Start with a real decision, ask an experienced colleague to explain the cues and exceptions behind it, and test whether another person can use that explanation on a different case. Shared practice needs reasons and limits, not just a transcript of what the expert did. AI can help organize and retrieve that material, but fluent answers do not establish that the judgment has transferred.
This is a practical task for a team lead or knowledge owner who hears “you learn that with experience” whenever a colleague asks how to handle an awkward case. Tacit knowledge is the understanding people use without fully spelling it out. Some of it can become useful guidance; some still needs observation, conversation and practice. The tools and workflows guide places this work alongside the team's wider operating decisions.
Ask about a real decision rather than everything the expert knows
A request to document everything often produces a broad procedure or a long recording. Begin with a recent case where the experienced person's choice was not obvious to someone else. Use approved material and remove information that does not belong in the learning record.
Suppose a retailer's purchasing coordinator is investigating a supplier invoice that does not match the recorded delivery. An experienced colleague suspects that the order was split across two shipments. The next useful action is to check the second delivery record with the warehouse and prepare a question for the buyer, rather than assume the invoice is wrong. The buyer retains the existing approval decision.
Ask the colleague to reconstruct that choice from the documents available at the time:
- What caught your attention? Identify the particular date, reference or missing detail.
- What else could explain it? Keep plausible alternatives beside the preferred explanation.
- What did you check next? Name the source that could confirm or contradict the interpretation.
- What would change your decision? Ask for a case where the same cue would be misleading.
- Where would you stop and ask someone? Preserve the boundary of the colleague's knowledge and authority.
These questions seek observable reasoning. “I could tell it was a partial delivery” is a starting point, not a reusable instruction. The useful detail is how the person distinguished a partial delivery from a duplicate invoice, an unrecorded receipt or a genuine quantity error.
A Reddit user in a May 2026 management discussion described keeping a few sentences after nontrivial decisions explaining why one option was chosen over another. The commenter reported finding this more useful than generic training, but supplied no independently verified results. The suggestion is still practical: collect reasoning close to the work, when the alternatives remain fresh, instead of waiting for a large documentation exercise.
Preserve exceptions and disagreements in the record
An expert's account may contain a dependable rule, a local habit and an unresolved assumption in the same paragraph. Separate them before putting the material into an AI knowledge source. Otherwise, a polished summary can make all three sound equally authoritative.
For the purchasing example, a small record could distinguish the following:
| Detail captured | How the team should treat it |
|---|---|
The order permits split deliveries | Check against the actual order terms |
This supplier often sends a later shipment | A clue to investigate, not proof of receipt |
The second delivery is not recorded | An unresolved fact requiring confirmation |
Colleagues disagree about the next step | Record both reasons and refer to the process owner |
A buyer must approve the payment decision | Preserve the existing authority boundary |
Ask another experienced colleague to challenge the explanation. A disagreement may reflect different customers, products or conditions. It can also expose an outdated habit. Do not average two conflicting accounts into a confident rule merely to finish the document.
NASA technical fellow Heather Koehler describes a related discipline in a September 2026 conversation about passing down knowledge. Alongside mentoring and guided practice, she emphasizes understanding the data, assumptions, applicability and uncertainties behind engineering methods. This is NASA's account of knowledge transfer, not an AI evaluation. Its relevance here is the attention to when an established method actually applies.
For shared team guidance, retain the case, reasoning, scope, source and reviewer. Mark disputed material as unresolved and keep it out of approved instructions until the responsible person decides. Use the team AI agreement to make that review responsibility clear.
Test whether a colleague can use the guidance
A document can accurately reflect an expert's explanation and still fail to help someone else. Test the guidance through a fresh, approved example. Ask a colleague to explain what they notice, what they would check and where they would seek help before showing the expert's answer.
In the retailer's exercise, use another invoice discrepancy with different circumstances. Perhaps the order allows split deliveries, but the warehouse confirms there was only one shipment. Does the learner investigate the missing quantity, or copy the earlier explanation because it appears in the guidance?
Review the exercise for four things:
- Relevant cues. Did the colleague notice the detail that changes the case?
