Tacit AI: what tacit knowledge in the AI layer is, and why it matters
What is tacit AI?
Tacit AI is the practice of giving AI agents the tacit knowledge of the people they work alongside. That means the judgment, the exceptions and the unwritten rules that never made it into a document. There is no single industry definition of the term, so this is how we use it.
The word is tacit, as in understood without being said. People also search for it as tactic AI, which is a common misspelling. A tactic is a plan of action. Tacit knowledge is what you know and have never spelled out.
The idea has a long history. The philosopher Michael Polanyi introduced it, and in his 1966 book The Tacit Dimension he put it this way: we can know more than we can tell. An experienced account manager can sense when a renewal is at risk. A finance lead knows which invoices to chase by phone. Ask either to write down how, and you get a fraction of it. For background, see this overview of Polanyi and tacit knowing.
Tacit knowledge is the opposite of explicit knowledge, which is written down: SOPs, contracts, tickets, reports. AI agents handle explicit knowledge well, because it is text they can read. They miss tacit knowledge entirely unless someone captures it and hands it over. Tacit AI is the work of doing that.
Where does tacit knowledge sit in the AI stack?
Think of an AI agent as a stack of layers, each holding a different kind of knowledge.
| Layer | What it holds | Example | Where it comes from |
|---|---|---|---|
| Model | General knowledge and reasoning | How to write, summarize, calculate | The model provider |
| Documents and retrieval | Explicit knowledge | SOPs, contracts, tickets, price lists | Your team, written over years |
| Context layer | Structured, current business context | Approval limits, client terms, approved sources | Your systems and your rules |
| Tacit layer | Judgment, exceptions and the reasons behind them | This client gets a call before every invoice | The people who do the work |
Most teams start with the first three. They pick a model, point it at their documents, and add retrieval so the agent can search them. That gets you an agent that can answer from the written record. For why that stops short, see why vector search is not enough for agent grounding and what a context layer for AI agents is.
The tacit layer is the easiest to skip, because it is the hardest to build. The knowledge is not in a system. It is in people, and getting it out takes conversation, not an export. In practice it is part of the context layer, the part that lives in people's heads.
Why does tacit knowledge matter more for AI agents than for people?
A new hire without tacit knowledge makes mistakes, but they also fix themselves. They ask questions, notice when a case looks odd, and learn from the correction. After a few months, they have picked up much of what the team knows.
An AI agent does none of that on its own. It rarely asks, it does not notice what it has not been told, and it answers with the same confidence whether it is right or wrong. That difference leads to three problems.
- The failures are quiet. The output looks right. The invoice goes out, the proposal is well written, and the mistake only shows up when someone who knows better reads it.
- Mistakes repeat at volume. A person gets a case wrong once and learns. An agent gets it wrong every time the case comes up until someone fixes the context.
- People end up as checkers. If every output needs a knowledgeable person to read it, the agent has added a second job rather than removing the first. We cover this pattern in why AI agents fail at work.
There is also a strategic argument. Writing in California Management Review in March 2026, Teresa Tung and Philippe Roussiere of Accenture argue that tacit knowledge is a company's next competitive moat. Any competitor can use the same models. What they cannot copy is how your experienced people decide. Their advice includes mapping where that knowledge lives and codifying it without flattening it, meaning without stripping out the nuance that makes it valuable.
What does tacit AI look like in practice?
The three examples below are illustrations, not case studies. Each shows the same pattern: the written record is correct, and the agent still gets it wrong.
| Situation | What the agent does with documents alone | What changes with the tacit layer |
|---|---|---|
| An invoice is ready to send. The SOP says Finance approves it. | Sends it after Finance approves. | Knows this client is paid only after an account lead phones first, and stops to ask. |
| Three documents describe the refund process. | Picks the top search result. | Knows two are outdated and which one the team really follows. |
| A long-standing client asks for a renewal quote. | Applies the standard price and uplift. | Knows legacy accounts have a capped uplift, and quotes accordingly. |
In each case, the missing piece is not a document. It is a decision someone makes without thinking, and that nobody has written down.
How do you capture tacit knowledge for AI?
The methods are old. What has changed is how fast they can run, and where the result goes.
- Structured interviews ask experienced people about decisions, not tasks. Which source do you trust, and why? When do you skip the standard process? Who do you check with first?
- Shadowing means watching someone handle a live case and asking why at each fork.
- Decision logs record the reasoning at the time a decision is made, when memory is fresh.
- Analysis of past work looks at documents, email and chats to find where undocumented steps and exceptions probably sit.
- AI interviews run the interview with an AI, after it has read what is already written, so the questions go to the gaps and not to what is documented.
Dactic takes the last route. You connect your sources, answer a 5 to 7 minute AI interview, and deploy your first agent with one click. Nothing deploys until you approve it. For the full method, including the questions that surface the why, read how to capture tacit knowledge with AI. If you are choosing among tools, see tools for capturing tribal knowledge before a key employee leaves.
How do you keep tacit knowledge current?
Tacit knowledge decays. A client renegotiates, a price changes, a new approver joins, and the unwritten rule you captured in spring is wrong by autumn. A one-time capture is a snapshot, and snapshots go stale.
Plan for a refresh cycle from the start.
- Name an owner for each process, with a review date.
- Re-run the interview when something big changes, not only on a calendar.
- Trace every agent mistake to a missing or stale piece of context, and fix it at the source.
In Dactic, context refreshes on a schedule: every 30 days on Launch, every 14 on Growth, every 7 on Pro, and daily or continuously on Enterprise. When your business changes in a big way, Dactic runs a short follow-up interview.
What can go wrong with tacit AI?
Capturing tacit knowledge carries real risks, and they are worth naming.
- Flattening. Turning judgment into a rule can strip out the nuance. A rule like "always call first" is wrong for the client who hates phone calls. Capture the reason, not just the rule.
- Misrepresenting the person. A capture is only as good as the interview. Let the expert review and correct what was recorded.
- Privacy. Capture can touch email and files. Check what a tool can access and how long it keeps data. Dactic does not store your raw data, which is processed ephemerally. Whatever tool you use, involve your legal and HR teams.
- Over-trust. Captured knowledge is a starting point. Keep a person approving what agents do until the results have earned trust, and keep the context readable and editable so you can see why an agent acted.
How do you start?
You do not need a program. Pick one task your best person does better than anyone else, then run these steps.
- Ask five questions about it. When three sources all look right, how do you pick one? When do you skip or change the standard process? Who do you check with before this goes out? Who reads the result, and what do they look at first? What would a smart new hire get wrong in their first month?
- Check whether your AI agent has the answers.
- Wherever it does not, you have found the gap and the first thing to capture.
The AI Agent Context Checklist turns this into a scored exercise across five areas, with a 30-day plan for closing the gaps it finds.