encrypts data at rest. A clinical validity questionnaire, as most of us currently write it, tells you an aggregate accuracy number and calls it done. Neither one asks the question this study says actually predicts whether the model will work for your patients: what population was this built on, and how far is my population from that one.

What This Means for Buyers, Not Just Builders

Practically, this changes what belongs in a vendor security and clinical-validity review before a contract gets signed. Training population characteristics and out-of-distribution performance data should sit next to SOC 2 reports and BAAs as baseline diligence items, not optional appendices a vendor provides if asked twice.

It also reframes what “bias in AI” means for people who spend their days procuring the stuff rather than training it. This isn’t an abstract fairness debate that lives in a data science paper. It’s an 8.8-point drop in the AI model’s own accuracy (from 82.9 to 74.1) the moment your patient population stops matching the one the model grew up on. Ask for that number before you ask for the roadmap.

1. Liberty Mutual preps legacy systems for AI scale: The insurer is reaping the rewards of a years-long modernization strategy to build systems that allow for as much flexibility as possible.

2. Why Employees Resist AI—and How Companies Can Win Them Over: Why do so many technology initiatives stall? Research by Das Narayandas and Shunyuan Zhang finds that employees often avoid tools like AI not because they don’t work, but because they threaten their expertise.

3. Accenture, Google bring enterprise AI to the midmarket: Midsize companies face familiar AI adoption setbacks but need services tailored to their scale, Accenture CEO Julie Sweet said.

4. CEOs fear they’re underinvesting in AI: More than half of chief executives are concerned their businesses will fall behind due to limitations in technology foundations, according to a new survey.

Ryan Pendell’s analysis of Gallup’s global workplace research gave me language for something I’d been circling for months without naming it. He calls it “the remote work paradox,” and once you see the shape of it, the remote-versus-office debate looks like it’s measuring the wrong variable entirely.

The Engagement Illusion

Pendell’s analysis sorts workers into four groups: fully remote, hybrid, on-site remote-capable, and on-site non-remote-capable. Globally, fully remote workers report the highest engagement of any group, at 31 percent, compared with 23 percent for hybrid workers, 23 percent for on-site employees whose jobs could technically be done remotely, and 19 percent for those whose jobs can’t.

Here’s the paradox: those same fully remote workers are less likely to say they’re thriving in their lives overall (36 percent) than hybrid workers or on-site remote-capable workers, both at 42 percent. The group posting the best engagement numbers is posting some of the worst wellbeing numbers. A leader who only watches an engagement dashboard is reading half the story and mistaking it for the whole one.

This is the trap in the remote-versus-office debate. Neither side of it predicts wellbeing cleanly, and neither side predicts engagement cleanly either. Location correlates with both outcomes, but it isn’t causing either one. Something underneath the location variable is doing the actual work.

Autonomy Is Not the Same as Connection

The engagement side of the paradox, Pendell argues, comes down to autonomy. Remote workers shape their own schedules, reach a flow state without interruption, and use their time more efficiently, and that shows up directly in how engaged they feel. But autonomy has a cost on the other side of the ledger: without shared structure, coordinating with a team you rarely see becomes its own source of stress, and digital collaboration isn’t always smooth, especially on work that requires heavy coordination.

This is where I think a lot of hybrid leadership goes wrong, including some of my own early instincts. We solve for autonomy (flexible hours, async-first communication, fewer standing meetings) and quietly assume that’s the same as solving for connection. It isn’t. Autonomy is about control over your own time. Connection is about whether anyone would notice, and care, if you disappeared for a week. A team can have abundant autonomy and almost no connection at all.

The leadership implication is uncomfortable: giving people more freedom doesn’t automatically make them feel more tied to anything. If coordination is where the friction actually lives, coordination is where the redesign has to happen, not by adding meetings, but by making handoffs, context, and checking-in deliberate instead of incidental.

The Camaraderie Nobody Puts on the Org Chart

The most useful point in Pendell’s piece isn’t about the calendar; it’s about what an office quietly hands you for free. Camaraderie. The running joke that started three weeks ago. The lunch nobody scheduled. Remote workers lose access to that scaffolding, and Pendell notes that isolation, left unaddressed, feeds loneliness and can tip into sadness and anger. He also points to something I hadn’t considered: sharing meals with other people predicts overall wellbeing about as strongly as income does.

That one reframed how I think about my own team’s rituals. We’ve built plenty of structure around delivery (sprint cadences, standups, retros) and almost none around the unstructured stuff that happens naturally when people share a building. Those informal moments aren’t a nice-to-have layered on top of the work. Pendell’s data suggests they function closer to a wellbeing input, one that hybrid and remote arrangements quietly remove unless someone deliberately puts it back.

Putting it back is a leadership decision, not a facilities decision. It has to be someone’s job to manufacture the equivalent of a hallway conversation on purpose: a standing unstructured call, a rotating pairing system, or just protecting the first ten minutes of a meeting for nothing work-related. The office never had to try. Hybrid leadership does.

The Leader Isn’t Immune

Gallup’s most recent State of the Global Workplace report extends the story in a direction that complicates it further. Leaders report the highest engagement of any group in the organization, and also the highest stress, anger, sadness, and loneliness of any group. Gallup’s own Vibhas Ratanjee named the gap directly in a recent Forbes piece: leaders come out roughly ten points lonelier than the people they lead.

This is the part of the human connection framework that usually gets skipped. We talk about connection as something a leader provides, a benefit designed and distributed downward. Gallup’s data says the leader is frequently the most isolated node in the system, visible to everyone and genuinely connected to almost no one. A framework that treats the leader purely as the supplier of connection, never as someone who also needs it, will eventually run out of a leader who can supply it.

This is particularly relevant in hybrid organizations, where the leader is often the one person expected to be everywhere: in the office when it matters, on the call when it doesn’t, present for every team’s rhythm but embedded in none of them. A human connection framework for hybrid leadership has to include the leader’s own connection, or it’s just a delegation of loneliness.

