Ask an experienced person how they know, and sometimes the answer disappoints.

“The sound is wrong.”

“The paragraph is leaning.”

“You can feel the timing.”

“It looks fine, but I would not trust it.”

These statements can sound mystical or evasive. Often they are compressed reports from years of perception.

Instructions end before judgment

Procedures are powerful because they make action repeatable. They allow knowledge to travel beyond the person who discovered it.

But procedures cannot specify every relevant feature of every future situation. At some point, the worker must decide which detail matters now.

This decision is tacit knowledge: perception shaped by repeated exposure, feedback, and consequence.

AI can produce instructions at extraordinary depth. It can explain common exceptions and generate examples. Yet the final situation still contains texture the instruction did not name.

Output hides perception

When AI produces a finished artifact, observers see the result without seeing what a skilled person would have monitored while making it.

A generated design may look polished but violate a physical constraint. A generated policy may be coherent but ignore the institution’s history. A generated repair plan may omit the one sound that would concern an experienced technician.

The missing element is not always information. It is salience—the ability to know which information deserves attention.

Tacit knowledge is trained by consequence

A person’s judgment becomes reliable partly because errors cost something. The material breaks. The customer returns. The audience becomes confused. The experiment fails.

Consequences update attention.

A model can learn from large patterns, but the user interacting with it may avoid direct contact with consequence. If the system completes the task, the user receives output without the feedback loop that would have trained perception.

This is how capability can rise while personal skill remains flat.

Capture stories, not only rules

Organizations often try to preserve expertise by asking experienced workers to document procedures. The result is useful and incomplete.

A better method also captures cases:

• a time the standard rule failed;
• the earliest sign that something was wrong;
• a mistake that looked correct at first;
• a tradeoff that cannot be optimized universally;
• a detail novices consistently ignore.

Stories preserve the shape of judgment better than abstract rules alone.

AI can help interview experts and organize these cases. It should not flatten them into one universal checklist.

Keep experts near the edge cases

Automation usually handles common cases first. This can make expert work more difficult because experts receive a higher concentration of exceptions while getting fewer routine opportunities to maintain basic familiarity.

Systems should preserve deliberate contact with ordinary cases, simulation, and review. Otherwise expertise becomes an emergency role separated from the material conditions that created it.

The value of saying “I cannot fully explain it yet”

Modern organizations prefer knowledge that can be documented, transferred, and measured. Tacit knowledge can appear suspicious because it depends on the person.

That dependence creates risk, but pretending the knowledge is fully explicit creates a different risk.

An honest system allows experts to mark uncertainty and incomplete articulation. This invites observation and training rather than forcing false precision.

Work is more than the artifact

AI will translate more craft into instructions, examples, and generated output. This is useful. It may also reveal more clearly what resists translation.

The untranslatable part of work is not magic. It is the accumulated ability to notice, feel, and care about differences before they become failures.

A prompt can request the result. A craft is the history of learning what the request forgot to mention.

Tacit knowledge appears at the edges

Ask an experienced worker what they notice first. The answer is often not a rule but a deviation: a sound that is slightly wrong, a customer who is agreeing too quickly, a paragraph whose logic is correct but whose confidence feels borrowed. Expertise lives in sensitivity to these edges.

AI can help name patterns after enough examples, yet the act of collecting examples may miss what practitioners have never needed to articulate. A system can reproduce the visible move while lacking the history of consequences that made the move meaningful.

Use the tool as an interviewer

Rather than asking AI to replace a craft, organizations can use it to help experts surface their judgment. Prompt the practitioner with contrasting cases, near misses, and “what would change your mind?” questions. Record not only procedures but stories of failure and repair.

The resulting knowledge base will still be incomplete. That incompleteness should remain visible. Newcomers need access to people, feedback, and real situations—not only a polished manual.

The untranslatable part of work is not mystical. It is knowledge embedded in bodies, relationships, places, and consequences. Technology can support its transmission best when it does not pretend the transmission is finished.

