Record short, structured interviews at the machine, transcribe the voice notes, and have an AI draft work instructions from the transcript. The veteran then reads every draft, corrects it and signs it off, and a trainee tests it on a real job. Start at least six months before he leaves.
The value is in the judgement, not the manual. Any machine has a manual; what disappears at retirement is "if the chips go blue, drop the feed by a tenth", "that fixture sits a thou off, tap it here", and which supplier's blanks behave. AI is good at turning spoken ramble into a tidy document, and poor at knowing which parts of the ramble matter, so your job is to design the questions.
Work out what is worth capturing before you press record
An open request such as "tell me everything you know" produces three hours of stories and no instructions. Narrow it first. Make a list of the jobs, machines and problems that depend on this one person, then rank them by what it would cost if he were unavailable tomorrow.
Sources for that list already exist in the building:
- Scrap and rework records: which parts, which operation, which fixes.
- Customer complaints and corrective-action reports, where the same name appears as the person who found the cause.
- The jobs colleagues send to him because he "has the touch".
- Maintenance notes on machines that only he coaxes through a bad week.
Paste anonymised extracts of those records into an assistant and ask it to cluster them: "Group these 40 scrap notes by operation and likely cause. Show the groups with the most repeats first." Treat the output as a prompt for conversation, not as a finding; the first meeting with the veteran should be spent correcting the list, and he will enjoy doing so.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA twelve-week plan for one machinist, with real numbers
This is an illustration of the effort involved, not a benchmark. A small job shop has a machinist with decades on a five-axis mill and two older lathes. He retires in about six months. The owner picks 12 topics from the ranking above.
| Week | Activity | His time | Output |
|---|---|---|---|
| 1 | Agree scope, consent, list of 12 topics | 45 min | One-page agreement and ranked list |
| 2-9 | Two 30-minute recorded sessions a week at the machine, one topic per session | 8 h total | 16 recordings (8 hours of audio) |
| 3-10 | Transcribe, draft instruction, send for review | 20 min per draft | 16 drafts, about 5 h of review |
| 10-11 | Trainee runs three of the instructions on live jobs | 1 h supervising | Marked-up corrections |
| 12 | Final read and signature | 2 h | Signed set in the shop's document system |
His total time comes to roughly 17 hours across 12 weeks, about an hour and a half a week. Transcription of 8 hours of audio is 480 minutes. At the estimated $0.0045 a minute OpenAI lists for its gpt-transcribe model (as of November 2026), that is about $2.16; a note-taker plan such as Otter Pro, listed at $8.33 a month on annual billing or $16.99 monthly with 1,200 transcription minutes a month, would also cover it. The AI cost is trivial. The real cost is his hours and the trainee's, so protect the calendar before you buy anything.
Interview prompts that draw out the judgement calls
The interviewer can be the production manager, the trainee, or the AI itself in text mode before a session. What matters is the shape of the questions. Questions that start with "how do you" give procedures; questions that start with "what tells you" give the signals a newcomer cannot see.
Ask the AI to prepare each session's questions from the topic and the records:
I am interviewing a retiring machinist about [topic: e.g. holding thin-wall
aluminium housings on the 5-axis mill]. Here are the scrap notes and
the current setup sheet for that job: [paste].
Write 12 interview questions in this order:
1. The normal sequence, in his own words.
2. Five questions that start "What tells you..." (sound, colour, feel, chip shape, gauge reading).
3. Four questions about what goes wrong and how he knows early.
4. Two questions about what he would do differently with a new starter.
5. One question to check for anything he does without thinking.
Keep each question under 20 words. No jargon he would not use himself.A realistic output, trimmed: "1. Walk me through the first ten minutes of this job. 2. What tells you the vacuum fixture has seated properly? 3. What do you listen for on the first roughing pass? 4. What tells you the cutter is about to go? 5. How do you check the part has not moved before finishing? 6. What is the first sign the housing is warping? ..." What you would change: the questions are good but generic, so add one that names a specific past failure from the records ("In March a batch came out 0.08 mm out of flat; what happened?"). Specific stories draw out better answers than general questions.