None of this resolves into a policy you can put in a handbook. Three days in the office, or five, or zero, isn’t the lever. The lever is whether someone has taken responsibility for the parts of human connection that don’t happen automatically anymore: the coordination autonomy quietly erodes, the camaraderie the office used to hand out for free, and the leader’s own need for the same thing they’re building for everyone else. Gallup’s data suggests the payoff for getting it right is durable: among fully remote workers who are both engaged and thriving, only 38 percent are job hunting. Everyone else is somewhere north of that. If you’re building the coaching-first leadership culture that tends to go with this, I wrote about the research behind that shift in Coach, Don’t Command.

1. The State of AI in HR 2026 Report: 87% of organizations forecast greater adoption of AI within HR processes, with 46% of organizations expecting to use AI in HR in 2026. AI’s organizational impact is 5.7 times more likely to shift job responsibilities and three times more likely to create new roles than to displace jobs.

2. AI’s Impact on Productivity in 2026: A Deep Dive for HR & Executives: The most successful organizations in 2026 are not replacing humans with AI; they are augmenting human capabilities and proactively upskilling their workforce.

3. Agentic AI to disrupt $234B in SaaS spending: AI agents are reducing reliance on traditional software interfaces, upending everything from legacy user dashboards to pricing models.

A 2019 study in Nature Communications by Luc Rocher, Julien Hendrickx, and Yves-Alexandre de Montjoye spent some time taking that password apart. Using a generative model, the team found that 99.98 percent of Americans could be correctly re-identified in any dataset using just 15 demographic attributes. Not a carefully chosen dataset. Any dataset. The number has stuck with me because the entire economic logic of de-identified health data assumes the opposite is true.

The Release-and-Forget Bargain

HIPAA gives you two ways to de-identify data, both laid out in the Privacy Rule’s de-identification guidance. The first is Expert Determination, where a qualified statistician documents that the re-identification risk is very small. The second, and the one almost everyone actually uses, is Safe Harbor: delete 18 specified categories of identifiers (names, dates, geography below the state level, and so on), and as long as you have no actual knowledge that what’s left could identify someone, the data is no longer PHI.

That’s the bargain. Once the 18 identifiers are gone, the data walks out from under HIPAA entirely and can be used for any purpose. It’s a clean, rule-based checklist, which is exactly why it’s popular (you can hand it to an engineer and they’ll apply it deterministically). The trouble is that the bargain rests on an assumption baked into the standard itself: that de-identification is a one-way door. Release the data, forget the patient.

What’s easy to miss is that HHS never actually promised the door was locked. The guidance says outright that both methods, even applied correctly, still leave data with some residual risk of identification. The two words that get dropped in practice are “some risk.” We treat de-identified as a synonym for safe. It was never meant to be.

Fifteen Attributes

Rocher’s team built a copula-based generative model that estimates how likely a given record is to belong to one specific person, even when the dataset is heavily incomplete (the partial, sampled extracts that get sold and shared all the time). Across 210 populations, the model predicted individual uniqueness with AUC scores between 0.84 and 0.97. Then came the headline result: 15 attributes, things as ordinary as age, gender, and marital status, are enough to single out virtually everyone.

The part that matters for anyone building in this space is how the re-identification happens. It’s machine learning doing the work. The same class of technique we use to pull signal out of messy health data is the technique that pulls a named individual back out of a dataset we swore was anonymous. De-identification and re-identification are running on the same engine, and the engine keeps getting better while the 18-identifier checklist stays frozen.

De Montjoye framed the practical objection well in the press materials around the paper. Organizations reassure themselves that the datasets they release are incomplete, so re-identification can’t really work, and his finding is that this “might not help.” Incompleteness was the last line of defense. The study walked right through it.

The Rule Got Modernized. The Premise Didn’t.

Here’s the contrast I can’t stop thinking about. HIPAA is, right now, in the middle of its first serious technical overhaul in over a decade. The proposed Security Rule update, published in early 2025 and still grinding through more than 4,700 public comments as of this writing, would make encryption and multi-factor authentication mandatory, require a living inventory of every system that touches ePHI, and strip away the old “addressable” wiggle room. It’s a real answer to a real threat environment. Large breaches hit more than 286 million people in 2024 alone.

But notice what’s being hardened. The Security Rule governs how you protect health data that’s inside your walls. All that encryption and segmentation and asset-mapping is about keeping the vault shut. The de-identification standard is a different animal. It’s the rule that decides which data gets to leave the building with no HIPAA protection at all. And that standard, the one machine learning most directly undermines, lives in the Privacy Rule and hasn’t fundamentally moved since the list of 18 identifiers was first written more than two decades ago.

We are reinforcing the vault and leaving the side door propped open. The data we hand out freely, because a checklist told us it was no longer PHI, is precisely the data Rocher’s model is best at unwinding.

Build For the Spectrum, Not the Switch

None of this means de-identified data is useless or that you should stop using it. It means treating “it’s de-identified” as the end of the privacy conversation is a mistake, and an expensive one to keep making. The honest mental model isn’t a switch that flips data from protected to safe. It’s a dial, and the dial reads “residual risk that climbs every year the models improve.”

For teams shipping healthcare AI, that reframing has teeth. Safe Harbor is cheap and deterministic, but for anything feeding a model (especially a model that might memorize and resurface its training data) Expert Determination is the more defensible path, because it actually reasons about your data, your recipients, and a realistic attacker. Minimize the attributes you retain, because every quasi-identifier you keep is one of Rocher’s 15. And write re-identification prohibitions into your data agreements even when HIPAA doesn’t force you to, because the regulation describes a world with fewer generative models in it than the one you’re operating in.

The number to keep on a sticky note is 15. That’s how many ordinary facts it takes to find almost anyone. The checklist deletes 18 identifiers and calls it done, but it was built for a time when re-identifying a person was hard. That time is over, and the rulebook hasn’t caught up.

I went back to that instinct after reading a systematic review of predictive models embedded in electronic health record systems, not models that were built and validated and written up, but models that were wired into live clinical use and then measured against patient outcomes. Of the 32 studies that reported clinical outcomes, 22 showed improvement after the model went live. A 69 percent hit rate is encouraging on its own. What’s more useful is what separated the wins from the misses, because it had very little to do with how good the model was.

