A digital health company signs a pilot with a health system. Everyone is pleased.
Then comes security review, which takes eleven weeks. Legal, which takes seven. IT integration scoping. Value analysis committee. A service line committee that meets monthly and defers the item twice.
Eight to eighteen months later the pilot runs. Some clinicians like it. A few are enthusiastic. The data is ambiguous in the way pilot data usually is.
And then it stops.
Not because anyone decided against it. Because nobody inside the system owned carrying it from pilot to a paid enterprise contract. The innovation office ran the pilot and does not hold the operating budget. The service line has a budget and did not run the pilot. The clinicians who liked it have no mechanism to advocate beyond saying so.
The vendor burns through another two quarters of runway. The system adds another entry to a portfolio of pilots that never scaled. And eleven months later, a different health system begins evaluating the same category of product, from scratch, learning nothing.
The variable that decides these outcomes is almost never the product. It is whether a credible practising clinician inside that system will stake their own reputation on it, in a room, in front of the people who sign contracts.
Vendors cannot find that person from outside. Systems frequently cannot identify them from inside. And nobody pays them.
The failure rate is now measured
For years this was an anecdotal complaint from vendors, easy to dismiss as sour grapes. Recent work has produced numbers.
MIT's NANDA analysis, drawing on 300 public enterprise deployments, 150 leader interviews, and 350 employee survey responses, found that roughly 95 percent of enterprise generative AI pilots produced no measurable profit-and-loss impact.
That is not healthcare-specific, and healthcare has no reason to think it performs better. Set it alongside the healthcare-specific picture:
Implementation research has a name for this. A Social Science and Medicine analysis of an NHS innovation intermediary was built around "the unnecessary repetition of pilots, so-called pilotitis," and demonstrated that pilot evidence does not travel to the next adopter without deliberate intermediary work.
A Digital Health study of one hospital implementation found barriers spanning every domain of the standard implementation framework, concluding that overcoming pilotitis requires strategies addressing all of them simultaneously rather than any single fix.
And the evidence base underneath the products is thin. A JMIR analysis of 224 venture-backed digital health companies found 44 percent had a clinical robustness score of zero, and that robustness was essentially uncorrelated with funding raised.
So: a category with weak underlying evidence, an implementation process that fails at very high rates, and a documented pattern of the same pilots being repeated across institutions because nothing learned in one travels to the next.
The champion is the mechanism
Anyone who has run a health system implementation knows what actually determines the outcome, and it is not in any procurement framework.
It is whether a respected clinician says, in the room, that this should be paid for.
Not a user. Not a satisfied pilot participant. A clinician with standing, who is willing to spend that standing, in front of a committee, on a product that might fail.
This is the same primitive that appears throughout this series under different names: sponsorship in academic careers, vouching in hiring, attestation in expertise. Someone with credibility putting it at risk on behalf of something else.
And it has all the same properties. It cannot be mandated. It cannot be purchased directly. It is invisible from outside. And it is unevenly distributed by proximity.
Look at how champions are currently identified. The innovation office recruits from the handful of digitally visible physicians at that institution: the ones who show up at innovation events, who are already known to the office, who have expressed interest in technology. That set is small, systematically unrepresentative, and frequently does not include the physicians who actually own the workflow the product must fit into.
Meanwhile the vendor has no way to verify that a self-nominated champion has any standing at all. A physician who volunteers enthusiastically may be someone the committee does not listen to, and the vendor will not discover this until the decision meeting.
The labor problem nobody names
Here is the part that gets consistently overlooked, and it is the reason this does not resolve on its own.
Being a clinical champion is a substantial, unpaid, uncredited job.
Consider what it actually involves. Attending vendor meetings. Learning the product properly. Training colleagues. Fielding complaints when it does not work. Collecting and interpreting data. Preparing and presenting to committees. Absorbing the reputational damage if it fails.
That is easily dozens of hours over a pilot period, on top of a full clinical load, for no compensation, no protected time, and no professional recognition.
