A chief nursing officer at a 400-bed community hospital is about to authorize a virtual nursing pilot. Cameras in forty rooms, a remote nursing station, a redesigned admission and discharge workflow, eighteen months of change management, and a budget with several commas in it.
She has done her homework properly. She has read the trade coverage. She has seen four conference presentations. She has spoken to three references the vendor supplied, all of whom were enthusiastic, as vendor-supplied references reliably are. She has a colleague from a previous job who thinks it went well at their place, probably.
What she actually wants is one specific conversation. She wants the name of the person at a similar hospital who ran this exact pilot eighteen months ago, and she wants to ask them four questions:
- What did you measure, and what did you wish you had measured?
- What broke that you did not expect?
- Where did the nurses push back, and was the pushback right?
- If you were doing it again, would you?
That conversation would be worth more than the entire vendor evaluation process. It would take twenty-five minutes.
She has no way to find that person.
Not because they are hiding. Because nothing in American healthcare records who ran what. The institution is recorded. The vendor is recorded. The press release is recorded. The human being who ran the thing, learned the lessons, and could tell her in twenty-five minutes what will take her eighteen months to discover, is indexed nowhere at all.
The number that should be a national scandal
Start with what happens to healthcare improvement work generally, because the pattern is worse than most people inside it realize.
Research published in Health Research Policy and Systems found that fewer than 40 percent of healthcare improvement initiatives transition to sustained spread beyond one area of one organization. Manager turnover is named explicitly among the barriers.
A review in BJA Education put the broader figure bluntly: up to 70 percent of attempted organizational change fails. The same literature gives us two wonderfully honest pieces of jargon: "not invented here," and "improvement evaporation."
So the base rate is this. Most improvement work does not spread. A large share does not even survive in the place that invented it.
Now consider what that implies about the information environment. If 60 to 70 percent of initiatives fail or fail to spread, then the majority of the useful knowledge in healthcare operations is knowledge about failure. What did not work, in what setting, and why.
And almost none of it is ever written down, because nobody has ever built a career on presenting the pilot they shut down.
Everyone is running the same experiment, simultaneously, in private
If this were merely inefficient, it would be a tolerable cost of a decentralized system. What makes it remarkable right now is the sheer synchronization.
Consider ambient AI documentation, the most rapidly adopted clinical technology in a generation.
A study published in JAMIA in 2025 surveyed 43 health systems on their status across 37 AI use cases. On ambient notes, the finding was unanimous: all 43 of 43 systems reported adoption activity, with roughly 40 percent actively piloting. Seventy-seven percent cited immature tools as a primary barrier, and the authors explicitly called for shared strategies and governance models.
An AHA Market Scan piece in 2026 examined six named health systems that had deployed ambient AI scribes: Emory, Mass General Brigham, Cleveland Clinic, Cooper, Mercy, and Intermountain.
Each ran a separate evaluation, with non-comparable metrics.
Six sophisticated organizations, with excellent people, evaluating substantially the same category of product, at the same time, and producing six sets of results that cannot be compared to one another.
Now scale that mentally to the roughly 400-plus US health systems and 6,100 hospitals, and hold it next to the governance picture:
- 88 percent of 233 health systems reported using AI, while only 18 percent had mature governance and a fully formed AI strategy (HFMA, 2025).
- A scoping review of 77 published AI governance frameworks found that only 19.5 percent specified an actual oversight mechanism such as a committee (npj Digital Medicine, 2026).
- Only 8 percent of physicians said their organization's AI decision-making process was clear to them (Doximity survey, spring 2026).
So we have hundreds of organizations independently writing the same intake process, the same risk tier definitions, the same vendor questionnaire, the same monitoring plan, and the same patient consent language for ambient recording. Simultaneously. In private. Mostly badly, judging by the maturity numbers.
Roughly 150 hours of committee, legal, informatics, and clinician time per organization for a three-vendor evaluation is a conservative estimate. At a blended $200 an hour across 400 systems, that single exercise is on the order of $12 million of duplicated professional labor, for one technology category, in one year.