The needle moved where someone could move it

The pattern that runs through the successful implementations is that the prediction arrived where a human was positioned to do something about it. The review’s authors highlight a heart-failure readmission model described by Amarasingham that was deliberately not run on weekends or holidays. It fired on weekdays only, specifically because that was when a heart-failure case manager was available to coordinate follow-up for the patients it flagged.

Sit with that for a second. The team intentionally narrowed when their model spoke, not to improve its accuracy, but to match it to the hours when its output could be converted into action. A flagged patient on a Saturday is a number with nowhere to go. The same flag on a Tuesday, handed to a case manager whose entire job is the next step, is a workflow. The model didn’t change between those two days. Everything around it did.

This is the part the demo slide never shows. Anticipating a patient’s needs before they arise is only half a sentence. The other half (and then a specific person does a specific thing in time to matter) is where the value lives, and it’s the half that almost never makes it into the procurement conversation.

An echo, not a warning

The clearest cautionary tale here is the Epic Sepsis Model, a proprietary tool that had been switched on at hundreds of US hospitals before anyone outside the vendor had rigorously checked whether it worked. When Andrew Wong, Karandeep Singh, and colleagues at Michigan Medicine externally validated it across more than 38,000 hospitalizations, the results were rough: an AUC of 0.63, well below the 0.76 to 0.83 Epic had reported to its customers, and a sensitivity of 33 percent. The model missed roughly two-thirds of actual sepsis cases while flagging enough patients to bury the clinicians who had to chase every alert.

The most instructive finding wasn’t the miss rate. It was why the alerts that did fire were so often useless. The model used antibiotic administration as one of its inputs, and antibiotics get ordered when a clinician already suspects sepsis. So the score frequently climbed after the bedside decision had been made. The “early warning” was, a lot of the time, an echo of a judgment call the team had already reached on their own. The validation found the model surfaced only about 7 percent of the sepsis cases clinicians hadn’t already caught.

A prediction that mostly confirms what you already decided isn’t anticipation. It’s latency with a dashboard. And it carries a real cost: every false alert spends a clinician’s attention, and attention is the most rationed resource in the building. Epic eventually overhauled the model and started telling hospitals to retrain it on their own patient data before going live, a quiet admission that a model shipped as one-size-fits-all wasn’t one.

Anticipation is a workflow decision

Put those two stories next to each other and the lesson is hard to miss. The heart-failure model worked because someone designed the conditions under which its prediction would turn into care. The sepsis model struggled because it was bolted onto a workflow that hadn’t been redesigned to receive it. It just fired into the same stream of alerts everyone was already learning to ignore.

The review names this directly. Alert fatigue, lack of training, and added work for the care team show up again and again as the reasons implementations stall, and the authors note that plugging in a previously validated model without local customization tended to make those problems worse, not better. The off-the-shelf score that performed beautifully somewhere else is not a head start. It’s an unvalidated assumption about your population, your staffing, and your existing alert load.

The implication for anyone building or buying one of these is less glamorous than the model work. The hard, valuable design happens around the prediction: deciding who owns the response, whether that person actually has the capacity to own it, what the alert threshold should be for this population rather than someone else’s, and (the step everyone skips) what gets pulled out of the alert stream to make room for the new signal. That’s operational design, not data science, and on most projects it belongs to nobody.

So the first question I want to ask about any predictive tool (ours included) is no longer “what’s the AUC.” It’s narrower and more annoying: when this thing is right, who acts, what do they do, and does the signal reach them early enough to change the outcome? If there’s no clean answer to that, a better model won’t save the project. It’ll just be a more accurate way to generate noise. The prediction was the easy part. The workflow it lives inside is the whole job.

1. AI-Powered Diagnostics and Digital Twins Reshape Healthcare in 2026: Healthcare technology is advancing through AI-powered diagnostics and personalized care with digital twins, enabling less invasive surgery and real-time detection of diseases like heart disease and colon polyps.

2. AI-Driven Mental Health Support Transforms 2026 Workplace: Organizations will further integrate AI and use data-driven tools to reduce workplace stigma around mental health.

3. AI-Enabled Collaboration Platforms Transform Team Workflows: Enterprise organizations using generative AI in planning functions are already accelerating project cycles and improving delivery consistency, with companies leading in generative AI adoption seeing both cost savings and revenue growth.

4. Foundation to tackle AI token cost management: The Tokenomics Foundation will bring together enterprises, hyperscalers and frontier model developers to address rising AI costs

So when MIT’s Project NANDA published The GenAI Divide: State of AI in Business 2025 this past summer, the headline number landed less as a shock than as a confirmation. Across more than 300 deployments, lead author Aditya Challapally and his colleagues found that roughly 95 percent of organizations were getting zero measurable return on their generative AI investment. Only about 5 percent had crossed into real value. The cause, they argued, was not model quality. It was the gap between a tool that demos well and a tool that survives contact with how an organization actually works.

The Failure Is at the System Layer

The most useful reframe in the NANDA research is also the least flattering: AI rarely fails because the model is bad. It fails because the model is the easy part. A pilot runs on a curated slice of clean data in a forgiving environment, proves feasibility, and earns a budget. Then it meets production, where every assumption about data quality, system reliability, and user behavior gets tested at once.

Healthcare makes this worse than almost any other industry, because the production environment is a museum. Most hospitals still run on electronic health records, PACS imaging platforms, and radiology systems that are ten to twenty years old and were never designed to hand real-time data to an algorithm. A 2025 scoping review of health-data interoperability found that across the studies it examined, legacy systems were the single most frequently cited barrier, ahead of limited IT capacity and semantic mismatches. The healthcare-specific numbers track with the broader picture: a Bessemer survey found that only about 30 percent of healthcare AI pilots reach production, with data readiness named as the thing holding the rest back.

None of that is a reason to slow down on AI. It is a reason to stop pretending the hard part is the model. The hard part is the seam between a new system and the twenty-year-old one it has to live next to.