So the supply of champions is limited to people with unusual personal enthusiasm and slack in their schedule. Which is not the same population as people with relevant workflow expertise, and is a small fraction of it.
And there is a compounding effect. The clinicians who do this well get asked repeatedly, because the innovation office knows them, until they stop. Exactly the concentration pattern that appears in peer review, in sponsorship, and in every other uncompensated professional contribution described in this series.
Pilot purgatory is usually analyzed as a procurement problem. It is more accurately a labor problem, and no amount of procurement reform touches it.
Why nothing is learned between pilots
The second structural failure is that the knowledge generated by a pilot never leaves the building.
When a health system completes a pilot, someone in that building knows things of substantial value:
- What broke that nobody anticipated.
- Which clinicians used it and which quietly stopped.
- What the value analysis committee actually asked about.
- What the vendor promised and what was delivered.
- Whether the integration worked as described.
- And why it ended.
None of that is written anywhere another institution can see.
The employer treats it as confidential and competitively sensitive. The vendor treats successful pilots as reference assets and unsuccessful ones as things not to discuss. The clinicians involved have no venue in which to record it.
So the next system re-derives the same finding at the same cost. And the vendor re-buys the same lesson at the price of another eighteen-month sales cycle.
The implementation literature identified this precisely: pilot evidence does not travel without deliberate intermediary work, and almost nobody performs that work.
The stakes just rose
Two developments have made this substantially more consequential in the last two years.
The AI wave industrialized the failure. Where a system might once have run three or four pilots a year, many are now running dozens simultaneously across ambient documentation, message triage, imaging, coding, and operational forecasting. The per-pilot failure rate has not improved and the volume has multiplied.
And buyer tolerance has collapsed. Industry reporting on 2026 digital health funding describes buyers with "zero appetite for failed technology rollouts." Where a demonstration once carried a deal, the reference call and the champion question now decide it.
Which produces a market where the deciding factor in most healthcare technology purchases is an unpaid, uncredited, proximity-recruited clinician, and no mechanism exists to find, verify, or compensate them.
What would actually work
Pay the champion. This is the most direct and least attempted intervention. If the role is genuinely dozens of hours of skilled work that determines a six or seven figure purchasing decision, treating it as a volunteer activity is an obvious mispricing. Protected time or direct compensation, disclosed to everyone including the committee, would expand the supply beyond people who happen to be enthusiastic and available.
Verify standing, not enthusiasm. A champion's value is entirely their credibility with the people who decide. Peer attestation from colleagues, rather than self-nomination to a vendor, is the only signal that distinguishes them.
Match on workflow ownership. The right champion for an ambient documentation product is someone who documents heavily in that setting, not a physician with a general interest in technology. That is a specific, findable attribute and nobody indexes it.
Record implementations at the person level. "I ran this, here, at this size, on this EHR, and here is what happened including what broke." Attached to a person who can be asked follow-up questions, because the value in this domain is almost entirely in the follow-up. A de-identified survey cannot answer "what did your committee ask about," and a named implementer can.
Include the abandonments prominently. The pilots that stopped contain more usable information than the ones that scaled, and they are systematically absent from every channel that exists.
And keep vendors out of the record. The moment vendors can see or influence a peer implementation record, it becomes marketing. This is the same constraint that applies to peer reference calls, for the same reason.
What you can do now
If you are a clinician asked to champion something
Ask what it involves, in hours. Then decide whether you have them. Most champions substantially underestimate the commitment and then resent it, which harms the product's chances more than declining would have.
Ask for protected time or compensation. This is a reasonable professional request for a real workload that shapes a large purchasing decision, and asking normalizes it for everyone who comes after you.
Ask what happens if you conclude it does not work. A champion who cannot say no is not evaluating anything, and both you and the institution should want to know the answer before you start.
Write down what happened, including the failures. You are the only person who will hold that knowledge, and it currently disappears when you change jobs.
If you run an innovation office
Widen your champion pool beyond the usual names. The physicians who attend innovation events are not necessarily the ones who own the workflows you are trying to change, and recruiting by proximity is the most common reason a well-liked pilot fails to convert.