- Source checking. Did they consult the right record instead of filling a gap with an assumption?
- Exceptions. Could they explain why the earlier answer might not apply?
- Escalation. Did they stop at the right boundary and identify who could decide?
Use mistakes to improve the explanation. The expert may have omitted a familiar abbreviation, a source location or a condition they normally recognize without thinking. Repeat the affected exercise after repairing that gap. The number and variety of cases should reflect the work's consequences; one successful example does not establish general competence.
Give AI a bounded supporting job
Once people have reviewed the material, AI can help retrieve related cases, suggest questions or draft a comparison of the available evidence. Tell it to distinguish source-backed facts from unresolved information and to point to the underlying records. A person still needs to check whether those records support the proposed next step.
There are real attempts to support this kind of capture. The August 2026 AquiLLM paper describes an AI architecture used in a UCLA astrophysics research group, including collections with project instructions and practical documentation. Its authors explicitly say that the current approach has not yet been benchmarked for tacit knowledge in informal material such as meeting minutes and team messages. A working collection and retrieval system therefore should not be presented as proof that expert judgment has been preserved.
Also check who benefits and who is being constrained. In the November 2024 version of Generative AI at Work, researchers studying customer-support agents found different effects by experience and skill. Less-experienced workers benefited more, while the most-skilled group saw small declines in quality measures. This is evidence from one support setting, not a prediction for purchasing teams. It challenges the assumption that a shared AI answer should become every expert's default.
For the retailer, an approved assistant might assemble the invoice, order terms and receipt references into a draft discrepancy note. It should leave the missing delivery confirmation visibly unresolved. It should not turn “this supplier often splits shipments” into “the goods were received,” or take over the buyer's approval. Test those specific failures before relying on the assistant.
Keep the guidance connected to the work
Give the reviewed record a named owner and a clear place in the team's existing knowledge system. Keep the original evidence accessible to the people who need it, with appropriate permissions. The handoff guide explains how to preserve usable sources and review status when work crosses tools.
Acknowledge the contributors in the internal record in the way they agree to. Make clear which parts they reviewed, and allow them to correct the account. Asking for expertise should include time for that work, rather than quietly adding documentation to an already full workload.
Update the guidance when a new case exposes a missing condition, a source changes or the process owner changes the approved method. Retire obsolete examples from active instructions while retaining required history. A growing folder is not necessarily an improving practice.
The useful result is a colleague making a better-supported decision with an appropriate route for uncertainty. Record that evidence directly. Counts of interviews, uploaded documents or generated answers describe activity; they do not establish that the team can handle the work.
Questions about sharing expert judgment
Is recording an expert interview enough to capture tacit knowledge?
No. A recording preserves what the person said, but may omit cues they did not think to mention or conditions they assumed the listener understood. Start with a specific decision, ask about alternatives and exceptions, then check whether another colleague can apply the explanation to a new case. Use the elicitation questions to make the interview more useful without treating it as complete capture.
What should we do when experts disagree?
Record each explanation and the conditions under which it applies. Compare the accounts with the relevant evidence and existing policy, then ask the accountable process owner to resolve any operational conflict. Keep unresolved material out of approved AI instructions; a generated compromise can conceal the disagreement. The record structure helps separate a useful clue from a confirmed fact or authorized rule.
How can we tell whether knowledge has transferred?
Ask a colleague to handle a different, approved example and explain their reasoning. Look for the relevant cues, source checks, exceptions and escalation boundary, not just an answer matching the expert's. Repair missing guidance and repeat affected checks. Use cases that represent important variations in the work, and avoid claiming general competence from one success. The transfer exercise shows how to make that assessment concrete.
Can an AI knowledge base replace experienced colleagues?
An AI knowledge base can make reviewed explanations and examples easier to find, but access to those materials is not proof that it can exercise the same judgment. Keep experienced people involved in reviewing difficult cases, updating limits and testing the assistant's output. Begin with a bounded supporting task, such as assembling evidence for a purchasing discrepancy, while preserving the existing decision owner. See the AI supporting role for the evidence and limits behind that choice.
