Why Undocumented Know-How Is Becoming Your Biggest AI Bottleneck

By Victor Hoang · Updated Oct 10, 2026

Every executive running an AI rollout eventually hits the same wall. The tool is licensed, the pilot team is keen, the model is plenty capable. And the output is still generic, slightly wrong, or confidently off. The usual diagnosis is "we need a better model" or "we need better prompts." Often the real problem is simpler and harder to fix: the thing the AI needs to know was never written down.

Our view is that AI turns undocumented know-how from a quiet inefficiency into the binding constraint on what AI can do for you. A model can only act on what is in front of it. If the real way your company qualifies a deal, approves an exception, or handles a difficult customer lives in the heads of five senior people, no model will reproduce it. That makes documenting how work is really done an AI investment, not an administrative chore. This piece explains why, how to tell which knowledge to capture first, and what to do in the next 90 days.

Key Facts

• Michael Polanyi's phrase for the core problem, we can know more than we can tell, comes from his book The Tacit Dimension. Much of what experts do well resists being put into words.
• In a study of 5,179 customer support agents, an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal impact on experienced workers. The authors describe suggestive evidence that the model disseminates the best practices of more able workers.
• In a randomized trial by METR, 16 experienced open-source developers took 19% longer on 246 real issues in repositories they knew well when AI tools were allowed.
• Anthropic's engineering team defines context as the set of tokens included when sampling from a model. If the knowledge isn't in the context or in a source the model can retrieve, the model can't use it.

The Constraint Moved

For two decades, the limit on what software could do for a company was usually the software. Buying a better system helped.

Generative AI changes where the limit sits. The model can already draft, summarize, classify, and reason over a contract. What it can't do is know your company. It doesn't know that "approved" in your legal workflow means two different things, or that a supplier named in the contract hasn't shipped on time in a year.

That knowledge exists. It just isn't anywhere a model can reach. It's in a senior ops manager's judgment, an old Slack thread, a spreadsheet one person maintains. We've made a related argument about what makes an AI strategy survive model churn: models are rented and replaceable, while your own operating knowledge is the part that compounds. The corollary is uncomfortable. If your operating knowledge isn't captured, you can't compound it either.

What "Tacit" Actually Means in a Company

The word gets used loosely, so it helps to be precise. Polanyi's observation was that we can know more than we can tell: people recognize a face, ride a bike, or sense a deal is about to stall without being able to state the rules they're following. Nonaka and Takeuchi carried the idea into management, distinguishing explicit knowledge (written, codified, shareable) from tacit knowledge (personal, experiential, hard to articulate), and arguing that organizations create value by moving knowledge between the two.

In a business, undocumented know-how comes in at least four distinct forms, and they behave differently:

1. Unwritten procedure. The steps are known, just never recorded. In an illustrative case: "Ask Priya, she knows how the month-end exception works." This is the easiest to capture and the most common.
2. Judgment and thresholds. The senior rep who knows when a prospect is real, or the controller who knows which variances matter. The rules exist but are fuzzy and weighted by experience.
3. Context and history. Why a policy exists, which customer got burned, what was tried and failed. Without it, AI proposes the exact thing you abandoned two years ago.
4. Relationships and politics. Who actually decides, who needs to be consulted first, which team will quietly block it. This is the hardest to document and the last one to automate.

The first one isn't really tacit at all. It's just undocumented, and a few interviews fix it.

Why AI Makes This Urgent Now

Companies have always run on undocumented know-how. It worked because the same humans were there every day and the senior person was one message away. AI breaks those assumptions.

AI has no hallway. A new employee picks up the unwritten rules by watching, asking, and getting corrected. A model working from your documents gets none of that. Whatever isn't in its context window or its retrievable sources simply doesn't exist for it. Anthropic's guidance on building agents is blunt about this: the practical job is curating the smallest set of high-signal tokens that give the model what it needs. Curating requires something worth curating.

AI amplifies whoever wrote the documents. The best evidence we have on this comes from a field study of 5,179 customer support agents. The AI assistant raised productivity by 14% overall and 34% for novice and low-skilled workers, with minimal effect on the most experienced. The authors report suggestive evidence that the tool passes along the practices of the more able workers. That's the optimistic version of the story: when top performers' habits end up inside the system, everyone benefits. But note the precondition. The system had conversation data from skilled agents to learn from. If your best people's judgment never leaves their heads, it can't be passed along, and the lift goes to whichever competitors did capture theirs.

Experts can get slower when the tool lacks their context. In a randomized trial, 16 experienced open-source developers working on repositories they knew well took 19% longer on real issues when allowed to use AI tools. The study's authors discuss implicit requirements, such as documentation standards, testing coverage, and linting rules, that take humans substantial time to learn. It's a small sample of one profession, so we wouldn't generalize the number. But the mechanism is worth noticing: the more of the work depends on context a person carries around, the less a general-purpose tool helps, and the more it costs to review what it produces.