Assign the conversion owner before the pilot starts. The most common failure is not a bad pilot but the absence of anyone whose job it is to carry a successful one to a contract. Name that person on day one.
Publish your abandoned pilots internally. Your own organization is probably repeating pilots it already ran, for the same reasons other organizations are.
If you are selling into health systems
Ask who owns conversion, in the first meeting. If the answer is unclear, the pilot is unlikely to convert regardless of how it performs, and you now know that before spending eighteen months.
Verify your champion's standing. Ask directly who else in the organization the committee listens to on this topic, and whether your champion is one of them. Enthusiasm is not standing.
Ask what happened to the last three pilots in this category. The answer tells you more about your prospects than any evaluation criteria document.
If you invest
Ask portfolio companies about champion recruitment specifically. Given that pilot conversion is the dominant failure mode in this sector, how a company identifies and supports champions is a more predictive diligence question than most product questions.
Frequently asked questions
What is pilotitis in healthcare? The unnecessary repetition of pilots across institutions, where pilot evidence fails to travel to the next adopter. It is a named phenomenon in implementation research, examined in a Social Science and Medicine analysis of an NHS innovation intermediary, which found that moving pilot evidence between adopters requires deliberate intermediary work that rarely occurs.
What proportion of AI pilots succeed? MIT's NANDA analysis of 300 public enterprise deployments, 150 leader interviews, and 350 employee survey responses found roughly 95 percent of enterprise generative AI pilots produced no measurable profit-and-loss impact. This is a cross-industry figure rather than healthcare-specific.
Why do healthcare pilots fail to convert into contracts? Frequently because nobody inside the organization owns the transition from pilot to paid deployment. Innovation offices run pilots without holding operating budgets, service lines hold budgets without having run the pilot, and enthusiastic clinician users have no mechanism to advocate at the decision level.
What does a clinical champion actually do? Attends vendor meetings, learns the product, trains colleagues, handles complaints, collects and interprets data, presents to committees, and absorbs reputational risk if it fails. This is typically dozens of hours over a pilot period, performed without compensation, protected time, or professional recognition.
How strong is the evidence base for digital health products? Weaker than funding levels suggest. A JMIR analysis of 224 venture-backed digital health companies found 44 percent had a clinical robustness score of zero, with robustness essentially uncorrelated with the amount raised.
Why can't vendors find the right clinical champion? Because standing is invisible from outside. A vendor can identify clinicians who express interest but cannot verify whether the decision-making committee actually listens to them, and health system innovation offices typically recruit from a small set of already-visible physicians rather than from those who own the relevant workflow.
The bottom line
Health systems are running more technology pilots than at any point in their history, at a documented failure rate that would be unacceptable in any other capital process, and repeating each other's pilots because nothing learned in one institution reaches the next.
The variable that most often decides the outcome is a single clinician with credibility, willing to spend it in a committee room.
That person is recruited by proximity rather than by relevance, verified by nobody, paid nothing, and given no recognition. When they do it well they get asked repeatedly until they stop.
And when the pilot ends, everything they learned about what actually broke, what the committee cared about, and why it did or did not convert stays inside the building until they change jobs, at which point it stops existing.
We have built a multi-billion dollar procurement process whose decisive input is an unpaid volunteer that nobody can find, and whose accumulated learning is discarded after every attempt.
Part of a series on the missing professional infrastructure of healthcare. Previously: Serendipity-Only Pathways
Evidence note: enterprise AI pilot failure figures come from MIT's NANDA "GenAI Divide" analysis (2025) as reported in the business press, covering 300 public deployments, 150 leader interviews, and 350 employee survey responses, and are cross-industry rather than healthcare-specific. Pilotitis analysis comes from Social Science and Medicine (2024) examining an NHS innovation intermediary, and from Digital Health (2025). Clinical robustness figures come from JMIR (2022) covering 224 venture-backed companies. Buyer sentiment and funding figures come from Rock Health reporting as summarized in trade press.