Recording on the floor without wrecking the audio
Noise is the main enemy. A phone placed on the machine will record the spindle and little else. Do the following:
- Record the interview beside the machine at a quiet moment (between jobs, during a tool change) if he needs to point at things, and a second copy of the key explanations away from the noise.
- Clip a cheap lapel microphone to his collar, or use the phone held about 20 cm from his mouth. Test with 30 seconds before every session.
- Photograph or film the thing he points at. "Put it in this slot and tap it here" means nothing in a transcript, but a photo timestamped to the recording solves it.
- Have him say part numbers, tool numbers and settings slowly and repeat them. It costs a few seconds and saves most transcription errors.
- End every session with "What have I not asked that I should have?" The best material often comes in the last two minutes.
Before a transcription tool touches the audio, check three things: whether it keeps the recording, how long, and whether it uses the audio to train models. These recordings describe how you make your products, so a plan with training turned off, or on-device transcription, is the safer choice. If you are unsure, the knowledge management tutorial covers where to keep the results.
Turning a transcript into a draft work instruction
Transcripts are messy: false starts, "you know", the machine alarm, and the same point made three times. A good prompt tells the AI to keep his words where they carry the judgement and to flag, not fix, anything it is unsure of.
Below is a transcript of a retiring machinist explaining [job].
Turn it into a draft work instruction with these sections:
- Purpose and when this applies
- Before you start (material, fixture, tools, gauges)
- Steps in order, numbered, one action each
- What to watch and listen for, with his own wording in quotation marks
- Typical faults and the fix he described
- Open questions: anything unclear, contradictory or missing
Rules: use only what is in the transcript. Never invent a number. If he
gave a figure, copy it exactly. If a figure sounded uncertain, put it
under Open questions with the exact words. Mark anything you inferred
with [CHECK].
[paste transcript]The last rule is the one that matters. With "use only what is in the transcript" and the [CHECK] marker, the output of an illustrative run reads like this: "Step 4. Rough the pocket at 60% of the finishing feed. [CHECK: transcript says 'about a third less, don't hold me to it'; 60% is my reading.] Listen for a steady hum; if the pitch rises, 'back it off before it chatters'." That is exactly where the veteran looks first, which is why you want those marks.
What the AI changes without telling you
Three habits show up in almost every first draft, and you can test for each of them.
- Tidying a rule of thumb into a rule. "Usually about a thou" becomes "0.001 in". Search the draft for exact figures and compare each with the transcript.
- Merging two jobs. When he moves from the mill to a lathe in the same story, the model may blend the steps. Ask it to keep one machine per document.
- Adding safety or quality statements nobody said. Some are helpful, but they are not his, and a guard rule invented by software can contradict your real procedure. Delete any line that is not in the transcript, then add real requirements from your safety lead.
Make a habit of asking the model for a second pass: "List every number and tool name in the draft with the minute of the recording where he said it." It will not always get the timestamps right, but gaps tell you where to look.
The sign-off meeting: how the veteran signs and what he checks
Do not send a PDF and ask for approval by email; that gets a thumbs-up and nothing else. Sit with him for 20 minutes per draft, with a printout, and ask him to mark three things in different colours: wrong (strike through), missing (insert) and "that's not how I'd say it" (rewrite in his words). A trainee who later reads "chatter" in his voice remembers it.
When the marked-up draft is final, add a short block at the bottom:
Reviewed and agreed as accurate by: [name], [role]
Date: [date] Revision: 1
Applies to: [machine, part family]
Not covered: [anything he said is outside this document]That "not covered" line matters. It stops a trainee applying the instruction to a different alloy or a different fixture and blaming the document. Then store the signed version where the team will find it; the ISO 9001 documentation tutorial shows how such instructions fit a controlled document system.
A trainee test that proves the instruction works
An instruction is finished when someone who was not in the room can follow it. Pick a trainee with some experience but not on this machine, give them the draft and a real job, and watch without helping. Count every moment they stop and ask, because each stop is a missing step.
In an illustration, a trainee asked six questions on the first run of a fixture-setup instruction: where the locating pin goes, which of two near-identical clamps, how tight "firmly" is, which coolant, what the first-off check is, and where to find the offset. Each was added to the draft, three with photographs. On the second run with a different trainee, there was one question. That drop is your measurement. If you can drive questions to zero or one, the instruction is ready.
