The operating risk
The warning is not simply that experienced people retire.
At the Made in Connecticut manufacturing-report panel, participants discussed retirement planning, cross-training, developing people internally, and the risk that working knowledge can leave with an experienced employee. I am using that as a session-level theme, not attributing automatic-transcript wording to an individual panelist.
The official report adds measurable context: Connecticut manufacturing has become more technically weighted, manufacturers report that open positions often take weeks to fill, and the state’s labor force is under pressure. None of that proves a specific manufacturer has a knowledge-continuity problem. It does explain why a fragile handoff can become an operating constraint rather than merely a documentation inconvenience.
Choose a boundary
Preserve one consequential decision or handoff first.
“Capture our tribal knowledge” is too broad to operate. A better starting point is one interruption, judgment, or handoff: how an estimator finds the governing requirement before quoting; how a planner recognizes an exception; how an inspector locates the approved instruction; or how a senior employee explains why a recurring condition requires escalation.
- Name the user and the decision, task, or handoff that needs continuity.
- Identify the controlled source that already governs the work and who owns it.
- Separate documented instruction from expert context, history, heuristics, and unresolved opinion.
- Ask another competent person to review captured context before it is made broadly available.
- Define who corrects, expires, archives, and escalates the material after the original expert moves on.
A practical framework
Use a controlled knowledge-continuity lifecycle.
| Stage | Required record | Control question |
|---|---|---|
| 1. Select | One named decision, handoff, user group, and operating consequence. | Is the boundary narrow enough to maintain and evaluate? |
| 2. Establish authority | Controlled sources, document owners, competent reviewers, and escalation path. | Which source governs when expert context and approved instruction differ? |
| 3. Elicit context | Examples, cues, exceptions, rationale, and uncertainty captured with provenance. | What is observation, inference, local practice, or unresolved judgment? |
| 4. Review | A second competent person checks meaning, scope, and dangerous omissions. | Could a new user apply this outside its intended conditions? |
| 5. Publish deliberately | Status, audience, source links, scope, owner, effective date, and review/expiration date. | Can the user tell guidance from controlled instruction? |
| 6. Retrieve with evidence | Search or AI responses show sources, status, and uncertainty. | Can the user inspect the underlying authority rather than trust a fluent summary? |
| 7. Correct and maintain | Feedback route, correction log, source-change trigger, and archive decision. | Who owns the answer after the original expert is unavailable? |
| 8. Measure | Onboarding time, expert interruptions, search success, corrections, exceptions, and source use. | Did continuity improve without hiding errors or authority? |
Assistance boundary
AI can improve access; it cannot create authority.
AI may help transcribe an interview, organize candidate themes, compare a draft against a template, retrieve relevant passages, present cited context, or suggest questions that remain unanswered. Those are assistance functions. They become useful only inside a maintained information process.
AI must not silently turn an interview into an approved work instruction, choose among conflicting sources, invent a missing exception, grant competence, replace required training or supervision, or approve a manufacturing decision. A generated answer should preserve links to the governing material and make uncertainty visible enough for the responsible person to act.
| AI may assist with | A competent person remains responsible for |
|---|---|
| Finding potentially relevant source passages. | Confirming applicability, revision, scope, and controlling authority. |
| Drafting a summary of reviewed expert context. | Approving meaning, limits, release status, and intended audience. |
| Suggesting related examples or missing fields. | Determining whether the example is correct, safe, complete, and relevant. |
| Collecting feedback and grouping correction themes. | Resolving the correction and changing the authoritative record. |
Measure the handoff
Look for continuity, not content volume.
A large knowledge base is not proof that knowledge transfer worked. For the selected handoff, compare practical measures: time for a new person to complete the task with appropriate review, interruptions to the expert, search attempts that reach the right source, corrections per use, repeated exception types, and the share of answers that expose an approved source.
Also monitor the failure side. If users bypass the source, apply context outside its scope, stop reporting errors, or treat generated text as instruction, the system may be creating confidence rather than continuity. The right response may be to revise the content, narrow the audience, improve training, change the tool, or stop the AI layer entirely.
The goal is modest but valuable: make one important handoff less dependent on being able to interrupt one particular person, while preserving the source, context, correction path, and human authority that make the knowledge trustworthy.
References
Sources and further reading
- 2026 Connecticut Manufacturing Report Report pages 8–10 · Workforce composition, hiring duration, and workforce-development context
- 2026 Connecticut Manufacturing Report: Results & Insights panel Panel discussion, 00:04:43–00:08:01 (session-level paraphrase) · Discussion themes are paraphrased without attributing them to an individual speaker
- AI for Manufacturing Knowledge Transfer Maintained-source and expert-context implementation guide