The Trap Is Semantic, Not Just Technical

Here is the part that catches teams off guard. You can build an integration that works perfectly and still feed your AI garbage. Moving data between systems is the syntactic problem, and we are reasonably good at that now. Getting the receiving system to understand what the data actually means is the semantic problem, and we are not.

Health information exchanges run into this constantly. When an exchange operates as a pass-through, it delivers a lab result or an allergy list exactly as the sending system encoded it, non-standard abbreviations and all. Analysts who study these exchanges have made the point bluntly: even small volumes of nonstandard messages can overwhelm manual review queues and create false confidence that data arrived intact when the receiving system could not actually interpret it. A human clinician squints at an ambiguous field and figures it out from context. An AI does not squint. It either fails silently or, worse, proceeds confidently on a misread.

This is why “we already have FHIR” is not the reassurance people think it is. More than 90 percent of EHR vendors now support FHIR, and the silos have not gone away, because supporting a standard and implementing it identically are very different things. Version fragmentation and local code mappings mean two FHIR-compliant systems can still disagree about what a given field means. The standard gets you a shared envelope. It does not guarantee a shared language.

Don’t Rebuild the Plumbing You Can Buy

One NANDA finding I keep coming back to: tools bought from specialized vendors or built through partnerships succeeded about 67 percent of the time, while internal builds succeeded roughly a third as often. Challapally noted that almost everywhere the team looked, enterprises were trying to build their own tool, and the data kept showing that the purchased solutions were the more reliable bet.

For a healthcare CTO, the lesson is not “never build.” It is “be honest about where your differentiation actually lives.” Your clinical logic, your domain models, the workflow your users live in for ten hours a day (that is worth building). The integration layer that maps a legacy HL7 feed into something a model can consume has been solved, repeatedly, by people who do nothing else. Rebuilding it in-house is one of the most reliable ways to turn a six-month project into an eighteen-month one.

The better posture is to meet the legacy system where it is. Put a FHIR facade in front of the old EHR instead of waiting to rip it out. Use middleware to normalize the messy feed before it ever reaches the model. Treat the twenty-year-old system as a constraint you design around, not a problem you have to solve before you are allowed to start.

Sequence Is the Whole Game

If there is a single operational takeaway, it is about order of operations. The organizations that cross the divide build the data foundation first and choose the use case second. The ones that stall do the opposite. They pick an exciting use case, build a model around it, and discover in production that the data underneath was never ready to carry it.

This is genuinely unglamorous work. Cleaning and standardizing clinical data, resolving semantic mismatches, and putting governance in place so the quality does not decay over time (none of it demos well in a board meeting). It is also, by every account in the research, the thing that separates the 5 percent from the 95.

It is worth sitting with how contested that 95 number actually is. Critics have rightly pointed out that the NANDA report was preliminary rather than peer-reviewed, and that the precise figure referred to custom enterprise tools reaching production rather than every pilot ever launched. Fine. Argue the number down to 80, or to Bessemer’s 70. The diagnosis does not move. AI in healthcare does not break at the model. It breaks at the seam — and in our industry, that seam is almost always a system older than the engineers being asked to integrate it.

1. House committee takes step toward blocking Medicare AI prior authorization pilot: It’s another sign of lawmakers’ concern about the pilot, which has been lambasted for delaying care to seniors.

2. The AI bill is coming due for businesses. Some are seeing sticker shock. Many companies are struggling to determine how much they are spending on AI and whether the technology has provided any benefit before the bill comes due.

3. Shutterstock, Prudential Financial share AI cost strategies: New models and agentic AI tools are leading to more token usage, driving AI costs higher for enterprises.

4. Oracle, AWS among vendors working to curb rising AI costs: Advanced AI models and agentic tools are leading to greater token usage, which is pushing into budgets as business executives look to manage spend.

I keep coming back to that story because of a paper I read recently from Shaheen Lakhan in Cureus. Lakhan ran a bibliometric analysis of the entire PubMed-indexed PDT literature from 2020 through May 2025 — 61 articles in total — and classified each one against four foundational scientific domains: clinical, engagement, regulatory, and implementation. The result is a near-perfect explanation of why Pear failed, and a warning for everyone now layering generative AI on top of digital therapeutics expecting a different outcome.

The Lopsided Literature

Clinical science showed up in 74% of the papers. Engagement science in 58%. Regulatory in 46%. Implementation — the science of getting a therapy into actual care, through actual workflows, with actual reimbursement — appeared in just 29%. Only seven of the 61 papers (12%) touched all four domains.

That’s the gap. The field has spent most of its published intellectual effort proving that software-based interventions can change clinical outcomes, and comparatively little effort working out how to embed them into the operational reality of American healthcare. Lakhan is direct about this in the discussion: “Implementation science remains the least represented domain, despite its central role in enabling reimbursement, provider uptake, and integration into clinical workflows.”

What stands out is the co-occurrence pattern. Regulatory and implementation science — the two domains you’d most need to pair to actually commercialize a PDT — co-occurred in only nine articles. A field can’t be said to be serious about real-world adoption when fewer than 15% of its papers think about clearance and deployment in the same breath.

What Pear Actually Proved

Pear’s story is the practical exhibit. The MedCity News post-mortem panel at INVEST Digital Health surfaced the real failure mode. Smit Patel of the Digital Medicine Society made the point that payers were “still learning,” and that there was no benchmark for comparing digital therapeutics against standard of care. Pear had the evidence. What it didn’t have was a payer pathway anyone could navigate cleanly.

The takeaway from that panel that has stuck with me: try to be only slightly disruptive. Pear tried to disrupt both how care was delivered and how it was reimbursed. Healthcare can absorb one of those at a time, occasionally. Two at once is a death sentence.

Translate that back to the Lakhan numbers. 74% of the field was building the clinical case. The clinical case wasn’t the binding constraint.

The AI Wedge — and the AI Trap

Here’s where it gets interesting for everyone (including us) layering generative and agentic AI onto digital therapeutics. The optimistic case is that AI changes the implementation math. A static CBT-I module that takes a patient through eight scripted weeks is one product. A model that adapts difficulty, surfaces the right exercise at the right moment, conducts a clinical-grade conversation, and reports back to the prescribing physician inside their EHR is something else entirely. The same prescription, with workflow gravity behind it.

