Record the person who does the task talking it through, or collect their rough bullet notes, then give an AI assistant those notes plus a fixed SOP template and tell it to structure what's there and mark gaps as questions instead of filling them. Before the owner signs it off, have someone unfamiliar with the task follow the draft.
The biggest risk is that AI writes a confident, plausible step nobody actually does, such as a torque check or a supervisor sign-off that isn't part of your process. Invented steps look professional, so they survive a quick review unless you forbid them in the prompt and check each step against the notes. A good first draft has questions in it; a suspiciously complete one probably has fiction in it.
Pick procedures where a mistake costs real money
Don't start with the easiest task. Start with the one where a missed step causes scrap, a complaint or a safety near-miss, and where more than one person does it. A quick score of frequency, cost of a mistake and number of people involved (1 to 3 each) sorts the list in ten minutes.
For an illustrative manufacturer of spare parts for packaging machines, the owner's list looked like this:
| Procedure | How often | Cost of a mistake | People | Score |
|---|---|---|---|---|
| CNC lathe changeover and first-off inspection | 3 | 3 | 3 | 9 |
| Goods-in check for bar stock | 3 | 2 | 2 | 7 |
| Packing parts for export orders | 2 | 3 | 2 | 7 |
| Weekly coolant concentration check | 2 | 2 | 1 | 5 |
| Month-end stock count | 1 | 2 | 3 | 6 |
The changeover wins: it happens several times a shift, a wrong offset scraps a batch, and three setters each do it slightly differently. That last point is common and matters. Writing an SOP often exposes that there are three procedures, not one, and somebody has to choose the version that becomes standard before AI writes anything.
Capture the task as it's really done
You need raw material, not prose. Three capture methods work, and the first is usually best:
- Talk-through while doing it (15 to 20 minutes). The setter records a voice memo on a phone during a real changeover, saying what they're doing and why. Transcribe it with the phone's built-in transcription or the voice input in your AI assistant's app.
- Rough bullets afterwards (10 minutes). Quicker, but people skip the steps they do on autopilot, which are exactly the ones a new starter misses.
- Photos of each stage (5 minutes extra). Screens, gauges, labels and settings. You can't paste photos into the final SOP here, but they help the reviewer check the draft and can go into the version on the shop floor.
This is what a raw talk-through transcript tends to look like, and it's fine that it's messy:
"Right so first I pull the last job's program, well actually first check the drawing revision matches the traveller, that's caught us before. Then chuck out, new jaws if it's the 40 mil stuff, soft jaws for the brass. Load program, check tool list against what's in the turret, usually T3 needs swapping. Touch off, set Z zero on the face. Run first one at 50% rapid, measure the diameters with the mic, the bore with the gauge, if it's out more than two thou... sorry, 0.05, adjust wear offset, run another. Then get Sam or whoever's on to sign the first-off sheet."
Notice what's valuable: the drawing-revision check ("that's caught us before"), the tolerance, the specific tool that usually needs swapping. Notice what's risky: "two thou... sorry, 0.05" mixes units, and "Sam or whoever's on" is a role that needs a proper name.
The SOP skeleton to give the AI every time
Using the same template for every procedure means staff learn where to look, and the AI can't wander into its own format.
SOP TEMPLATE
Title:
SOP number / version / date:
Owner (role):
Purpose (one sentence):
Scope (when this applies / when it doesn't):
Safety and quality warnings (only ones specific to this task):
Tools, materials and documents needed:
Steps (numbered; one action per step; decision points as IF / THEN):
Checks and acceptance criteria (with units):
Records to complete:
What to do if something goes wrong (who to tell):
Open questions for the owner:
Change log:
A drafting prompt that forbids invented steps
Turn the transcript below into an SOP using the template exactly.
Rules:
1. Use only information in the transcript. Do not add steps, checks,
warnings, tools or sign-offs that are not mentioned.
