Key takeaways
- Capture the conditions around expert advice, not just the final instruction.
- Approved documents remain authoritative; interviews and notes provide context.
- Every answer should show its source, revision and escalation path.
- Measure time to trusted information, repeat questions and successful handover, not chatbot usage alone.
Why manufacturing knowledge disappears
Critical plant knowledge is often distributed across SOPs, maintenance history, drawings, shift notes, vendor manuals and the memories of experienced people. The problem is not simply that information is undocumented. It is that a skilled worker knows which document applies, which symptom matters, what changed after the last shutdown and when a familiar workaround is no longer safe.
An AI knowledge assistant can make this context easier to find, but only when the content has ownership and boundaries. A model that gives a fluent answer without showing the source can make poor information more persuasive.
Choose one knowledge bottleneck
Start with a recurring question that delays a real workflow. Good examples include finding the correct inspection procedure, understanding an alarm history, preparing a maintenance job, comparing the current symptom with previous work orders or assembling evidence for shift handover.
Define the user, the moment of need and the action that follows. A maintenance planner preparing tomorrow's work needs different evidence from an operator responding to an abnormal condition. Keep emergency and safety-critical instructions outside the first pilot unless the relevant engineering and safety owners design the controls.
Build a trustworthy knowledge set
Separate content into three classes:
- Authoritative records: - approved SOPs, controlled work instructions, manuals, drawings and current policies.
- Operational evidence: - work orders, inspection records, alarms, deviations and shift logs.
- Expert context: - interviews, annotated examples, decision cues and lessons learned.
Record document owner, revision, effective date, equipment or process scope and access classification. Expert interviews should capture the question being solved, the conditions under which the advice applies, warning signs, exceptions and the person or role that must approve action.
Design answers for the factory floor
A useful response should begin with the direct answer, cite the exact source, show the applicable equipment and revision, identify uncertainty and offer the next safe action. If evidence conflicts, the assistant should show the conflict rather than silently choosing one version.
For multilingual teams, translation can improve access, but the approved source language should remain available. Technical terms, units, tags and warnings need controlled terminology. Voice access may help gloved or mobile workers, provided noisy environments and confirmation of critical details are tested.
Keep people in authority
Industry 5.0 places worker wellbeing and agency at the centre of technology design. In practice, that means the assistant helps people retrieve and compare evidence while qualified roles retain decisions about safety, quality, release, maintenance and process changes.
Create a visible escalation path. A low-confidence answer, missing revision, unresolved contradiction or request outside the approved scope should go to a named role. Feedback should correct the source or retrieval logic, not simply teach the model to repeat a preferred answer.
A 60-day pilot pattern
- Select one role, one site area and the ten questions that consume the most expert time.
- Inventory authoritative documents and remove obsolete or duplicate versions.
- Interview two or three experienced workers using real recent examples.
- Build a read-only assistant with citations, access control and an escalation path.
- Test with normal questions, ambiguous wording, outdated documents and deliberately missing evidence.
- Compare time to trusted information, answer acceptance, escalations and corrections against the baseline.
What success looks like
Success is not the number of conversations. Look for faster access to approved information, fewer repeated interruptions to experts, more complete shift handovers, better preparation for supervised work and a growing list of resolved knowledge gaps. Also monitor unsafe confidence, stale sources, bypassed escalation and the workload placed on content owners.
The long-term asset is not the chatbot. It is a governed knowledge system that experienced workers can improve and newer workers can challenge, understand and use responsibly.