That’s the wedge. The pessimistic case is that AI gets used the way the literature has historically been used: to keep stacking clinical wins. We’ll see another generation of randomized trials proving an LLM-augmented therapeutic beats sham at p<0.05, while the same payers ask the same questions about coding, claims, and clinical workflow that they asked Pear. The 29% implementation number doesn’t get better just because the underlying technology got smarter. If anything, AI raises the bar, because now the model needs to be governed, audited for drift and bias, and validated against a label the FDA is still writing.

The pattern I’d watch for in the next round of PDT companies is whether they treat implementation as a primary scientific discipline or a go-to-market afterthought. The tells are concrete: implementation pilots running inside payer systems before clearance submission, workflow integration baked into the product rather than bolted on after the launch press release, and published evidence on adoption and provider uptake at anywhere near the rate of published efficacy data.

What This Means for Operators

For those of us building AI in healthcare — not just in DTx, but anywhere a model is going to touch a patient or a clinician — the Lakhan paper is a useful mirror. It’s easy to over-invest in the science we know how to do (clinical validation) and under-invest in the science we don’t (real-world integration). It’s also easy to assume that because the model performs well on a benchmark, the model will perform well in a clinic. Those are different evaluations.

The implementation lift in healthcare AI is not a marketing problem and not a sales problem. It’s a research problem in its own right, one that deserves study designs and published outcomes the same way efficacy does. Workflow and adoption work needs to sit alongside the trial work. If that body of evidence doesn’t exist alongside the efficacy data, the AI doesn’t reach the patient regardless of how good it is on paper.

Lakhan ends the paper with a line worth holding onto: PDTs need to be not only effective but “scalable, fundable, and embedded into routine care.” That’s the whole assignment. Software is the easy part. The hard part is everything that happens after the software works.

1. ‘Bots have now passed human traffic online,’ Cloudflare boss laments — says agentic traffic wasn’t expected to eclipse real people until next year: Bot (automated) vs. human HTTP requests are split 57.5 vs. 42.5 percent, according to the firm’s latest data.

2. Tech jobs grew in May despite AI layoffs: The shift to AI is creating an uneven tech job market, driving major cuts even as overall hiring ticks up and in-demand skills evolve.

3. How to Deploy AI Projects That Create Business Value: How can you develop AI projects that create measurable business value? This is the trillion-dollar question on everyone’s minds. Especially with the stories floating that most enterprise LLM projects never leave the demo/prototype phase.

4. Engineers are becoming advisers to agent-powered work: Tech teams must change how they manage people-led operations to truly benefit from agentic AI, according to a Bain & Company report.

5. AI Doesn’t Have ROI: If you can’t measure how good something is, how much it might cost, or what your return on investment might be, it’s fair to ask why you’re even paying for it in the first place.

6. Retailers turn to AI for productivity, personalized shopping: Best Buy, Gap and Dick’s Sporting Goods said their AI investments were yielding rewards and reshaping their businesses during their Q1 earnings calls.

7. What CIOs should watch for in Trump’s AI oversight order: Clear standards will be key to making the process worthwhile for AI providers and the companies that contract with them, tech experts say.

8. AI deployment plans are catching leaders underprepared: Embedding control and creating visibility into AI systems will allow leaders to scale confidently, Matt Lyteson, IBM CIO said.

The headline finding is the one everyone has cited: 65% of US hospitals are using predictive AI, but only 44% locally evaluate those models for bias. That gap is the story most coverage stopped at. The line Nong dropped in her Tradeoffs interview is the one that should stick with anyone who builds, buys, or governs healthcare AI: “A biased tool is a bad tool.” Read that again. She is not framing bias as a separate ethics question that sits next to quality. She is saying bias is a quality defect. Which means explainability (the thing that lets you find bias in the first place) is not a compliance overlay. It’s a precondition for the product working at all.

Transparency is a vendor problem before it’s a hospital problem

One of my favorite findings in the Nong paper, and the one that lands closest to home for those of us on the vendor side: 79% of hospitals using predictive models get those models directly from their EHR vendor. The bias evaluation gap isn’t really a hospital governance failure in isolation. It’s a downstream consequence of how little the buyer knows about what they bought. You can’t locally evaluate a model for bias if the vendor never told you what data it was trained on, what populations it was validated against, or what drift looks like in your environment.

This is the part of the conversation that gets uncomfortable for vendors, and it should. When Nong talks about the Biden-era transparency rules requiring developers to share information with their customers, she calls it “a really important first step.” First step, not finished work. The implicit standard she’s pointing at is one where the buyer can answer four questions before deploying anything: what was this trained on, who was it validated against, what does it predict and what does it explicitly not predict, and what monitoring tells me when it stops working. If a vendor can’t answer those four questions in plain English to a clinical lead and a CTO at the same table, the tool is not ready. It does not matter how good the AUC is on the original dataset.

The numbers are encouraging — and the gap that’s left is the dangerous one

The follow-up data from the AHA’s late-2025 analysis of the 2024 IT Supplement is the part of this story that hasn’t gotten enough play. Predictive AI adoption in EHRs climbed from about 66% to 71% in a single year. Local accuracy evaluation moved from 61% to 82%. Bias evaluation moved from 44% to 74%. Post-implementation monitoring landed at 79%. In a year. The transparency regulations and the published research worked faster than I’d expected.

But the variation underneath those averages is what should keep CTOs and CMOs awake. 86% of system-affiliated hospitals use predictive AI; only 37% of independents do. Rural adoption sits at 56% against 81% for urban hospitals. Critical access hospitals are at 50%. The AHA report calls this a “persistent digital divide” and they’re being polite. What it really is: the hospitals with the most resources are the ones doing the most rigorous evaluation of the tools they’re deploying, while the hospitals serving the most vulnerable patients are either skipping the tools or deploying them without the governance muscle to catch a bad model. Neither option is good for those patients.