That is the visible tip. The same pattern is running right now on virtual nursing, in-basket message triage, sepsis prediction, capacity command centers, hospital-at-home, and discharge lounges.
"But we have peer networks"
Healthcare has more peer networks than almost any other industry: Vizient, the Scottsdale Institute, AONL, CHIME, HFMA, ACHE, HIMSS, the American Hospital Association, plus every state association and specialty leadership group.
They are genuinely valuable and they do not solve this. Understanding precisely why is the whole point.
They are institution-keyed, not person-keyed. Membership belongs to the organization. The Scottsdale Institute has roughly 67 member systems. If you leave that system, you leave the network, and your knowledge leaves with you.
They operate on non-attribution by design. The Chatham House norm that makes candid discussion possible in the room is the same norm that prevents anyone from later finding out who said the useful thing. You cannot follow up with a person who was deliberately not named.
They are synchronous and scheduled. A quarterly call is enormously valuable if your question happens to arise near the call. Most questions arise on a Tuesday in the middle of a budget cycle.
They are gated by seniority and dues. The person who actually ran the pilot is frequently a nurse manager, an informaticist, or a project lead, not the executive who holds the membership.
And then there are the commercial substitutes. Analyst firms sell peer insight, and they sell it well. But their unit of analysis is the institution, their revenue includes the vendors being rated, and their model is a subscription to curated research rather than a phone call with the person who actually did it.
The result is that the industry has many mechanisms for exchanging conclusions and almost none for exchanging operators.
The specific thing that gets lost
Let us be precise about what knowledge disappears, because "institutional memory" is vague enough to be useless as a concept.
When a pilot ends, the organization retains: a slide deck, a business case, some metrics, a vendor contract, and possibly a lessons-learned document that nobody will read.
What it loses is everything that actually determines whether the next attempt works:
- The reason step three exists. Every mature protocol contains a step that looks unnecessary and exists because of an incident nobody documented.
- Which metric was misleading. Almost every pilot has one measure that looked great and was measuring the wrong thing.
- Where the resistance came from, and whether it was correct. Frontline pushback is sometimes obstruction and sometimes the most valuable signal available, and telling them apart is a judgment held by a person.
- What the vendor said versus what the vendor delivered. Never written down anywhere, for obvious contractual reasons.
- The workaround. Every deployed system has undocumented adaptations that make it work in practice, invented by staff, invisible to leadership.
- Why it was actually shut down. Rarely the reason in the final report.
Now put that against the turnover rate.
Registered nurse turnover runs 17.6 percent with average bedside RN tenure of roughly 6.8 years (NSI, 2026). Hospital chief executive departures ran to 78 through July of 2025, up 15 percent year over year. Physicians are leaving clinical practice at a mean age of 48.1 by one Permanente Journal analysis.
Every departure is a deletion, and the deletion is silent. Nothing in any system flags that the only person who understood why the sepsis alert threshold was set where it is has resigned.
This is what institutional amnesia means concretely: healthcare preserves the documents and deletes the people who make the documents mean anything.
The negative results file that does not exist
Medicine understands publication bias in research. Everyone knows that trials showing no effect are less likely to be published, that this systematically distorts the evidence base, and that registries and mandatory reporting were created as partial correctives.
Healthcare operations has exactly the same problem, in a more extreme form, with no corrective at all.
Think about what gets presented at conferences and written up in trade press. Successes. Always successes. Nobody submits an abstract titled "Our Virtual Nursing Program Failed and Here Is Why." Nobody issues a press release about the AI tool quietly switched off after eight months. The vendor certainly does not.
So the information environment that a CNO is making a multi-million dollar decision inside is composed almost entirely of survivorship bias, curated by parties with an interest in the outcome.
And the evidence suggests the failures are where the information is. That JAMA Network Open commentary from December 2025, drawing on roughly 900 bedside nurses across ten states, found over half reported no change in workload from virtual nursing, and concluded that effects "depend on local implementation and environment."
