Knowledge management in healthcare: what AI can actually hold

The phrase covers at least three different things inside a health organisation, and an agent that writes its own records has no business in two of them. The third has no owner at all, which is why it is where the administrative week disappears.

Knowledge management in healthcare is usually framed as the problem of getting what an organisation knows to the place where it is used. The familiar split is between explicit knowledge, the part already written down as policies, pathways, protocols and procedures, and tacit knowledge, the part that lives in people's heads: judgment, workarounds, who to call when the usual route is closed, which committee will send a submission back and for what. The long-standing complaint about the first is that it is easy to store and hard to keep current. The complaint about the second is that it determines how the work actually gets done and is almost never captured at all.

Healthcare has a structural reason for finding this harder than most industries. Care is continuous and staffed in shifts, so handover is not an occasional event, it is the basic unit of organisational memory. What the organisation knows has to survive being passed between people several times a day, across departments that keep their records separately, on a rhythm set by clinical need rather than anybody's convenience. That is a genuinely hard problem, it has a large literature, and it is not the problem this page claims to solve. What follows is narrower and it is worth being explicit about which part.

Three layers, and only one of them is loose

The first two layers have owners, oversight and systems that were chosen deliberately. The third has meeting minutes written late, a shared drive, and the memory of whichever administrator has been there longest. It is also the layer where a program stalls: not because anyone was clinically wrong, but because the decision a committee reached in March was never written down, the vendor's commitment sat in one person's inbox, and the deadline that mattered was in the body of an email nobody read twice.

What an agent can hold here, and what it must not

The concession comes first, because it is what makes the rest of it credible. No protected health information goes into this. It is not a validated clinical system, it has not been qualified or certified for anything, it is not a medical record, and it is not a clinical decision tool. Nothing about it should touch care. Any material that identifies a patient belongs in the systems your organisation already runs under its own governance, and pointing this at them would be both wrong and pointless, because the layer that is worth holding has no patients in it at all. It is contracts, committees, vendors, budgets, deadlines and promises between adults in offices.

Inside that boundary the material is real work and it is genuinely unowned. A service line that depends on outside suppliers, each with a contract, a renewal date and a service commitment somebody made verbally on a call. A capital purchase moving through three departments, where the reason a specification changed sits in a thread rather than in the paperwork. A committee that decides something and then drifts from it over two quarters because the minute was thin. A research administration thread, where a submission to a review board, a sponsor's reporting deadline and a site's outstanding document each have their own clock running, and the study data itself stays where it already is. An inspection preparation that is really a list of promises with owners and dates. None of that is clinical, and none of it is in the record system.

How it works, and where it stops

The mechanism is unremarkable, which is the point. Mail arrives from a vendor, a colleague, a sponsor, a committee chair. It is read, filed against the program or the relationship it belongs to, and the record for that thread is updated: what the message resolves gets closed, what it creates gets opened, dates that moved get moved. The next step, usually a reply, is drafted and waits for approval. Nothing outbound goes on its own. The records are plain text files on the machine already being used for the work, which is a statement about where the material sits rather than a claim about any standard.

Its limits are worth stating before anyone builds a business case on it. It runs on a schedule rather than continuously, hourly at best, and only when the machine is awake, so it will never be a live operational feed. It is one seat: one install, one owner, one mail identity. It does not assign work to anyone else and it does not notify anyone, so a commitment another person made is recorded as your own item to check, never as their task. That single-seat shape is why the honest description of it is an administrator's or a program lead's own system rather than a departmental deployment.

Which layer your question was about

If you came looking for how a health system moves clinical knowledge into practice, this is not that, and the papers on the subject are better company. If you came because you are the person holding twenty operational threads and no system holds any of them, the layer above is the one to look at. The distinction is close to the one drawn on the neighbouring page for knowledge management in pharma, where a validated system owns half the problem and nobody owns the other half. The industry-neutral form of the argument is AI knowledge management, the concept underneath it is what an AI company brain is, and the version for a profession where confidentiality drives the whole design is knowledge management for law firms. What accumulates over years, if it works, is AI institutional memory.

Two honest notes to close on. This was built by someone running his own company in life sciences, so the demo is a drug program and the expertise behind it is operational rather than clinical. And nothing here is evidence that it works in a hospital, because no health organisation is running it. What is on offer is a mechanism you can inspect and judge, which is a weaker claim than the category usually makes and the only one that is true.