2. Where something seems missing or ambiguous, do not fill it in.
Add it to "Open questions for the owner".
3. One action per numbered step, starting with a verb.
4. Write decision points as IF ... THEN ... OTHERWISE.
5. Keep all numbers and units exactly as stated. If units conflict,
flag it as an open question.
6. Plain English for a new starter. No filler, no generic safety text.
TRANSCRIPT: [paste]
An illustrative extract of what comes back:
3. Check that the drawing revision matches the job traveller. IF they don't match, THEN stop and tell the production supervisor.
7. Compare the tool list in the program with the tools in the turret. Replace T3 if it doesn't match.
10. Run the first part at 50% rapid.
11. Measure outside diameters with a micrometer and the bore with a bore gauge. IF any dimension is out by more than 0.05 mm, THEN adjust the wear offset and run another part. OTHERWISE go to step 12.
Open questions: (a) Transcript mentions "two thou" and "0.05"; confirm the tolerance is 0.05 mm. (b) Which jaws are used for materials other than 40 mm steel and brass? (c) Who is authorised to sign the first-off sheet?
That's a good draft. What you'd still fix: step 7 says "replace T3 if it doesn't match", but the setter meant "check every tool; T3 is the one that usually differs". The model turned a helpful aside into a rule. It's the most common kind of drafting error, and you'll only catch it by reading each step against the transcript.
Make the AI list what's still missing
Once the draft exists, a second prompt finds the gaps a new person would hit. Run it in the same chat:
Read the SOP as if you were a competent machinist who has never worked here.
List every point where you would have to ask someone a question before
continuing. For each, quote the step and say what you'd need to know.
Do not answer the questions yourself.
Illustrative questions it might return: "Step 5: where are the soft jaws kept?", "Step 9: which face is Z zero set on for parts with a shoulder?", "Step 12: where is the first-off sheet, and is it paper or on the tablet?" The owner answers these in ten minutes, and each answer becomes a line in the SOP. In a spare-parts shop, those three answers are the difference between an SOP a new setter can follow on nights and one that needs the day-shift lead standing next to them.
The new-starter test
An SOP isn't finished until someone who doesn't normally do the task has followed it. Ideally that's a recent starter or someone from another area; the author is the worst possible tester. Ask them to follow the steps exactly, and to say out loud wherever they hesitate, while the owner watches without helping.
In an illustrative test of the changeover SOP, the tester stalled twice. At step 9, "Set Z zero on the face" didn't say whether to face the part first; the experienced setters always take a skim cut, so nobody had mentioned it. At step 13, the first-off sheet asked for a signature from "QC", a role that doesn't exist on the night shift. Both fixes took minutes. Without the test, both would have surfaced as a scrapped batch at 3am.
A second example: sample receipt at a testing laboratory
Procedures with decisions in them benefit most from AI's IF / THEN formatting. An illustrative lab's rough notes for receiving client samples read: "check paperwork, check temp if chilled, photo if damaged, log in LIMS, label, fridge or shelf, email client if anything's wrong". Seven fragments that everyone in the lab understands and a new technician doesn't.
After the drafting and gap prompts, and the owner's answers, the core of the SOP read like this (illustrative):
- Match each container to the chain-of-custody form by client reference. IF a container has no matching line, THEN place it in the quarantine tray and email the client the same day.
- For chilled samples, measure the temperature of the temperature blank with the probe thermometer. IF it is above the limit on the client's method sheet, THEN record the reading, photograph the cool box and contact the client before testing.
- IF any container is cracked or leaking, THEN put it in secondary containment, photograph it and tell the duty supervisor.
- Log each sample in the lab system and print its label; stick the label on the container body, not the lid.
Two questions from the gap prompt changed the procedure itself: nobody had written down that the temperature limit comes from the client's method sheet rather than a single lab-wide figure, and "the lid" labelling rule existed only in one technician's head. Good SOP work regularly improves the process, not just the paperwork.