Combine it with photographs and short clips if you can; the tutorial on work instructions from photos and video goes through that step.
Capturing what he knows about suppliers, customers and the odd machine
Not all tribal knowledge is a machine procedure. Some is relationship knowledge: which supplier's bar stock runs soft in the third week, which customer's drawings always have the tolerance on the wrong view, who to ring at the carrier. These are quick to capture and often worth more than the setup sheets.
Use a different format: a 15-minute voice note, one subject, then a summary table.
| Subject | What he said (illustrative) | Action for the team |
|---|---|---|
| Supplier of 6082 bar | "Their Friday batch is harder; check hardness before cutting" | Add incoming hardness check for Friday deliveries |
| Customer drawings | "Always ring them about the flatness note on sheet 2" | Add a query step to the quote checklist |
| Older lathe | "It wanders after two hours, give it a warm-up cycle" | Write a 10-minute warm-up routine |
Ask the AI to turn each row into a checklist line and show him the result; the corrections tend to be short and honest. Link these items to your gauge and calibration records where they overlap.
Ranking the topics: a scoring rubric you can fill in on one page
Twelve topics in twelve weeks is the capacity of one person's attention, so choose them on evidence. Give each candidate topic a score from 1 to 3 on four questions and add them up. The topics with 10 or more go first.
- Cost if nobody knows (1-3). A 3 is a job that stops a line or ships a bad batch to a customer. A 1 is an inconvenience.
- Written down already (3 is "nowhere", 1 is "fully documented"). If a decent instruction exists, you only need his corrections, not a new document.
- How often it comes up (1-3). A weekly job beats an annual one for training value.
- How hard it is to learn by watching (1-3). A 3 means the knowledge sits in sound, feel or timing, which cannot be seen.
An illustration: "Hold thin-wall housings without distortion" scores 3, 3, 2 and 3, total 11, and goes first. "Change the coolant" scores 1, 1, 2 and 1, total 5, and waits. Ask the AI to score your list if you describe each topic in a line, then overrule it wherever you know better. A model does not know that the coolant change caused a month of scrap two years ago, but you do.
Three veterans, three different capture methods
The method above suits someone who is happy to talk at the machine. Not everyone is, and a plan that forces one style loses the quiet people.
The talker. He explains well, loses the thread, and needs a listener to bring him back. A recorded walk-through with a trainee asking "why?" works best. The risk is volume: a 90-minute story contains 10 minutes of instruction, so the AI's job is to cut hard and flag what it dropped.
The doer who hates being recorded. Film his hands instead and talk afterwards. A 5-minute clip of a setup, shot from over his shoulder, followed by a short voice note recorded while he watches the clip, gets the knowledge without the mic in his face. Ask the AI to match each sentence of the voice note to the part of the clip it describes.
The writer. Some veterans keep a notebook or a folder of scribbled setup sheets. Photograph the pages and use the AI to transcribe the handwriting, flagging every word it cannot read. Then interview him about the gaps. Handwritten abbreviations such as "T3 -.02 after 50" mean everything to him and nothing to the model, so expect to explain each one once and build a glossary that goes with the instructions.
A first draft that went wrong, and what it taught the shop
In one illustrative case, a shop's first AI-written instruction for a lathe job read well, and the veteran skimmed and signed it. A week later a trainee scrapped two parts. The instruction had left out the ten-minute warm-up cycle, because he had mentioned it in the second minute of the recording as a throwaway remark, and the model had treated it as chat. Nobody had checked the draft against the transcript for what was missing, only for what was wrong.
The fix had two parts. First, the prompt gained a line: "List every instruction-like remark in the transcript that you did not use in the draft, with the minute it was said." Second, the sign-off meeting added a question: "What would a newcomer do wrong if they followed this exactly?" That question draws out omitted steps better than "is this right?", because it asks the expert to imagine failure rather than read for typos.
Keeping the instructions alive after he leaves
The signed set is a snapshot. Machines get new tooling, materials change, and the next operator will find a better way. Name an owner for each instruction, usually the person who now runs that machine, and put a review date in the header. Ask people to add a note whenever they deviate from the text and why; the monthly job is a short read of those notes, with the AI summarising which instructions attracted the most edits.
