How Do You Capture an Expert's Knowledge Before They Walk Out the Door?

How Do You Capture an Expert's Knowledge Before They Walk Out the Door?

You capture an expert's knowledge by recording them doing real work with their reasoning spoken out loud, then letting AI draft the playbooks, decision criteria, and judgment guides from that raw material. The expert corrects and approves what the AI produces. Start months before the departure date, because the work itself is the knowledge.

The problem

Why do handover documents fail?

Every business has someone whose judgment holds the place together. A senior operations lead who knows which supplier promise is real and which is a stall. A master technician who hears a failure coming before any diagnostic catches it. A veteran project manager who can tell which delay is noise and which is the start of a slide. A lending expert who senses a bad deal in the first conversation. A head of quality who knows exactly which complaint is the canary.

When that person leaves, the documented procedures stay. The discernment leaves with them. The next person inherits a wiki full of steps and none of the sense of when to break them.

The standard responses are the exit interview and the handover document, and they fail for the same reason. Both ask the expert to summarize. Summaries produce what the expert can consciously recall in a conference room, which is the procedure. What made them valuable was never the procedure. It was what they checked first, which exception they chased, and what a bad situation smelled like three weeks before it showed up in a report.

This is getting urgent. A large share of the most experienced operators are heading toward retirement in the next few years, and in the engagements we run, the succession conversation almost always surfaces one name the whole plan quietly depends on. The plan looks fine on paper right up until you ask what happens when that name is gone.

What you are actually losing

What is expert knowledge actually made of?

Ask a veteran what they know and you get the procedure. Watch them work and you see something else entirely. You see the order they check things in. You see which anomalies they ignore and which one makes them stop everything. You see when they pick up the phone instead of replying to the email, and who they call.

That layer is judgment, and it has an awkward property: the expert cannot simply write it down. Most of it is tacit. They do not know what they know until a live situation demands it. Ask them to "document your role" and you get thin, generic pages that read like a job posting, because that is genuinely all that conscious recall can reach. The knowledge is real. The access path is the problem.

Expertise is not a document. It is a decision-making pattern, and patterns only show up in decisions.

This is why the failure of handover documents is structural, not a matter of effort. A more diligent expert writes a longer document that fails the same way. You cannot summarize your way to judgment. You have to catch it in the act.

The method

How do you capture judgment instead of procedure?

The approach that works now has four moves, and the first one changes everything downstream: capture the expert working, not summarizing.

First, record the real work. Sit in on the actual calls, the actual reviews, the actual approvals, and have the expert narrate the reasoning out loud as it happens. Where live capture is awkward, do walkthroughs of recent real decisions instead: pull up the deal, the incident, the project, and ask what they checked first, what almost changed their mind, and what they would have done if one detail were different. Record all of it. The transcripts are the asset.

Second, feed that raw material to AI. This is where the method became practical in the last couple of years. A model can turn hours of messy narrated work into structured drafts: a playbook for the recurring situation, explicit decision criteria for the judgment call, a "how they think" guide for the next person, a catalog of the exceptions that matter and the tells that precede them. The transcription and drafting work that used to make this economically impossible is now the cheap part.

Third, the expert corrects and blesses the drafts. Not writes. Reviews. Experts are poor authors of their own thinking and excellent editors of it. Handed a draft that says "you always check the timeline against the vendor's payment schedule first," they will immediately say "yes, but only after the second milestone," and that correction is exactly the judgment you were trying to reach.

Agent-Does-the-Work is the Vista principle behind this method: the AI does the heavy documentation lift and the human validates it. The expert never faces a blank page. They face a draft of their own reasoning and fix what it got wrong, which is work a busy veteran can actually do in the margins of a normal week.

Fourth, treat the output as a living corpus, not a binder. Keep it queryable, keep it versioned, and keep adding to it while the expert is still there. Every corrected draft makes the next draft better, and the corpus compounds in a way a one-time handover never can.

If you want to see this drafting loop demonstrated live before you try it on your own expert, the free Vista AI Lab runs working sessions on exactly this kind of AI-does-the-lift workflow.

Side by side

How does working capture differ from a traditional handover?

The two approaches sound similar and produce completely different assets. The differences line up on four dimensions.

