Tacit knowledge: what it is and how to capture it

There's a quick test for whether this affects you. Think of the person on your team who gets asked the most. Now imagine they don't come in tomorrow, or the day after, or for three months. Make the list of what would stop. If that list is long, almost nothing on it is written down anywhere in your company. What isn't written down has had a name for more than sixty years: tacit knowledge.

It's the least-discussed topic in knowledge management, and the only one that really matters. Manuals get copied. Processes get inherited. What doesn't transfer is judgement, and judgement is what separates a company that works from one that keeps putting out fires.

What tacit knowledge is

It's the knowledge a person uses to do their job well but has never written down, and which in many cases they wouldn't be able to explain fully even if you asked. The judgement to decide, the context of why things are the way they are, the shortcuts, the map of who to call. The term was coined by the philosopher Michael Polanyi, who summed it up in one line: we know more than we can tell.

The classic example is riding a bicycle: you can do it perfectly and you couldn't write a manual that would help someone who has never got on one. In a company the equivalent is less showy and much more expensive. It's the salesperson who knows, from the tone of an email, that this client is going to ask for a discount next week. It's the technician who hears the problem over the phone and already knows which of the three possible causes it is. It's the person in finance who knows which invoices can be chased by phone and which need writing to for the record.

None of those three people has ever written that down. None of them considers it «knowledge». To them it is, simply, doing their job.

Tacit and explicit: the difference that decides what you can document

Explicit knowledge is already codified and can be passed to another person by handing it over: a manual, a procedure, a spreadsheet, a contract. Tacit knowledge cannot be handed over, only drawn out: you have to ask the right questions, watch the person who has it working, or record them while they explain it. Almost everything a company considers its documentation is explicit, and almost everything that actually makes it work is tacit.

The distinction isn't academic: it completely changes which tool you need.

Explicit. It already exists in some format. The problem is finding it, and that's why a decent search engine solves it. If your pain is «we have the information but nobody can locate it», you have a search problem.
Tacit. It exists in no format. The problem is getting it out, and no search engine can find what has never been written. If your pain is «only Marta knows how to do that», you don't have a search problem: you have a capture problem.

Confusing the two is the most expensive mistake in this field, and it explains half the knowledge management projects that fail. A search engine is bought for a capture problem. The tool works perfectly, indexes everything there is, and the team keeps asking Marta, because what they needed was never inside.

There's a third state worth naming, because it's where most of an SME's real knowledge lives: what has been expressed but not saved. The answer Marta gave over chat eight months ago, the explanation she gave in a meeting, the «watch out for this» said in passing. Technically it became explicit for thirty seconds. And then it was lost, because nobody was collecting it.

Why your company loses it without noticing

The loss never arrives as an invoice. It arrives diffuse: decisions that take longer, mistakes that repeat, a new hire who needs six months to rebuild judgement that already existed in the company. And the accounting records it backwards, as the newcomer's «learning curve» instead of knowledge lost from the person who left. The cost is real but it falls into another budget, later and under another name.

I'm not going to give you a market figure for what it costs. The ones circulating are hard to verify and this argument doesn't need them. The sum that matters is your company's: how many people hold the answers, how many hours a day they lose to interruptions, how long someone new takes to perform. If you want to do it with your numbers, that's what the knowledge loss test is for.

And there's an asymmetry that makes this worse than it looks: tacit knowledge is lost all at once and rebuilt slowly. An unexpected absence, a resignation, a retirement: what took six years to form disappears in an afternoon. What goes with a senior isn't recovered by hiring another senior: the new one will bring their own judgement, not your operation's.

Why documenting more doesn't solve it

The instinctive reaction is «we need to document better». A wiki is opened, ten pages are filled in enthusiastically and three months later it's out of date and nobody consults it. It happens every time, and it isn't a discipline problem in the team.

They are two distinct failures and it's worth separating them, because they have different solutions:

The maintenance one. Documenting is a separate job that competes with the real work, and the real work always wins. Any system that depends on someone finding time to feed it is dead by month three. I've developed this in AI knowledge management.
The extraction one. This is the serious one. When someone sits down to document, they write the steps and leave out the judgement — not out of laziness, but because the judgement seems obvious to them. What is evident to the expert is exactly what everyone else doesn't know. Nobody documents what they think everybody already knows.

Asking someone to write down what they know is asking them first to identify which part of what they know isn't obvious. Almost nobody can do that about their own work.

That's why the exit interview (two hours in the final week, when the person's head is already out the door) collects four urgent things and loses everything else. And that's why the problem isn't fixed with more willpower or a better template.

How it's really captured: four methods

These four work, and none requires buying anything to start. They're ordered by the ratio between what they cost and what they rescue.