That sentence is the whole ballgame. If effects depend on local implementation, then the only knowledge worth having is implementation-level knowledge held by people who implemented, in settings comparable to yours. That is exactly the knowledge that never gets published, never gets attributed, and evaporates when the operator moves on.
The negative results file of US hospital operations is the single most valuable unwritten document in the industry. It cannot be written by institutions, because institutions cannot publish their own failures. It can only be written by individuals, and only if those individuals have somewhere to write it that follows them across employers.
The attribution problem
There is a reframe here worth stating clearly, because it changes what you would build.
Healthcare does not primarily have an innovation problem. It has an attribution problem.
The innovations exist. They are running right now, in hundreds of places, generating real evidence about what works in what setting. The failure is that none of that evidence is attached to a findable person.
Compare it to how other fields solved analogous problems. Open source software has commit history: you can see who wrote a piece of code and contact them. Academic research has authorship: imperfect, gamed, and still enormously better than nothing. Both fields made contribution personally attributable, and both consequently developed functioning expert-location systems as a byproduct.
Healthcare operations attributes everything to institutions. "Cleveland Clinic implemented X." That sentence is useless to the person who needs to know what actually happened, because Cleveland Clinic cannot take your call. A named person can.
What a functioning implementation record would look like
The design follows directly from the diagnosis.
Keyed to the person, not the institution. A record that says "I ran this, here, then" and that travels with the individual through every subsequent job. This single property is what makes it survive turnover, and it is the property no employer-owned system can have.
Including abandonment as a first-class outcome. Scaled, sustained, modified, abandoned. Abandoned entries are the most valuable in the entire dataset, which means the system must make recording them low-cost and reputationally safe.
Structured enough to search, loose enough to be honest. Intervention, setting, size, EHR, timeframe, role held, outcome. Enough structure to answer "who has done this in a 150-bed community hospital on this EHR," without demanding a formal write-up nobody has time for.
With a willingness-to-talk flag. The record's purpose is to produce the twenty-five minute phone call. Everything else is metadata.
Respecting real confidentiality boundaries. Members share their own professional experience at the level their employer permits. No protected health information. No vendor-confidential material under non-disclosure. No pricing discussion, which is both an antitrust concern and unnecessary. The lesson is not the contract.
With no vendor access. The moment vendors can see or shape the registry, it becomes marketing and its value collapses. This is the design decision that determines whether the thing is trustworthy in year three.
None of this is technically difficult. The entire difficulty is that it has to be owned by the people who generate the knowledge rather than by the institutions that employ them or the vendors that sell to them. Which is precisely why none of the existing players has built it: every one of them is one of those two things.
What to do this quarter
If you are about to run a pilot
Write the negative results document at the start. Before you begin, write down what would constitute failure, what you would need to observe to stop, and who is authorized to call it. Most abandoned pilots are never recorded as abandoned because no definition of failure was ever agreed, so the program merely fades. A pre-registered stopping rule is the cheapest research hygiene available and almost nobody in operations uses it.
Name the operator in the record. Not the sponsoring executive. The person who actually ran it. Put their name on the internal write-up. This is how you make your own institution's memory personally attributable.
Ask vendors for a customer who stopped using the product. This request is revealing regardless of the answer. A vendor confident in their product can usually produce a churned customer with a legitimate reason. One that cannot has told you something.
Do a genuine exit interview when the operator leaves. Not the HR exit interview. A structured knowledge handover on every program they ran: why the thresholds are where they are, what broke, what the workarounds are. Ninety minutes, recorded, indexed. Almost no organization does this and every organization should.
If you are evaluating technology
Find one unofficial reference. Vendor references are selected. Spend the effort to find one person, through your own network, who used the product and was not offered to you by the vendor. That single call routinely outperforms the entire structured evaluation.
Ask about the workarounds specifically. "What did your staff start doing that you did not design?" That question surfaces more truth than any feature comparison.
Compare evaluation designs before comparing results. Given that six major systems produced non-comparable ambient AI evaluations, the first question with any peer's results is what exactly they measured and against what baseline.
If you lead a system
Count your active pilots. Many executives cannot produce this number. It is usually much larger than expected and contains duplicates.