Nong’s point on this was sharper than the report’s. The under-resourced hospitals, she said, are stuck between two bad choices: use tools they can’t validate, or skip AI entirely and miss whatever real value it provides. That framing (two bad choices) is what makes this an industry-level problem and not just a hospital governance problem. We do not solve this by writing better policies at the well-resourced systems. We solve it by making explainability cheap enough, standardized enough, and portable enough that a 25-bed rural hospital can evaluate a vendor’s model without hiring a data science team.

Explainability standards are a competitive advantage, not a tax

Here is where I lose some of my peers on the vendor side. The instinct in this industry is to treat transparency requirements as cost — a thing the regulators are imposing that we have to absorb. That framing is wrong on two levels. First, the buyers who are governing their AI well are already asking these questions, and they’re going to keep asking them. The vendors who can answer cleanly will win those accounts. The vendors who can’t are going to find renewals harder than they used to be. Second, the explainability work isn’t a tax on the model. It’s a forcing function on the model. If you can’t explain why your tool flagged this patient as high-risk, the odds are good that you don’t actually know either. The work of building an explanation surfaces the brittleness in the model. That is genuinely useful, even when it’s uncomfortable.

The Nong paper notes that hospitals using vendor-supplied models from their EHR are doing less local evaluation than hospitals that built their own. That’s not a coincidence. When you build it, you own the explanation. When you buy it, you inherit whatever explanation the vendor chose to ship (and if they shipped none, you have nothing to evaluate). The vendors who treat explainability as a first-class deliverable — documented training data, validation cohorts, monitoring telemetry surfaced at the clinician’s point of decision — are building a different product than the vendors who ship a black box with a confidence score next to it. Those are not two versions of the same product. They are two different products serving two different markets, and one of them is going to age very badly.

The standard worth aiming at

If I were writing a transparency requirement for healthcare AI tomorrow, I would not start with the model. I would start with the buyer’s ability to ask four questions and get four answers, in writing, before the contract is signed: what data, what validation, what predictions are in scope, what monitoring fires when the model drifts. Everything else (SHAP values, feature attribution, attention maps, the whole technical menagerie of explainable AI methods) is implementation detail in service of those four questions.

The Health Affairs data tells us we moved 30 percentage points on bias evaluation in 12 months when the regulatory and research pressure was steady. That’s a real signal. The work left is closing the gap for the hospitals that can’t afford to do it themselves, and pushing vendors to make the four-question answer the easy path, not the expensive one. Trust in healthcare AI is not going to be won by a press release about responsible AI. It’s going to be won, hospital by hospital, by vendors who can show their work.

1. Interaction Models: A Scalable Approach to Human-AI Collaboration: AI labs often treat the ability for AI to work autonomously as the model’s most important capability. Measuring AI Ability to Complete Long Tasks. METR, 2025. As a result, today’s models and interfaces aren’t optimized for humans to remain in the loop.

2. Mistral AI pitches sovereign security model to European banks: The discussions centre on a tool intended as a counterpart to Anthropic PBC’s Mythos.

3. LLMs Can Get “Brain Rot”!: “Brain rot” burst into public discourse as a shorthand for how endless, low-effort, engagement-bait content can dull human cognition—eroding focus, memory discipline, and social judgment through compulsive online consumption.

4. When building an AI strategy, don’t forget the humans: In tech, all roads lead back to people. AI is no different.

5. Enterprise data is creeping its way into shadow AI tools: Executives and employees are clashing over usage policies as AI security concerns rise, an Okta report found.

6. OpenAI’s Altman says AI unlikely to lead to ‘jobs apocalypse’: OpenAI CEO Sam Altman said on Tuesday the rapid development ‌and adoption of AI would not lead to a global “jobs apocalypse” and the technology had not claimed as many white-collar jobs as he had feared.

7. Token Maxing: The AI Industry is Struggling with Measuring Value: The AI market has a measurement problem. A year ago, you could evaluate a new model by looking at its MMLU score or open-source benchmark rankings. Today, those numbers are compromised.

8. Healthcare AI Doesn’t Have a Model Problem. It Has a Dr. Smith Problem. Somewhere in a health system’s Epic environment right now, there is a surgical procedure code called something like “Dr. Smith’s urology procedure.”

The data comes from a JMIR study by Dai, Li, and colleagues at Peking University Third Hospital, published last August. They used the Unified Theory of Acceptance and Use of Technology (UTAUT) to model what actually predicts a clinician’s intention to use AI. The findings are sharper than I expected, and they reframe how I think about “converting skeptics to champions.” Most of what we instinctively reach for, including the standard town hall about job security, isn’t what’s blocking adoption.

The Fear We Worry About Isn’t the Fear That Matters

If you’d asked me before reading this study what clinical staff fear most about AI, I’d have said job displacement. Every keynote on the topic seems to lead with it. But in the Peking study, only 28.17% of medical staff worried that AI would replace their jobs — the least common concern measured.

What ranked higher? Medicolegal liability (53.1% of physicians), AI providing incorrect information (47.3%), and the transparency and explainability of AI outputs (47%). The order matters. The dominant fear isn’t replacement. It’s being held accountable for a system clinicians can’t interrogate.

This tracks with what the AMA’s Augmented Intelligence Research survey found among U.S. physicians. When asked what would actually increase their adoption of AI tools, the top three attributes weren’t motivational. They were operational: a designated feedback channel (88%), data privacy assurances (87%), and EHR integration (84%). These are the asks of people who want to use the tools but don’t trust the surrounding plumbing.

Effort Expectancy Is Doing More Work Than We Think

The single largest predictor of physician intention-to-use in the Peking study was effort expectancy — the perceived ease of using the AI. The regression coefficient was 0.681, larger than performance expectancy, social influence, and facilitating conditions combined for that population.

That finding rearranged something in my head. We tend to sell AI on outcomes: better diagnoses, fewer errors, time savings. Those matter (performance expectancy was significant too). But the gating factor for whether a clinician will actually use the tool tomorrow morning is whether they believe they can use it without it costing them time they don’t have. Effort expectancy is the load-bearing wall.

