Making AI decisions in healthcare without repeating the last twenty+ years
Healthcare has a habit of stepping over dollars to pick up nickels.
The clearest example is still Meaningful Use. Federal incentives arrived, organizations reorganized around capturing them, and the money was real. What the money never covered was the productivity it cost to get. That trade wasn't irrational: the incentive was legible, sitting on a line item, arriving on a schedule. The cost was illegible: distributed across every clinician's day in fragments too small to invoice and too constant to ignore.
We are still paying it. Between 1998 and 2019, physician hours spent on direct patient care fell by roughly 12%, while hours spent on indirect patient care, including charting, documentation, messages, the paperwork wrapped around the visit, rose by 61%. That figure comes from Canadian data under a different payment system, which arguably makes the pattern more striking rather than less; the accounting differs, the substance does not.
The consequence is visible in every clinic in America. A January 2026 commentary from the National Academy of Medicine named it the 1.2-FTE problem: clinicians employed at one full-time equivalent who effectively work 1.2, the extra fifth being an invisible second shift of administrative work. The reality is worse than the headline. Per visit, primary care clinicians spend a median of six minutes on after-hours documentation and eight minutes on their in-basket. At eighteen patients a day, that's roughly 108 minutes of pajama time and 144 minutes of in-basket work per clinic day. A third of surveyed family medicine residents spend three or more hours a night in the EHR. The burden is now being trained into the next generation before they finish learning medicine.
Everything since has been an attempt to recoup the loss. Billing for messages. Billing for phone calls. New codes for work that used to be free because it used to be fast. These are not innovations. They are compensation for a productivity decline that standardization caused and never gave back.
I raise this not to re-litigate 2011, but because the same accounting failure is available again, and the incentive this time is less government mandate and more technological revolution.
The wolf isn't coming
For twenty+ years, folks have been predicting healthcare's collapse. It hasn't arrived, and I don't believe it will.
The reason is uncomfortable: patients are told they have choice and agency, and in the moment of actual need, they don't. They go where the need can be met. That structural fact insulates healthcare from the market discipline that forces other industries to innovate or die.
Most people treat that as an indictment. I'd argue it's the most useful fact available, because it means you are not in a race. Nobody is going to be disrupted out of existence next quarter for failing to adopt AI. The urgency in your non-clinical inbox is manufactured by vendors with a quarter to close, by consultants with a framework to sell, by trade press that needs something to publish on Tuesday.
You have time to think. Almost nothing in your inbox wants you to know that.
What makes healthcare actually different
Before any build-or-buy conversation, one constraint has to be understood, because getting it wrong is what produces expensive failures that look inexplicable afterward.
Healthcare does not permit mistakes. Not in the regulatory sense, but in the human one. Clinical judgment sits between the ears of a licensed provider whose name is on the chart, whose license is on the line, and who is trained to treat error as unacceptable. Give that person a tool that gets one thing wrong, and they will never open it again. They'll also tell their colleagues.
Call it single-strike abandonment, and understand that it doesn't work this way elsewhere. A salesperson whose AI tool invents a statistic shrugs, fixes it, and keeps using the tool. A clinician doesn't. The adoption curve in healthcare has a cliff at the edge, and most of the AI pilots that quietly died in the last two years died there rather than in procurement.
This is the real constraint. Not compliance, not integration, not budget. Trust, granted once, withdrawn permanently.
For many, there is no shock here. But for some, especially those hell-bent on adopting AI at all costs and at consumer-level speed, there is a further point to make. It's not just clinical error at risk, as the market has matured rapidly to avoid AI 1.0-level hallucinations. It is about operational error. The difference is that we are already dealing with overflowing in-boxes, decreasing reimbursement, long hours, and a level of fatigue that permeates from front to back office and all stops in between. For decades we've asked healthcare to do more with less, alongside an irrational belief that one magic app would solve all our ills.
The line: shared and outward-facing versus role-specific
Which brings me to the only build-versus-buy heuristic I've found that holds up.
If the objective is shared across the organization, and especially if the output faces outward, to a patient, a payer, or a regulator: buy it.
Not because vendors build better technology. Because at that boundary you need consistency, defensibility, and a validation burden someone else carries. Every clinician touching that workflow gets exactly one chance to lose faith in it, and a vendor with thousands of deployments has already absorbed the failure modes you'd otherwise discover in production, on your own people, once.
If the work is specific to one role or one person, build it.
How someone works moment to moment is intensely personal. The organization's objective is shared; the path each employee takes to it is not. A prior authorization coordinator, a practice manager, a care manager working a panel: each has a workflow shaped by their own habits, their patients, their local reality, their training, the way they attack their work. Generic tools fit these poorly, which is why enterprise platforms bought to solve personal-productivity problems rarely end up used as designed.
The economics have moved here, and this is the part most organizations haven't absorbed. Building genuinely useful role-specific tooling used to require a development team. It doesn't anymore. Modern tooling has collapsed that cost far enough that the vendor markup on a generic product no longer makes sense for work that only one role performs.
And there's a tier in between that almost nobody has on their map. A role has many occupants: every prior auth coordinator, every practice manager. Build for the role rather than the individual and you get the fit advantage of custom tooling with far better economics, without crossing into the shared, outward-facing territory where you should be buying. That middle layer is where most of the unclaimed value currently sits.