Sign-off, version numbers and where SOPs should live
Give every SOP an owner (a role, not a person's name, so it survives staff changes), an approver, a version number and a review date, typically 12 months, or sooner after any incident or equipment change. Keep a one-line change log: "v1.2, drawing-revision check moved to step 1 after rework on batch 3187".
Store the master copy in one shared folder, with the date in the file name, and put the working copy where the task happens: a QR code on the machine guard that opens the current version, or a laminated card for steps done with gloves on. Old versions go into an archive folder that nobody links to. Once SOPs live in one place, you can also point an internal assistant at that folder so staff can ask "what's the coolant check limit?" and get the answer with the SOP number.
Updating an SOP when the process changes
SOPs rot when updating them is a chore. AI makes the update quick if you give it the current version and a short note of what changed, and ask for a marked-up revision plus a change summary, rather than a fresh rewrite that quietly alters things you didn't mention.
Here is SOP v1.2 and a note of what has changed. Produce:
1. The revised SOP, changing only what the note requires.
2. A change list: each changed step, old wording, new wording.
3. Any other step that the change might affect, as a question.
CHANGE NOTE: [paste]
An illustrative case from a courier firm's proof-of-delivery SOP: the change note said "drivers now take the doorstep photo in the new app before capturing the signature, and 'leave safe' needs the customer's written instruction in the job notes". The change list came back with steps 4 and 6 reworded and a new IF / THEN at step 5. Its question was the useful part: "Step 8 says 'if no one answers, card the address and return the parcel'; does 'leave safe' now come before carding when the job notes allow it?" It did, and nobody had thought to change step 8. Publish the new version with the change list at the top for a fortnight, so drivers who knew the old version see what moved.
Spin-offs: a checklist, a pocket card and a quiz from one SOP
A finished SOP is the source for three shorter formats, and AI produces them in a minute each:
From the SOP below, produce:
A. A tick-box checklist of the critical steps only (max 12 lines).
B. A pocket card: the 5 things most often got wrong, one line each.
C. Five multiple-choice questions to check understanding, with answers
and the step number each one tests.
Use only what's in the SOP.
An illustrative pocket card for the changeover: "Drawing revision matches traveller. Every tool checked, not just T3. Skim the face before setting Z zero. First part at 50% rapid. Tolerance 0.05 mm: out means adjust and rerun." The checklist feeds straight into building opening and closing checklists with AI, and the quiz questions slot into a staff training matrix so you can record who has been signed off on which procedure.
Five ways AI-drafted procedures fail on the shop floor
- Generic safety text buries the real hazard. "Always wear appropriate PPE" appears above the one warning that matters ("never open the guard while the spindle is coasting"). Delete generic lines; keep specific ones, ideally next to the step they apply to.
- The order follows the talk-through, not the task. People recording a memo jump back ("oh, before that I..."). Check the sequence against a real run.
- Decisions get flattened. "Adjust the offset" hides the fact that some deviations mean stop and call someone. Ask the model to list every IF in the procedure and confirm each has a THEN.
- Numbers drift. A tolerance becomes 0.5 instead of 0.05, or gains a unit it never had. Compare every number in the draft with the notes; it takes two minutes.
- The SOP gets too long to use. More than two pages and people stop reading. Split it: "Changeover" and "First-off inspection" work better as two short documents than one long one.
Where a procedure is safety-critical or regulated, such as lifting operations, pressure systems or accredited test methods, the AI draft is just a starting point, and a competent person has to review it against the relevant standard. The same care applies to drafting risk assessments with AI, which often sit alongside SOPs.
Checking the SOP works a month later
Pick one number the procedure was meant to move before you publish it, and look at it after four weeks. For the changeover SOP that was first-off rejects: an illustrative shop logged 9 in the month before and 3 in the month after, and two of the three came from a setter who hadn't yet been walked through the new version. Other good signals are the questions people still ask the supervisor (if the same one comes up twice, the SOP needs a line), and whether the working copy on the shop floor matches the master. If nothing measurable exists, ask the three people who do the task most what they skip and why; the answers tell you which steps are unrealistic.