Agents raise the stakes. An agent that acts on its own works from whatever rules it was given, so autonomy raises the cost of every unwritten rule. We cover why the copilot-versus-agent distinction matters separately.

The Real Bottleneck Is Rarely the Model

When an AI project disappoints, leaders tend to look at the tool. A faster diagnostic is to ask four questions about the work the AI was supposed to take over:

• Could a smart new hire do this task from the written materials alone? If not, neither can the model.
• Who do people actually ask when the written process doesn't cover a case? That person is your bottleneck.
• Are the exceptions documented, or only the happy path? Models handle the happy path well. Exceptions are where the money is lost.
• If your three most senior people left tomorrow, what would break first?

The answers usually point to the same place: the organization's real operating system is oral tradition. This is also why measuring AI ROI so often disappoints in the second quarter: the easy, well-documented tasks got automated first, and the remaining high-value work is the undocumented kind.

A Decision Framework: What to Capture First

You can't document everything, and shouldn't try. Prioritize by two things: how much the work matters, and how much of it currently lives only in people's heads. The table below is a simple way to sort a process or role before you commit effort.

The ordering is deliberate. The first two rows are cheap and produce visible wins quickly, which builds the credibility you'll need to ask senior people for more of their time.

Why Documentation Is Now an AI Investment

Most companies treat documentation as overhead with no owner and no deadline. Documentation used to pay back slowly: faster onboarding, fewer repeated questions. Now it also determines what you can automate and how well. Every process you describe clearly becomes something an assistant can draft against, an agent can follow, and a reviewer can check. Every process you leave undescribed stays manual, regardless of what software you buy.

There's also a competitive case. A rival that codifies its top performers' habits lets everyone on the team work at that level. An undocumented company can't point AI at anything specific.

Documentation has limits. Some knowledge really is tacit in the strict sense, and writing it down captures only part of it. The goal is to move the codifiable majority out of private heads, mark what remains human judgment, and keep a person in the loop for that remainder. That's the same logic that underlies a minimum viable governance layer: decide explicitly where people decide, rather than leaving it implicit.

How to Capture Know-How Without Stalling the Business

The failure mode is a six-month documentation initiative that nobody finishes. Smaller and tied to a real AI use case works better.

Start from one workflow you want AI to handle. Pick a single process with clear volume and a visible cost, such as quote approvals, customer onboarding, or pipeline reviews. For a sales example, see how AI agents in the sales pipeline depend on defined stage rules and qualification criteria. Document only what that workflow needs.

Interview, then correct. Have someone else draft from a short conversation, then ask the expert to mark it up. People fix a wrong draft far better than they write from a blank page.

Capture exceptions, not just the standard path. Ask for the last 20 times the process went off-script and what the expert did. The rules behind those cases are the highest-value content in the whole exercise.

Write for a reader that has no context. Name the systems, define the terms, and state the reason behind each rule. A good test: could a competent contractor on day one follow it? That's also what a model needs. A reference like the process documentation guide covers the mechanics, and a well-formed standard operating procedure is a reasonable target format. Written rules also make AI output cheaper to check, which shrinks the verification tax on every workflow built on them.

Put it where AI can reach it. A document in someone's private drive helps nobody. Store it in a shared, searchable place with an owner and a review date. If you're building an assistant on top of it, the AI knowledge base agent article explains how retrieval from a maintained source fits together.

Treat it as a living asset. Assign each document an owner and a review cadence. A procedure that was right last year is a hazard if the business changed.

What to Do Next

Here's a sequence that fits inside a quarter.

1. Weeks 1-2: Find the dependency map. List your ten most important processes and, for each, name the one or two people who actually know how it works. The pattern will be obvious. Any process where the answer is one person goes on the priority list.
2. Weeks 3-4: Pick one AI use case and its knowledge gap. Choose the workflow you most want AI to take on. Run the four diagnostic questions above and write down what is missing.
3. Weeks 5-8: Capture the first two rows of the table. Interview, draft, correct, publish. Aim for one usable page per process, not a manual.
4. Weeks 9-12: Test with the AI and fix the documents, not the prompt. When the output is wrong, ask what a person would have known that the page didn't say, and add it. This loop is where most of the value shows up.
5. Ongoing: Make it someone's job. Name an owner, give them time, and tie new AI deployments to a documentation checkpoint. If the knowledge isn't written, the deployment waits.

One more point for leaders: make this a recognized part of senior people's roles. The people who hold the most valuable know-how are usually the busiest, and unpaid extra work doesn't happen.

The bottleneck is no longer which model you pick. It's whether your company can describe how it works. The organizations that can will find AI gets useful faster, because it has something real to work from. The ones that can't will keep tuning prompts around a gap that only people can close, and that coordination cost will keep landing on the same few senior shoulders.