A maintenance team introduces an assistant that retrieves service notes and suggests possible causes of equipment faults. Newer technicians can find relevant information more quickly. An experienced colleague still notices details that are poorly represented in the records: how a vibration changes under load, which previous repair matters, or when an apparently familiar fault needs another test.
The assistant may help transfer useful knowledge, and it may reduce some of the work through which technicians previously learned. Those effects can occur together. The organisation needs to assess what the tool contributes and how people will develop the understanding required to check it, use it well, and work when it is unavailable.
Tacit knowledge describes aspects of knowing that are difficult to express fully in rules or written instructions. In organisational work, it can include practical perception, contextual judgement, and an understanding of how to investigate a situation. It should not be confused with every undocumented fact, or treated as proof that experienced intuition is always correct.
Knowledge That Is Hard to Describe Can Still Be Partly Learnable
A technician can explain a diagnostic procedure without completely describing every cue used during an inspection. Observing examples and receiving feedback can convey something that a written rule omits. That is relevant to both human learning and machine learning.
David Autor’s discussion of Polanyi’s paradox examines the gap between what people can do and what they can explicitly describe. It also discusses learning from examples as a route around some limits of explicit programming. The inability to state a rule does not establish that a model cannot learn useful patterns associated with the task.
The actual limit depends on the information and learning process available. A model given only a short service note cannot use a sound or physical observation that was never recorded. A model trained on richer information may use some of those cues, though its performance still needs evaluation in the intended setting.
An undocumented supplier restriction is a different problem. If the company avoids a part because of a verified quality issue, that restriction should not depend indefinitely on one person’s memory. Recording the decision, its basis, and its review conditions can make it available to people and systems. Calling the restriction tacit knowledge should not excuse leaving an important operating rule inaccessible.
Examine Where Expertise Actually Comes From
Experience can produce useful judgement when the environment contains learnable patterns and people receive feedback that helps them distinguish good decisions from mistakes. It can also reinforce habits that were never tested. Confidence, seniority, and a compelling account of a case do not settle the quality of the underlying judgement.
In Conditions for Intuitive Expertise, Daniel Kahneman and Gary Klein emphasise the predictability of the environment and opportunities to learn its regularities. Their analysis cautions against treating subjective confidence as sufficient evidence of expertise.
For the maintenance team, investigate which judgements experienced technicians make reliably and what evidence supports that finding. Compare their initial assessments with later inspection findings and repair outcomes. Where outcomes remain uncertain, preserve that uncertainty rather than turn an expert’s preference into a permanent rule.
The learning opportunity includes more than the final answer. A newer technician needs to see which evidence was considered, which alternatives were rejected, and what finding changed the assessment. Explaining a decision after the event can help, but it may omit information that was available during the work. Observing and discussing the investigation gives the learner a fuller reference.
AI Assistance Can Spread Knowledge and Change the Practice Available
There is evidence that assistance can help less experienced workers. In Generative AI at Work, Brynjolfsson, Li, and Raymond found productivity benefits from an assistant used by customer support agents, with effects differing across workers. The study supports useful knowledge transfer in that setting; it does not establish that every assisted workflow preserves long-term independent expertise.
For the maintenance team, faster retrieval could give technicians more time to inspect equipment and discuss difficult cases. Alternatively, a process organised around accepting suggestions could reduce the investigation they perform themselves. The tool does not determine which arrangement the organisation chooses.
Examine the tasks before and after adoption. Who forms the initial hypothesis? Who chooses the next test? Who compares the proposed explanation with the physical evidence? A reduction in manual searching may be beneficial while a loss of access to source evidence creates a problem. Counting tasks removed does not distinguish those effects.
Review work can itself be demanding and educational. Comparing plausible diagnoses, finding missing evidence, and challenging an unsafe proposal require judgement. A workflow limited to confirming that required fields are present offers a different kind of practice. Describe what the reviewer actually does before concluding that review either preserves or destroys expertise.
