AI Work Instructions: Capture Know-How Before It Retires

AI Work Instructions: Turning Tacit Knowledge Into Procedure

AI work instructions have become one of the most practical uses of generative AI on the factory floor, for a reason that has little to do with technology. Experienced operators, maintenance technicians, and process engineers are retiring, and much of what they know has never been written down. In October, AGC described a voice tool that pairs speech-to-text with generative AI to turn shop-floor explanations into reusable records, claiming large reductions in the time spent analysing work. Aviation maintenance organisations are turning to AI for the same reason, as experienced staff leave faster than they can be replaced. Capturing know-how quickly is valuable. But a work instruction is not a summary. Someone will follow it with their hands, and if a step is wrong, the consequence lands on a part, a machine, or a person.

Ways to Capture Expert Knowledge

Each method trades speed against how much of an expert's real judgement makes it into the record.

Why Knowledge Capture Became Urgent

Manufacturing has warned about a skills gap for years, with industry studies projecting large numbers of unfilled roles over the coming decade. The more immediate problem is concentration. In many plants a handful of veterans know how to recover a particular machine from a fault, how to adjust a process when raw material varies, or which supplier's parts need an extra check. When those people retire, the knowledge leaves with them, and the first sign is often a longer outage or a quality problem nobody can diagnose.

Traditional knowledge capture asked experts to write procedures or sit with technical writers. Both are slow, and experts often struggle to articulate what they do instinctively. Recording them talking through a task, then letting AI structure the result, lowers that barrier dramatically.

How AI Work Instructions Are Made

A typical workflow records an expert explaining and performing a task, transcribes the audio, and uses a language model to draft structured steps, warnings, tools, and quality checks. Some tools also pull from existing manuals, maintenance logs, and engineering specifications to fill gaps, and generate translated versions for multilingual workforces. The result can be a draft work instruction in hours rather than weeks.

Where AI Work Instructions Go Wrong

In our view the errors that matter are predictable:

• Dropped safety steps. An expert mentions isolating energy in passing, and the summary omits it.
• Wrong values. Torque settings, temperatures, or tolerances get misheard, rounded, or mixed between variants.
• Merged procedures. Steps for two similar machines or product variants end up in one instruction.
• Invented detail. The model fills a gap with a plausible step the expert never described.
• Lost conditions. "Only if the part is above room temperature" disappears, turning a conditional step into a universal one.
• Translation drift. Technical terms translated literally can change the meaning of a warning.

Tacit Knowledge Is Not Always Right

One uncomfortable truth surfaces in almost every knowledge capture project: experts sometimes do things differently from the official procedure, and sometimes they are right and the procedure is wrong. Sometimes it is the reverse. A veteran's workaround may compensate for a design flaw nobody fixed, or it may be a habit that bypasses a control for convenience. AI will faithfully capture both.

That makes reconciliation essential. Every captured instruction should be compared with engineering specifications and existing approved procedures. Where they differ, the difference is valuable information. It might reveal a needed engineering change, a training gap, or an unsafe habit, but it should be resolved by engineering and quality, not by whichever version was written most recently.

A Validation Workflow

1. Capture the why as well as the how. Ask experts to explain reasons, warning signs, and what goes wrong.
2. Draft with AI, then compare against specifications. Check every value, tolerance, and sequence against controlled documents.
3. Review with the expert. The person who was recorded should confirm the draft says what they meant.
4. Have engineering and safety sign off. Especially for energy isolation, lifting, chemicals, and quality-critical steps.
5. Trial on the floor with a less experienced worker. If they cannot follow it successfully, it is not ready.
6. Release through document control. Version, approve, and distribute like any other controlled instruction.
7. Collect feedback and revise. Treat the first release as a draft that will improve with use.

Pros and Cons of AI-Drafted Instructions

• Pro: speed. Knowledge gets captured before experts leave, not after.
• Pro: lower barrier for experts. Talking through a task is far easier than writing a procedure.
• Pro: consistency and translation. Instructions follow a standard format and reach multilingual teams.
• Con: confident errors. Polished drafts can hide missing steps and wrong values.
• Con: captured bad habits. Workarounds get documented as if they were approved methods.
• Con: validation effort. The time saved drafting must partly be reinvested in checking.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how AI work instructions play out in practice rather than presented as verified case studies.

A plant records its most experienced maintenance technician explaining how to recover a press after a jam. The AI draft is clear and well structured, but it omits the lockout step the technician mentioned only briefly at the start. The floor trial catches it, and the plant adds a rule that every draft is checked against the site's energy isolation standard.

A packaging line operator explains a workaround for a feeder that misaligns with one material. The captured instruction reaches engineering, which realises the workaround compensates for a worn guide. The guide is replaced and the workaround is retired, a better outcome than either documenting it or ignoring it.

A multinational translates AI-drafted instructions into four languages. A native speaker review finds that a warning about residual pressure has been softened to a suggestion in one translation. The review step becomes mandatory for any safety-related content.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

Document Control Still Applies

Quality management systems such as ISO 9001, and sector standards in aerospace and medical devices, require work instructions to be controlled: approved, versioned, and available where they are used. AI drafting does not change that. An AI-generated instruction that bypasses document control is an uncontrolled document, and in regulated sectors that is an audit finding waiting to happen.

The cost of a wrong value travels further than it seems, a pattern we described in AI in manufacturing. And where instructions relate to machinery with safety functions, the wider obligations covered in the EU Machinery Regulation also come into view.

Why Talkory Wins

Reviewing a drafted work instruction is partly a matter of asking the right questions: what could be missing, which step is ambiguous, where a warning should go. Talkory runs the same question across GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 in one pass. An engineer can paste a draft instruction and ask each model to identify missing safety steps, ambiguous wording, or likely errors, then see where six independent reviews agree. Consistent flags deserve attention. Disagreement often points to the steps where context matters most, which are exactly the ones to check with the expert and on the floor.

Final Verdict

AI work instructions are one of the most useful tools manufacturers have for capturing know-how before it walks out of the door. They make it easy for experts to share what they know and fast to turn that into structured procedures. The risk is that speed hides errors: dropped safety steps, wrong values, merged variants, invented detail, and bad habits documented as good practice. Capture the reasons as well as the steps, reconcile with specifications, review with the expert, sign off through engineering and safety, trial on the floor, and release through document control. Capture fast. Validate properly.