The interface problem I could describe only after seeing it fail

I repeatedly asked for consistency across several workspaces: shared command headers, statistics, query toolbars, bulk actions, quick previews, and AI sidebars. The requirement sounded precise. Yet components that matched on paper could still feel wrong in use. A sidebar might be technically fixed but anchored to the viewport instead of the content area. A mobile action could exist but have the wrong visual weight. Two pages could share colors and spacing while presenting different mental models of selection.

These were not secrets hidden in the code. They were judgments formed by looking, clicking, comparing, and noticing the moment an interface stopped feeling coherent. I could write rules after the fact, but the rules did not fully contain the perception that produced them.

The same tacit layer appears in editorial work. A source can be reputable and still be wrong for a particular paragraph. An essay can meet every structural requirement and still sound mass-produced. I know the difference partly through accumulated contact with the publication: reading several pieces in sequence, hearing repeated cadence, and feeling where a sentence performs certainty rather than earning it.

AI is strongest where work can be represented. It becomes less reliable at the edge where representation depends on embodied comparison, local consequence, and a history of small corrections.

Capture the story around the rule

When I discover a design rule, I now try to preserve the case that created it. Instead of recording only “AI sidebar must be consistent,” I note what failed: on one page it covered content; on another it changed height; on mobile the close action competed with send. The story explains the purpose and helps a future maintainer handle a new edge case.

For articles, I preserve the observation behind the claim. “AI can weaken ownership” becomes more useful when tied to the moment generated code worked until a requirement changed and I could not explain the architecture. The concrete case does not prove a universal law. It gives the reader material with which to compare their own experience.

Tacit knowledge should not become an excuse for unaccountable authority. “I know it when I see it” can hide bias, gatekeeping, or poor communication. Experts should make reasoning as visible as possible, invite challenge, and test whether others can apply the rule. The boundary is that some remainder will persist even after excellent documentation.

Work is therefore more than the final artifact and more than the instructions used to produce it. It includes attention trained by consequence: the ability to notice what the checklist did not name. AI can help us document that attention. It cannot inherit the full weight of having been responsible when the detail went wrong.

Experts notice before they can explain

A technician hears a change in a motor before an instrument confirms the fault. An editor senses that a paragraph is evasive before naming the missing claim. A nurse notices that a patient’s ordinary answer does not fit their posture. These perceptions are not mystical. They are compressed histories of consequences, comparisons, and feedback.

Automation can make the final action reproducible while leaving the perception that triggered it undocumented. If organizations capture only procedures, they risk preserving what experts do after noticing and losing how they know where to look.

Build an exception library beside the manual

For difficult cases, record the situation, competing explanations, cue that changed attention, action taken, and outcome. The library should include false alarms and disagreements, not only success stories. Its purpose is to expose the boundary of the rule.

AI can retrieve similar cases and summarize patterns. A human still decides whether the resemblance is meaningful. Classic research on automation distinguishes appropriate use from misuse and over-reliance; a searchable archive becomes dangerous when similarity is treated as authority.

Train perception with consequences

Novices need more than access to expert answers. Give them cases before outcomes are revealed. Ask what they notice, what they would inspect next, and what result would change the plan. Then compare their sequence with the expert’s and with what actually happened.

This process makes tacit knowledge partially discussable without pretending it can be fully translated. The goal is not to replace apprenticeship with a database. It is to create more occasions where judgment can be observed, challenged, and refined.

The boundary of documentation

Some knowledge remains local because the environment changes faster than the manual, or because social context cannot be reduced without harm. A responsible organization marks those boundaries explicitly. “Requires experienced review” should name the reason and the available escalation path, not function as a badge of prestige.

Work is more than the artifact because competence includes knowing when the artifact is not enough.

AI can help expose tacit knowledge without claiming to own it

Interview prompts can ask experts to compare near-miss cases, explain what drew attention first, and describe a time when the usual rule failed. A model can organize those accounts and identify unresolved differences between practitioners.

The result should remain a training resource, not an automated verdict. Its value lies in making judgment discussable and giving novices more cases to inspect. The organization still needs observation, supervised practice, and feedback from real consequences.

Used this way, AI does not translate all work into rules. It helps people notice where the rules end.