Dimension	Traditional handover	Working capture
What gets captured	Procedures the expert can consciously recall	Decisions as they actually happen, with reasoning attached
When it starts	Weeks before departure, after notice is given	Months out, while the expert is still fully in the work
Who does the writing	The expert, staring at a blank template	AI drafts from recordings; the expert corrects and blesses
What survives	A static document nobody opens after month two	A living, queryable corpus the next person and your AI tools can use

The pattern we keep seeing with the operators we advise is that the handover document gets written, filed, and abandoned, while a working-capture corpus gets used precisely because it answers the questions people actually have: not "what is the process" but "what would she have done here."

The payoff

What does a captured corpus make possible?

The obvious payoff is continuity. The next person ramps against real decisions instead of folklore, and the questions they would have asked the departed expert now have recorded answers. Onboarding stops depending on whoever happens to remember.

The larger payoff is what the corpus does for your AI systems. Generic AI gives generic answers. An assistant grounded in your expert's actual reasoning, their criteria, their exceptions, and their tells gives answers that sound like your best person on their best day. That corpus is the difference between AI that performs like an intern and AI that performs like the veteran who trained it.

The documented procedures were never the moat. The accumulated judgment was, and now it can outlast the person who accumulated it.

This is Context-as-Moat, another Vista framing, applied at the organization level: the business's accumulated context is the moat, and it should belong to the business, not to one person's memory. Owner dependency and key-person risk are, at bottom, context living in the wrong place. Working capture moves it.

There is a succession angle here that advisory readers will feel immediately. A business whose critical judgment exists as a transferable asset is simply worth more, and easier to hand over, than one whose critical judgment commutes home every night. But you do not need a sale on the horizon to justify the work. You need one load-bearing person.

Operators who want to build this muscle deliberately, with structure and other people running the same play, do it inside the Vista AI Collective, our membership for operators putting AI to work on problems exactly like this one. The capture method above is the kind of system members stand up in weeks, not quarters.

The decision rule

When should you start capturing?

Here is the rule we give operators: if one person's judgment is load-bearing and their horizon is under a few years, start now. Not when they give notice. Notice starts the clock at the exact moment access to real work begins to wind down, and the work is the knowledge. Retirement, promotion, poaching, burnout: the horizon rarely announces itself politely.

The starting move is small on purpose. This week, pick the one expert your operation quietly depends on. Pick one recurring piece of their real work, a weekly review, a deal call, an approval queue. Record one hour of it with the reasoning said out loud. Have AI draft the playbook from the transcript. Then give the expert thirty minutes to mark up the draft.

That single loop will teach you more about what you stand to lose, and what you can now keep, than any succession memo. Run it once, and the case for running it every week makes itself.

Frequently asked questions

What is working capture?

Working capture is the practice of recording an expert doing real work, with their reasoning spoken aloud, then using AI to turn those recordings into playbooks, decision criteria, and judgment guides the expert corrects and approves. It replaces the traditional handover document, which captures procedure but loses the judgment that made the expert valuable.

How long before a departure should knowledge capture start?

Months at minimum, and ideally before any departure is planned at all. Judgment only shows up in live decisions, so you need enough real work cycles for the important patterns to surface. A few weeks of hurried interviews at the end produces summaries of expertise, not the decision-making pattern itself.

Do experts resist being recorded?

Less than you would expect, when the framing is legacy rather than replacement. Most veterans want their judgment to outlast them, and reviewing an AI draft of their own reasoning is flattering work rather than a chore. Persistent resistance usually signals a trust question about how the material will be used, so answer that directly.

Can AI accurately document expert judgment?

AI drafts it; the expert certifies it. Accuracy comes from the loop: the AI produces playbooks and decision guides from recordings of real work, and the expert corrects and blesses every page before it counts. Draft errors are cheap precisely because the one person who knows the truth reviews the material while they are still around.

What if the expert has already given notice?

Start anyway and triage hard. Skip general documentation and go straight to recorded walkthroughs of their hardest recent decisions, their live reviews, and their exception calls. Even a few weeks of working capture beats a handover document, because a small corpus of real reasoning transfers judgment that a summary never will.

Who should own the captured corpus?

The business, unambiguously. Store it in company systems rather than personal drives, and treat it as an operating asset with a named owner and an update rhythm. The entire point of capturing context is that it stops living in one person's memory, so do not let it quietly start living in another's.