This is why the AMA finding on EHR integration lands so heavily. An AI tool that lives outside the EHR is, functionally, a second workflow. A second workflow during a 12-minute appointment is a tax. And no clinician volunteers to pay a tax — no matter how enthusiastic they are about the cause.

Nurses and Physicians Are Not the Same Skeptic

One of the most useful findings buried in the Peking analysis is that perceived risk had no significant effect on physicians’ intention to use AI but had a significant negative effect on nurses’ intention. Same hospital, same tools, different psychology.

The authors offer a plausible explanation: physicians tend to use AI in assistive diagnosis, documentation, and research, where the consequences of an AI error are mediated by physician judgment. Nurses are in direct patient monitoring, where AI errors translate more immediately into patient harm. The risk isn’t abstract for them. It’s the alarm that didn’t sound, the deterioration that wasn’t flagged.

The practical implication is that a single change management playbook will fail. For physicians, the unlock is performance and ease of use — show them the tool saves time and works inside their workflow. For nurses, the unlock is risk mitigation — clear protocols for when to override, audit trails, escalation paths, and visible accountability for when the AI is wrong. Treating these populations identically is how programs that look good on a slide produce 19% utilization in the field.

Social Proof Beats Mandates

Here’s the part of the Peking analysis I keep coming back to as I think about our own AI transformation at Xsolis. Working in a hospital with high attention to medical AI had a significant positive effect on intention to use (B=0.998 for physicians, 1.066 for nurses). Having prior experience using AI also had a significant positive effect. Optimism about AI’s prospects had the largest demographic effect of all (B=1.914 for physicians).

These three things — institutional attention, personal experience, and optimism — feed each other. A peer uses the tool, doesn’t regret it, and word travels. The peer only had the chance to use it because leadership funded and signaled the work. And the resulting optimism is what gets the next person to try, which is where personal experience starts again.

The corollary, which is harder to swallow, is that a top-down mandate without those conditions in place mostly produces compliance theater. People will attend the training. They won’t change the workflow. The Peking authors land on it directly in their conclusions: medical staff are more inclined to adopt AI when they believe it enhances performance, is easy to operate, has environmental support, and offers favorable conditions. All four, not any one.

The Stuck Are Not the Skeptics

The “skeptic to advocate” framing is useful, but the Peking data suggests most clinicians aren’t actually skeptics. They’re already willing (73% of them). They’re stuck, which is a different problem with different mechanics. You don’t convert someone who already agreed with you. You unblock them.

What unblocks them is straightforward, even if it isn’t easy. Effort expectancy is the largest lever, so the AI tool has to be genuinely easier than the alternative and live inside the EHR rather than next to it. The risk conversation has to be differentiated by role, because the Peking regression makes clear that physicians and nurses are not the same skeptic — physicians want transparency on liability and explainability, while nurses want the safety scaffolding around AI failure modes. And the strongest predictor of all in the demographic data, optimism about AI’s prospects, mostly comes from one place: watching a peer in your own department use the tool without regret.

The AMA’s U.S. data offers a hopeful counterpoint. Physician AI use in the U.S. jumped from 38% in 2023 to 66% in 2024. The intention-behavior gap can close. It doesn’t close because we wished harder at people. It closes when the surrounding system catches up to what the clinicians were already willing to do.

1. Microsoft may discontinue Claude Code internally as it looks to push users towards GitHub Copilot: Microsoft is reportedly canceling most of the Claude Code license it uses internally, despite the Anthropic tool becoming somewhat of a hit among both workers and the community in general.

2. How to Build a Self-Improving Company with AI: In this recent batch talk, YC General Partner Tom Blomfield breaks down how to build a self-improving company using AI. He’ll cover how to create a series of recursive, self-improving AI loops and explain why the founders who get this right will run companies that can improve while they sleep.

3. AI Is Too Expensive: AI is, as it stands, not economically viable for anybody involved other than the construction firms, NVIDIA, and the surrounding hardware companies benefitting from the irrational exuberance of a data center buildout that doesn’t appear to be happening at the speed we believed.

4. OpenRouter raises $113M to bring order to enterprise AI inference routing: OpenRouter provides a quick and easy interface for individual developers and enterprise businesses to access, route to and optimize for centralized closed-source AI model providers and open-source models.

5. US spy agencies race for Nvidia AI chips under new $9 billion plan: The White House has approved a $9 billion AI funding initiative for US intelligence agencies as Washington races to secure Nvidia chips, boost surveillance capabilities, and stay ahead in the global AI competition.

6. If AI Knows Your Next Trade, What Happens to Money Managers?: Fund managers with novel strategies tend to outperform their more predictable peers, whose trades AI can mimic, says research by Lauren Cohen. How might the technology reshape asset management?

The good news is that academic research has finally caught up. Charles Binkley and colleagues at Hackensack Meridian Health published a four-component governance framework in npj Digital Medicine that I think every health system CIO, CTO, CMO, and AI governance lead should read with a highlighter in hand. It is the first paper I’ve come across that treats vendor AI selection as a structured, weighted process rather than a vibes-driven exercise (which, let’s be honest, is what most of these decisions still are).

Four prongs, not twelve dimensions

Binkley’s framework rests on four components: alignment with the health system’s strategic goals, assignment of executive sponsorship and accountability, projected impact and value, and risk assessment with multi-stakeholder input. What I appreciate about this structure is that it front-loads the executive decision (strategic fit, sponsorship) before the analyst work begins. Too many vendor evaluations I’ve seen start with a spreadsheet and end with a question of strategic fit that nobody can answer at procurement time.

The executive sponsor in Binkley’s model is not a rubber stamp. The sponsor is “accountable for securing financial support” and proactively identifies mitigation strategies, leading the integration into existing workflows. This is the part everyone nods through and nobody enforces. If a health system can’t name the senior leader who will personally own the outcome of an AI deployment (and accept the operational and political consequences of its failure), the selection process should stop right there. No risk model compensates for absent accountability.