Ask what happened to the pilots from three years ago. If nobody can say, you have discovered your institutional memory problem, and it is not unique to you.
Publish something that failed. Genuinely. The first health system to publish a serious, non-defensive account of an abandoned major implementation will get more credibility from peers than a decade of success announcements, and will make it easier for everyone else to do the same.
Frequently asked questions
How often do healthcare improvement initiatives actually spread? Rarely. Research in Health Research Policy and Systems found fewer than 40 percent of healthcare improvement initiatives transition to sustained spread beyond one area of one organization, with manager turnover named among the barriers. A BJA Education review reported that up to 70 percent of attempted organizational change fails.
Why do health systems all evaluate the same AI tools separately? Because evaluation experience is not shared even when frameworks are. All 43 systems in a 2025 JAMIA survey had ambient notes adoption activity. Six major systems examined in a 2026 AHA analysis each ran separate evaluations with non-comparable metrics. Standards bodies publish documents; nobody publishes the name of the person who ran the pilot and found the failure mode.
Is AI governance in health systems mature? Generally not. HFMA research found 88 percent of 233 health systems using AI but only 18 percent with mature governance and a fully formed strategy. A scoping review of 77 governance frameworks found only 19.5 percent specified an oversight mechanism. A Doximity survey found only 8 percent of physicians said their organization's AI decision process was clear to them.
Does virtual nursing work? The honest answer is that it depends heavily on implementation. A JAMA Network Open commentary drawing on roughly 900 bedside nurses across ten states found more than half reported no change in workload, and concluded effects depend on local implementation and environment. That conclusion is precisely why peer implementation knowledge matters more than published effect sizes.
Why can't peer networks like Vizient or the Scottsdale Institute solve this? They are valuable and structurally limited. Membership is institutional rather than personal, so knowledge does not follow the individual through job changes. Non-attribution norms that enable candid discussion also prevent later follow-up with the person who knew. Meetings are scheduled while questions arise continuously. And the operator who ran the pilot is often not the executive holding the membership.
What is institutional amnesia in healthcare? It is the routine loss of the tacit knowledge that explains an organization's own documents and systems: why a protocol step exists, which metric misled, what the workarounds are, why a program was really stopped. With RN turnover at 17.6 percent and average bedside tenure around 6.8 years, this loss is continuous, silent, and unmeasured.
The bottom line
Somewhere in the United States right now, a nurse manager is running the pilot that a chief nursing officer four states away is about to authorize. One of them knows what is going to happen. The other is about to spend eighteen months and a large budget finding out.
They will never speak. Not because either is unwilling, but because the profession has never built the thing that would let them find each other.
Meanwhile more than 60 percent of improvement work fails to spread, hundreds of governance committees are writing the same policy in parallel, six major systems evaluated the same technology in ways that cannot be compared, and the people who learned the real lessons are quietly changing jobs and taking those lessons with them.
Healthcare does not have an innovation problem. It runs enormous numbers of experiments. It simply throws away the results of most of them, because the results live in people rather than documents, and nothing in the industry is designed to keep track of people.
The most valuable document in American healthcare operations is the one nobody has written: the honest, attributed, searchable record of what has already been tried and what actually happened.
It could be written. It would just have to belong to the people who ran the pilots, not to the organizations that employed them at the time.
Part of a series on the missing professional infrastructure of healthcare. Previously: You Were Served. Now You Cannot Talk to Anyone.
Evidence note: sources include Health Research Policy and Systems (2019) on improvement spread, BJA Education (2018) on organizational change failure, JAMIA (2025) on health system AI adoption across 43 systems, HFMA research on AI governance maturity, npj Digital Medicine (2026) on governance framework review, AHA Market Scan (2026) on ambient AI evaluations at six named systems, JAMA Network Open commentary (2025) on virtual nursing, NSI Nursing Solutions (2026) on turnover and tenure, and Advisory Board reporting on hospital CEO departures. The duplicated labor estimate is arithmetic from stated assumptions, clearly identified as an estimate rather than a measured figure. Some counts of US health systems are approximate.