The simplest safeguard is a standing offer: the veteran, if he is willing, is available for an hour a month for six months after he retires, paid at an agreed rate, to answer questions the instructions raise. Whether that is possible depends on his wishes and your employment terms, so agree it early and in writing, and take advice if there is any doubt. If the sign-off meetings were done well, expect few questions, and fewer each month.
Protecting the recordings and the person
The audio and transcripts hold his voice and your methods. Agree, in writing, who can hear them, where they live and when they are deleted (a sensible default is to keep the signed instruction and delete the audio after the trainee test). Do not paste a customer's confidential drawing into a consumer chat tool to help a transcript make sense; describe it instead. Use a business plan, or the tool's setting that turns off training on your content, and check that setting yourself rather than relying on memory.
Treat his name as credit, not a risk. "Setup method as described by [name]" in the header tells the next person whose judgement this is, and tends to make the veteran more generous with detail. Any question on employment terms, ownership of the recordings or use of his voice beyond training belongs to an employment adviser, not an AI.
Checking the programme worked a year later
Three measures tell you whether the effort paid off: scrap and rework on his machines before and after he left (compared on the same part families), the number of times someone asks for "the number to ring him", and how many instructions have been edited since sign-off. Edits are a good sign when they come from users adding to the text; a document nobody changes has probably not been read. If you want to plan the capture around your own people and machines, that is the kind of job my AI implementation consultation covers.
Questions before you start recording
Do I need the veteran's permission to record?
Yes, and get it in writing. Explain what is recorded, who can hear it, where files are kept and when they are deleted. Recording someone without agreement damages trust and may break rules where you operate, so ask an employment adviser if you are unsure of your obligations.
How long does the whole capture take?
Plan on 8 to 12 weeks for one person's core knowledge: two half-hour sessions a week at the machine, plus about 20 minutes of review per draft. Starting six months before the retirement date leaves room for a trainee to test every instruction.
Is a voice recorder enough, or do I need a transcription tool?
A phone's voice recorder is enough to capture. You need a transcription step to turn speech into text before the AI can structure it. Check how the tool stores audio and whether it trains on it, because the recordings contain your processes.
What if the veteran does not want to share everything?
Do not force it. Start with the jobs that hurt most when he is away, make clear the aim is to keep his judgement valued, and let him choose the order. People share more once they see their own words come back accurately and with credit.
Will AI get the technical details wrong?
It can mishear numbers, tool names and part numbers, and it may tidy a rule of thumb into something firmer than he said. That is why every draft goes back to the veteran for a line-by-line sign-off before anyone uses it on the floor.
Further reads
- AI Quality Control for Small Manufacturers: Where to Start — Use the captured defect knowledge in your checks.
- 8D Report With AI: Answer a Customer Complaint Step by Step — Turn past complaints into cause-and-fix records.
- RMA Process for a Small Manufacturer: Returns, Repairs, Credits — Record how returns are diagnosed and repaired.
- Where Should a Small Manufacturer Start With AI? — Pick the first job worth giving to AI.
- Capacity Planning Template: Can You Take That Order? Check With AI — Check capacity while a key person is away.
- 5S Audit Checklist: Build One for Your Shop Floor With AI — An area-by-area 5S scoring sheet drafted with AI from photos and notes, scored monthly by a supervisor, with AI turning the scores into a one-page trend.
- AI for Industrial Manufacturing: Read Downtime Logs to Find Repeat Stops — A practical CSV walkthrough for finding repeated line stops, checking the arithmetic and choosing a maintenance action.
- AI for Small Manufacturing Companies: Build a Weekly Production Schedule — A spreadsheet-based production schedule with worked capacity maths, changeover checks and a daily review routine.
- AI in Plant Engineering: Breakdown Logs and Maintenance Plans — Turn a messy breakdown log into ranked repeat faults and a draft preventive maintenance plan, with the safety items left to a named engineer.
- Machine Shop Quote Template: AI Turns RFQ Emails Into Draft Quotes — Extract part, quantity, material and finish from RFQ emails, price from your rate card, and let an estimator approve every number.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: figures in the worked example are illustrative; transcription and note-taker prices were taken from OpenAI's API pricing page and Otter's pricing page in November 2026.