The other thing this framework gets right is making impact and risk parallel workstreams, not sequential ones. The Impact and Value Case has six domains, including a sober assessment of whether AI is even the right answer (sometimes a better non-AI risk calculator wins). The Risk Assessment runs alongside it. You don’t earn the right to deep risk analysis by passing impact first; both have to land.

Twelve risk domains, three that matter most

The risk assessment in the paper is the part that turned my reading into a working document. Binkley’s team identified 12 risk domains, each weighted by contribution to overall risk of harm. Nine of the twelve carry an average weight of about 6 percent. Three carry 15 percent each: enterprise risk, clinical documentation and decision support, and quality and patient safety.

That weighting is not arbitrary. The paper argues that enterprise risk is weighted high because the harm from an AI failure cascades: financial, reputational, operational, and trust effects ripple across the health system, and one failed model can damage adoption of every other model in the portfolio. Clinical documentation and decision support is weighted high because of how autonomously the model operates without clinician oversight. Predictive sepsis models that produce outputs without recommending interventions create confusion, frustration, and decreased adoption (the Epic Sepsis Model is the canonical cautionary tale here, and Binkley cites it). Quality and patient safety is weighted high because the paper is unwilling to let “we’ll monitor it after deployment” substitute for pre-deployment safety analysis.

If you do nothing else with this framework, weight your evaluation criteria with these three at the top and let the rest fall in line. The fastest way to find out if a vendor takes patient safety seriously is to ask them to walk through their incident reporting process and their pre-deployment safety identification methodology. The answer (or lack of one) tells you everything.

Maturity beats cost

Layered on top of the Hackensack framework, the 2025 Menlo Ventures healthcare AI report (drawn from a Morning Consult survey of more than 700 healthcare executives) corroborates with one finding that should be printed on every procurement officer’s monitor: leading systems like Mayo Clinic, Cleveland Clinic, and Kaiser Permanente are prioritizing maturity of technology, level of risk to patient care, and short-term value delivery. Cost is secondary.

This is a real change. Healthcare procurement has historically been a price negotiation with a feature comparison stapled to it. The Menlo data suggests that the most sophisticated buyers are now willing to pay a premium for production-ready solutions because the cost of failure (operational disruption, patient harm, reputational damage) dwarfs the procurement savings. Kaiser deployed Abridge across 40 hospitals and 600 medical offices in what the report describes as Kaiser’s fastest implementation of any technology in more than 20 years. They did not get there by running a price-led RFP.

Provider buying cycles have compressed from 8 months to 6.6 for health systems, and from 6 months to 4.7 for outpatient providers. That compression is not a relaxation of rigor. It is the result of better rigor up front: strategic alignment, executive sponsorship, and weighted risk analysis done well enough to make the downstream decision fast.

The vendor questions that separate signal from noise

Three questions, drawn directly from the risk domains in the Binkley framework, that I’d put at the top of any vendor scorecard:

First, on Data Acquisition and Integrity: what access do we have to your training and testing data? Full access is low risk per the framework. Access to statistics only is intermediate. No access at all is, by extension, the position vendors should have to defend explicitly before you proceed. The paper is direct on this point: a vendor’s willingness to show measurable impact data (not just training-set accuracy) is the single best proxy for whether they’re worth a deeper conversation.

Second, on Interpretability and Transparency: how do you measure interpretability, and can you produce SHAP or LIME outputs for the predictions your model is making in our environment? “It’s a black box, but it works” is not a clinical answer. Clinicians who can’t trust an AI output won’t act on it, and an unactioned prediction has an ROI of zero.

Third, on Enterprise Risk: is the model’s output retained in the EMR? If yes, you have four downstream issues — clinician misinterpretation of a prediction as fact, propagation of model error into the medical record, HIPAA and FDA and CMS regulatory exposure, and eDiscovery and storage cost. This single question reframes the entire conversation about model governance.

The unsexy work pays the bills

The reason this framework matters is not that it produces better vendor decisions on any single transaction (though it does). It matters because it produces a defensible, repeatable selection process that scales. When your AI governance committee is reviewing the seventeenth vendor pitch of the quarter, the framework is what prevents the process from collapsing into “I liked their demo.” When the board asks why you chose Vendor A over Vendor B, the framework is what gives you an answer that doesn’t sound like rationalization.

Binkley’s paper closes with a line I’ve reread several times: any vendor unwilling to share measurable impact data (not just training-set accuracy) should be approached with suspicion. That sentence is going on a sticky note above my monitor. The vendors worth buying from will welcome the rigor. The ones who can’t survive it were going to fail in production anyway — better to find out at evaluation than at go-live.

The market has produced eight healthcare AI unicorns and counting. The selection problem isn’t going to get easier. The framework you bring to it is the only thing that scales.

1. Medicare’s new payment model is built for AI, and most of the tech world has no idea: Neil Batlivala has spent seven years building a healthcare company that most of the tech industry has never heard of and that serves a patient population most of Silicon Valley ignores. But last month, that work put him at the center of something much bigger.

2. Engineering roles shift from developing code to managing AI: AI is producing code faster, but 81% of engineering leaders say the time saved is now spent auditing AI output, according to a new report.

3. Why Clinician-Led AI Strategies Are Gaining Momentum in Healthcare: The future of healthcare AI may depend less on algorithms and more on credibility. Presbyterian Healthcare Services is testing that theory with a clinician-led deployment strategy.

4. Complete Guide to AI Transformation: Most AI transformations fail because they start from the outside in — with tools, vendors, and pilots. This guide introduces the Core-and-Orbit Model: a framework for building AI transformation around your organization’s unique intelligence, not someone else’s technology.

5. 2026 AI Trends: What Enterprises Need to Know: Autonomous AI agents represent one of the most significant 2026 AI trends transforming how organizations operate. Unlike traditional AI tools that respond to prompts, agentic AI systems take initiative, make decisions, and execute complex workflows with minimal human intervention. These intelligent agents function as digital employees capable of managing multi-step processes across different systems.

6. How to Move from AI Experimentation to AI Transformation: Many companies are investing heavily in AI but failing to translate isolated productivity gains into meaningful business results. The problem is a “micro-productivity trap,” where firms optimize tasks without rethinking workflows or value