1. The rescue interview, recorded and not written. Forty-five minutes with your most interrupted person, with a single instruction: have them go through the ten questions they get asked most and how they answer them. Record it. Don't ask them to write anything: writing is the step that never happens. The transcript of that recording is already knowledge rescued, and forty-five minutes produces more than six months of wiki.
2. Have someone watch the person who knows do the work. It's the oldest method there is, the apprentice beside the master, and it's still the only one that transfers what the expert doesn't even know they know. Nonaka and Takeuchi called it socialisation in their knowledge creation model, and it remains expensive in hours and hard to scale. Save it for the genuinely critical.
3. The «write it once» rule. Any question answered for a second time over chat gets written somewhere with author and date. One rule, zero new tools. The second time is the signal: if two different people have needed the same thing, there will be a third. It's the cheapest thing on this list and the one most people skip.
4. Teach on video instead of writing. Whoever explains best by showing (an ERP, a CRM, a process with fifteen clicks) is never going to write that document. But they will record five minutes of screen talking through it. Video captures the tone, the hesitations and the «watch out for this» that a manual loses. And it's five minutes against the two hours of the document that, honestly, was never going to be written.

What all four have in common: none of them asks anyone to sit down and document. They all capture while the person does something they were going to do anyway, or nearly.

What part of this AI can do, and what it can't

It's worth being honest here, because this is where there's most hot air right now. No AI can extract what nobody has ever expressed. If the knowledge hasn't come out of a head, there's nothing to capture, and connecting more integrations doesn't change that: what isn't written isn't in Slack, or Drive, or your document manager. Anyone selling you «connect your tools and capture your company's knowledge» is selling you a search engine for a capture problem.

What AI does solve well is the surrounding work, which turns out to be nearly all the work:

Transcribing. The bottleneck in the rescue interview and in video was never recording, it was turning it into text and organising it. That is now free and automatic.
Collecting what was said once. The answer someone gave over chat is explicit knowledge for thirty seconds. A system that saves it with its author and date turns everyday conversation into memory.
Detecting the gap. This is the interesting and least obvious part: a system that answers can notice that it doesn't know something. If three people ask the same thing and there's no material to answer with, that's a signal that tacit knowledge with real demand exists, and it also tells you the topic. It's the closest thing to knowing what to ask.
Asking the right person. Once the gap is detected, the only way to close it is to go to whoever knows. Automating that circuit (gap, question to the expert, review, and notifying whoever asked) is what makes the base grow on its own instead of dying.

That is exactly the architecture Savia is built on: capture what is already being said, always cite where each answer came from, and when something isn't covered, don't invent it — log it and go and find the expert for that area. The human part is still human. Someone has to talk. What changes is that now talking once is enough.

Where to start this week

Without buying anything and without a project:

Make the list of your «only ones». One sheet, two columns: critical process and who is the only person who masters it. Every row with a single name is a point of failure. It takes an hour and shows you exactly where it hurts.
Document by pain, not alphabetically. Start with what stopped the last time someone was missing. That process, first.
Book a rescue interview. One, this week, with the person on the first row of your list. Forty-five minutes, recorded. It's the step most people postpone indefinitely and the only one that produces a result the same day.

Tacit knowledge doesn't disappear because nobody cares. It disappears because it's invisible while the person who has it is still sitting in their chair, and it becomes visible the day they get up. Everything above amounts to not waiting for that day.

Frequently asked questions

What is tacit knowledge? It is the knowledge a person uses to do their job well but has never written down, and which in many cases they wouldn't be able to explain fully even if you asked. The judgement to decide, the context of why things are the way they are, the shortcuts, the map of who to call. The term was coined by the philosopher Michael Polanyi, who summed it up in one line: we know more than we can tell.

How does tacit knowledge differ from explicit knowledge? Explicit knowledge is already codified and can be passed to another person by handing it over: a manual, a procedure, a spreadsheet, a contract. Tacit knowledge cannot be handed over, only drawn out: you have to ask the right questions, watch the person who has it working, or record them while they explain it. Almost everything a company considers its documentation is explicit, and almost everything that actually makes it work is tacit.

Why isn't documenting more enough? For two reasons. The first is that documenting is a separate job competing with the real work and it always loses, so the wiki is opened enthusiastically and three months later it's out of date. The second is deeper: when someone sits down to document they write the obvious part, the steps and the processes, and leave out precisely the valuable part, the judgement and the why. It isn't negligence: nobody knows what to ask to get it out, starting with the person who knows it.

Can artificial intelligence capture tacit knowledge? Partly, and it's worth being honest about the limit. No AI can extract what nobody has ever expressed: if the knowledge hasn't come out of a head, there is nothing to capture and connecting more tools doesn't change that. What AI does do well is the surrounding work: transcribing what is said in a recording, remembering what was answered once in a chat, detecting that there is a question nobody can answer and alerting the person who does know. The human part is still human: someone has to talk.