Every retirement is a library that burns. The person who knows why the mutual-aid agreement reads the way it does, which vendor's invoices deserve a second look, how the last flood got fought street by street: the morning they clean out their desk, the institution loses all of it. The binder they leave behind answers none of the questions that matter, because the binder was always the least of what they knew.
Organizations lose this knowledge on a schedule they can see coming and mostly ignore. The most experienced people in a county, a clinic, a utility, or a machine shop are, as a rule, the closest ones to retirement. The industry keeps selling these institutions a chatbot to draft their emails faster. The knowledge quietly walking out the door is worth more than every email they'll ever write.
What leaves when they leave
The part of the job that made it into the SOP is the part that was easy to write down, which is exactly why it survives. The loss that hurts is the other kind. Michael Polanyi named it tacit knowledge and put it in six words: "we know more than we can tell." It's the pattern the claims adjuster reads on the first pass that says this file is fraud and that one's just a bad week. It's the feel for which rule bends under pressure and which one snaps. It's the system operator who hears the substation before the alarm board does, and the machinist who knows why that fixture runs backwards and always has. None of it lives in the documentation, because it barely lives in language.
That's the knowledge that runs the institution, and it's the knowledge nobody captured, because capturing it was always somebody's someday project, and someday doesn't survive a retirement party.
The fixes everyone tries, and why they don't take
The sector's answers to this are old and thin. There's the exit interview, scheduled two weeks out, as if thirty years of judgment downloads over a conference-room lunch. There's the job-shadow, which works right up until the expert retires mid-lesson. There's the knowledge-management initiative that produces a wiki nobody updates and nobody reads. Every one of them recovers a rounding error of what's leaving, and anyone who's run one knows it.
I came to this work out of emergency management, where the loss isn't a bullet on a strategy slide. It's operational, and it compounds with every separation the HR office processes. Lose the emergency manager who fought the last three fires in that canyon and you haven't lost a headcount. You've lost the map.
What AI changes about the arithmetic
The thing that shifted isn't the model getting smarter. It's that a system sitting inside the workflow while the expert is still running it can catch the work in working form. Not a deathbed download two weeks before the party. A living record that builds up across months of ordinary operation: the intake logic, the drafting patterns, the exceptions and how each one got handled, gathered while the person does the job the way they really do it and not the way the manual pretends they do.
Tuned on the institution's own corpus, a local model turns that record from a dead archive into a conversation. The new hire on their worst Tuesday asks the institution itself how this used to get done, and gets an answer with citations back into the real record, instead of a shrug from three desks over. That's what capturing institutional knowledge looks like when it's built into the work instead of bolted on after the fact.
The line between capture and surveillance
This capability has an evil twin, and the difference between them is the whole game. A system that watches an expert to preserve what they know and a system that watches an employee to score what they produce can run on the same log format. What separates them is design, stated out loud and enforced.
The staff know what's captured and why. The record belongs to the institution, not to a vendor holding it on someone else's servers and not to a manager quietly building a case. And nothing captured turns into a productivity ranking of the people who fed it. Get that wrong and you don't get an archive. You get a workforce that gives the thing exactly as much as it takes to look cooperative, and the real knowledge walks out on schedule anyway.
Dignity isn't a nicety you add if there's budget. It's a design requirement here, the same standing as the security posture, and for the same reason: the system doesn't work without it. People protect what they know when they suspect it'll be used against them, and they're right to.
What you don't get
It captures a fraction. The judgment retires with the person, the way it should; no system inherits thirty years of instinct, and any vendor promising that is selling you the brochure. What the institution keeps is the context that judgment ran on, which happens to be most of what the successor is missing on the hard days. Smaller than the pitch. More valuable than it sounds.
The archive is the treasure, and a mess in stays a mess. If the records are chaos and the expert's real reasoning never makes it into the corpus, the model has nothing worth being tuned on. What comes out tracks what the institution was willing to gather.
And it can't be a second job. The moment knowledge capture becomes a documentation duty stapled onto an already-full day, it dies, and it deserves to. It works as a byproduct of the work, with deliberate depth added on the expert's own schedule and under the expert's own review, or it doesn't work at all.