Design a Learning Loop Alongside the Production Workflow
The organisation can retain deliberate opportunities to make an assessment, receive feedback, and revisit it. These need not involve exposing customers or equipment to unreviewed decisions. Historical cases, supervised work, and simulations can support practice where they represent the task adequately.
A technician might record an initial diagnosis before viewing a suggestion on selected training cases. An experienced colleague can then discuss differences and identify what further evidence would resolve them. The purpose is to examine reasoning, not reward agreement with either the person or the model.
Choose a representative mix of work. Routine cases can establish foundations, while difficult cases expose the boundaries of familiar approaches. Only examining cases the assistant flags as uncertain can leave gaps in that mix. Conversely, excluding every difficult case can prevent learners from seeing how experienced staff respond when a procedure is insufficient.
Feedback needs an appropriate timescale. If a repair’s effectiveness becomes clear only after several weeks, retain a way to connect the later result to the original assessment. A high volume of decisions without usable outcome feedback may provide less learning than the activity count suggests.
Retain Context Without Treating Every Memory as Policy
Case reviews can identify knowledge worth recording: the equipment version, operating conditions, a misleading symptom, or the reason a standard test was unsuitable. Record enough to help another technician recognise when the case is relevant. A bare conclusion without those boundaries can become another source of inappropriate advice.
Separate confirmed constraints from recollection and unresolved suspicion. An informal view that a supplier is unreliable may warrant investigation, but it should not automatically become an exclusion rule. Establish who can validate the information, decide its operational meaning, and review whether it remains current.
Documentation and apprenticeship serve complementary purposes. Records make evidence and decisions available. Working alongside others helps people use that evidence, ask questions, and recognise where it is incomplete. Improving one does not eliminate the need to design the other.
Knowledge capture can begin with a specific incident rather than a request for an expert to list everything they know. Ask what they noticed first, which finding made the case unusual, what alternative they considered, and what would have changed their next step. Where possible, compare the account with the records and observations available at the time. This can identify information worth preserving without assuming that a retrospective explanation is a complete account of the judgement.
The resulting material should be tried by another practitioner. Give them a comparable case and see whether the record helps them ask the relevant questions. If it does not, the gap may involve missing context, an unclear concept, or a practical skill that needs demonstration. This makes knowledge transfer an activity with feedback, rather than a documentation count. It also gives the team a way to improve material used by an assistant without claiming that the model or the document now contains every aspect of the expert’s capability.
Lisanne Bainbridge’s Ironies of Automation discusses problems created when operators are left to handle abnormal conditions while automation changes their ordinary work. That is a reason to examine practice and takeover requirements. It is not evidence that AI necessarily causes an irreversible loss of every skill it assists.
Check Capability Before It Is Needed for Recovery
The team should periodically examine whether staff can perform the tasks they are still expected to own. Can they diagnose a representative case with the authorised source material? Can they recognise when the assistant lacks relevant evidence? Can they manage the agreed fallback when the service is unavailable?
Use those checks to improve training and workflow design. If a particular task has become unfamiliar, identify whether practice, better information, or a different allocation of responsibility would address the gap. Avoid treating a single exercise as proof of either permanent skill loss or complete readiness.
The related issue of worker autonomy concerns whether people have the discretion and authority to use their judgement. Expertise development concerns how that judgement is learned and maintained. A technician can retain authority while lacking practice, or remain highly skilled while being unable to act. Both need attention.
Operational control also depends on what the organisation can support at scale. A few expert technicians may provide valuable backup without having capacity to absorb every automated case. Recovery planning should reflect both competence and workload.
AI adoption can make useful knowledge more accessible and alter the conditions that produce new expertise. The maintenance team can assess both by looking at representative practice, feedback, source context, and demonstrated capability. Preserving that learning process gives the organisation a firmer basis for adaptation than assuming either that the model has captured everything or that human expertise must inevitably disappear.