Two things people get wrong on the way
Compliance is usually an alibi. "Legal won't allow it" is doing more work as a refusal than as a fact, and the person saying it frequently can't name the rule. The tooling to handle PHI responsibly exists: compliant environments, sandboxed processing, appropriate agreements. These are infrastructure decisions that sit underneath both the build path and the buy path, not a reason to avoid either. What they require is someone willing to build, implement, and structure it correctly. That's work. It isn't a prohibition, and the two get conflated constantly by organizations that find the confusion convenient.
Your EHR is a billing system, still. Every one of them is, whatever the marketing says. The data model was built to capture charges, and clinical utility is inference layered on top. This is why AI bolted onto an EHR so often disappoints: the substrate wasn't designed for the question being asked. It also means the dominant vendor's roadmap is oriented toward revenue capture, which may or may not overlap with what would actually help your clinicians or your patients.
And it requires a reminder of something important for most of American healthcare: the majority of care moments still happen outside hospitals, and outside the outpatient clinics owned by hospital systems. Independent practices, urgent care, behavioral health, home health, community organizations, ancillaries. These organizations are not the target customer for the vendors courting large systems, and the interest they do attract is usually a function of revenue capture rather than improving anyone's life, patient or employee. Meanwhile the burden they carry has grown, and the common implementations continue to fall woefully short at exactly the edges where these organizations live.
This means you need a sharper eye, not a bigger budget. Need versus want. Specific versus general. And some creativity while emerging categories mature.
What working looks like
Here is the part I most want to convey: there are organizations thriving right now. They are hard to see, because thriving quietly doesn't generate press releases, but they exist and I've watched them.
They share a small number of traits.
They were thoughtful, and they were not swayed: not by trade press, not by a board member's golf conversation, not by a great salesperson. They did the hard work first: an honest assessment of where they actually were and where they actually wanted to be. Not from a spreadsheet. From actually understanding operations and watching the work happen. Then they pinpointed two or three areas for improvement and stayed in that lane while everyone around them chased the next thing.
The "killer app" idea in healthcare is dumb. There isn't one and there won't be. But incremental improvements are real, available now, and some of them are worth a great deal to evaluate and pursue.
The published evidence agrees. Hawaii Pacific Health ran a campaign inviting staff to nominate low-value EHR steps for elimination, "Getting rid of stupid stuff," and recovered roughly 1,700 nursing hours a month. Atrius Health measured in-basket volume by folder, set a baseline, and applied a simple eliminate/automate/delegate/collaborate discipline, cutting overall primary care physician messages by 25%, with some categories down between 30 and 98%.
Neither of those is artificial intelligence. Neither required a platform purchase. Both beat what many organizations got from a decade of enterprise investment, pilots, committees, evaluations, RFPs, and demos.
Where AI has delivered, it has looked the same in shape: narrow, measured, aimed at one specific burden. Stanford deployed AI-drafted replies to patient in-basket messages across 162 clinicians and found significant reductions in burden and burnout, and no change in time at all. A Dutch academic hospital found the same thing: solid adoption, response times statistically identical. Read that carefully, because it may be the most useful finding in this whole space. The tool didn't make the work faster. It made the work lighter. That distinction matters, and it's exactly the kind of result that gets flattened into a press release about efficiency gains. Ambient documentation shows a similar pattern; a 2025 randomized trial, still in preprint, found reductions in burnout, work exhaustion, and task load. Not a transformed organization. One job made less heavy. The PR wires oversimplify where AI belongs in healthcare, but tools that are specific, well-scoped, and built around the actual work are safe, manageable, and worth pursuing.
Win a battle, not the war
The lesson of the last twenty years isn't that technology failed healthcare. It's that healthcare kept buying the war and losing the battles.
Nobody wins the whole war. Not Epic, not the best-funded startup in the category, not the health system with the largest IT budget. But anyone, a vendor, a clinic, a two-physician practice with no IT department, can win a specific battle if they're willing to do the analysis and planning first and stay focused afterward.
Pick the thing that is actually costing you. Watch the work to find it rather than inferring it from a report. Ask whether the objective is shared and outward-facing or specific to a role. Buy in the first case, build in the second, and run both on infrastructure that handles PHI properly regardless.
Then stay in that lane long enough for it to matter.
The wolf isn't at the door. That's not permission to do nothing. It's permission to do this well.
4ealth Consulting Group (4CG) provides technology and product due diligence for healthcare investors and interim operating leadership for portfolio companies. Evan Frankel is the principal.
Sources
Lee SK, Mahl SK, Rowe BH. "The Induced Productivity Decline Hypothesis: More Physicians, Higher Compensation and Fewer Services." Healthcare Policy, 2021;17(2):90–104. — direct vs. indirect patient care hours, 1998–2019
Atabeygi A, Mitchell C. "The Real Driver of Burnout: The 1.2-FTE Problem." NAM Perspectives, January 2026. — 1.2-FTE framing, pajama time and in-basket minutes, resident EHR hours, Hawaii Pacific Health, Atrius Health
Garcia P, Ma SP, Shah S, et al. "Artificial Intelligence–Generated Draft Replies to Patient Inbox Messages." JAMA Network Open, 2024;7(3):e243201. — burden and burnout reductions, no change in time
Bootsma-Robroeks CMHHT, et al. "AI-generated draft replies to patient messages: exploring effects of implementation." Frontiers in Digital Health, 2025;7:1588143. — 58% adoption, response times statistically unchanged
Lukac PJ, et al. "A randomized clinical trial of two ambient artificial intelligence scribes." medRxiv preprint, July 2025.