Time, tools and what a batch of ten SOPs costs
Any of the main assistants handles this well: ChatGPT, Claude, Gemini or Copilot. Use a business account if your procedures contain trade secrets or customer details; business plans such as ChatGPT Business and Claude Team don't train on your content by default, and both need at least two seats at around $25 a seat on monthly billing. Transcription on a phone is free.
| Stage | Typical time per SOP (illustrative) |
|---|---|
| Capture (talk-through plus photos) | 20 minutes |
| Draft prompt and first read | 15 minutes |
| Gap questions answered by the owner | 15 minutes |
| New-starter test and fixes | 30 minutes |
| Sign-off, versioning, filing | 10 minutes |
| Total | About 1.5 hours |
Writing the same SOP from a blank page usually takes a supervisor half a day, and it tends not to get done at all. At 1.5 hours each, ten core procedures is a two-week side project for one person working an hour or two a day. If the procedures you're writing are also candidates for automation, mapping the process before you automate it uses the same notes and saves doing the interviews twice. And if your team's knowledge lives in undocumented habits across several sites or shifts, documenting processes before adding AI covers how to agree the standard version first.
Further reads
- Writing Pharmacy SOPs With AI: A Step-by-Step Workflow — The same method applied in a tightly regulated setting.
- How to Build a Company Knowledge Base AI Can Answer From — Store finished SOPs so an assistant can answer questions from them.
- AI Employee Onboarding: Plan a New Starter's First Two Weeks — Build SOPs into a new starter's first two weeks.
- How to Build a Shared Prompt Library for Your Team — Save the drafting and gap-finding prompts for the whole team.
- AI Quality Control for Small Manufacturers: Where to Start — Where inspection SOPs meet AI-assisted quality checks.
- How to Create SOPs From Screen Recordings With AI — For computer-based tasks, record the screen instead of notes.
- Reduce Owner Dependency: Use AI to Capture What Only You Know — Get what's in your head into a form staff can use: an interruption log, an AI interview prompt, decision rules with thresholds, and a shared assistant.
- How to Build a Kitchen Training Manual With AI — Capture how your kitchen really works, let AI shape it into station cards and quizzes, and keep every temperature and allergen under the head chef's control.
- Turn Fixed Tickets Into a Searchable AI Knowledge Base — Turn a year of closed IT tickets into tested, searchable fix articles: selection rules, redaction, root-cause clustering, a template and upkeep.
- How Surveyors Use AI to Turn Site Notes Into Reports — A dictation routine for site visits, transcription options, a report prompt with sample output, a paragraph library and the checks behind your signature.
- How Funeral Directors Can Use AI for Admin With Care — Where AI can take funeral admin off your desk, where it must never speak for you, and the read-back routine that stops a misspelt name reaching print.
- Where Should a Small Manufacturer Start With AI? — Why a small manufacturer's first AI project belongs in the office, how to score the options, and a six-week RFQ pilot with real numbers.
- Can Farmers Use AI for Farm Records and Paperwork? — How farmers can hand the typing of spray logs, receipts and forms to AI, and the records that must still be checked against the source.
- Safe AI Uses for Counsellors That Never Touch Client Notes — Twelve ways a counsellor in private practice can use AI for the business side, with examples, plus the settings that keep client data out.
- How to Write an Employee Handbook With AI and What to Check — A staged method for drafting a small-firm handbook with AI, plus the checks that catch imported legal figures, accidental promises and contradictions.
- How to Write a Health and Safety Policy With AI and What to Check — Draft a three-part health and safety policy with AI from your real hazards, then check it on a walk-round before anyone signs it.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: vendor documentation for ChatGPT, Claude and Gemini business plans (training defaults); facts sheet for plan prices. The procedures, timings and outputs shown are illustrative.