An AI SOP generator is good at exactly one thing: turning a vague description of a task into a structured, sequenced first draft. It is reliably bad at everything that makes a procedure yours — your equipment, your numbers, your escalation path, and the safety callouts that only exist because something went wrong once. Get that split right and a usable procedure takes twenty minutes. Get it wrong and you publish a document that reads beautifully and cannot be followed.
I built one of these tools and then watched people misuse it, which is a specific kind of education. The failure is never the draft. It is that someone reads a plausible draft, recognises nothing obviously wrong, and hits publish. What follows is the method that works, and it works with any AI tool — ours, a general chatbot, whatever your company has already paid for.
What does an AI SOP generator actually get right?
Three things, and they are worth more than they sound.
Structure. It knows a procedure needs a purpose, a scope, numbered steps, and a review date, and it will produce all of them without being asked. Most hand-written SOPs are missing at least two.
Sequence. For any common operational task, the ordering will be broadly right. Setup before operation, checks before service, clean-down before close. You are editing an order, not inventing one.
The blank page.This is the real one. The reason your procedures are not written is not that writing them is hard; it is that starting is hard, and a manager with eleven minutes between shifts will not start. An AI draft converts the job from “write a procedure” into “correct a procedure,” and people will do the second task on a Tuesday afternoon when they would never do the first.
What does it always get wrong?
Five categories, in the order they will bite you.
- Specifics. It does not know your machine, your till system, your supplier, or which of your three fridges is the one that runs warm.
- Numbers. It will produce times, temperatures, tolerances and pressures that sound authoritative and are generic at best. Every number in an AI draft is a placeholder.
- Escalation.It will write “inform the duty manager” without knowing whether you have one, whether they are on site at 6am, or what happens when they are not.
- Scarred knowledge. The callouts that exist because someone was burned, a delivery was rejected, or an inspector asked a question you could not answer. Those live in people, not in training data.
- Judgement. The draft tells you what to do when things go normally. Your team already knows that. What they need is what to do when the reading is out of spec.
Treat every number as a placeholder. If an AI draft says a machine takes fifteen to twenty minutes to reach temperature, that is a guess about machines in general, not a measurement of yours. Someone stands there with a timer once, and then the number is real.
What should you put in the description?
The single biggest quality lever is the input, and almost everyone writes it badly. A description like “cleaning procedure” produces a generic cleaning procedure, because that is all you asked for.
Write four things, in two to four sentences:
- Who does it and when.“The closing supervisor, at the end of every trading day.” Role and trigger, not a name.
- Where it starts and where it ends.Boundaries stop the draft from sprawling into three adjacent processes. “Starts when the last customer leaves, ends when the alarm is set.”
- The equipment and systems by name.Not “the till” but the system you actually use. Named things anchor the draft to something concrete.
- The failure you are preventing.“We have twice left the back door unlocked overnight.” This is the line that changes the draft most, because it tells the model what the procedure is for.
Longer descriptions do not produce better drafts. They produce longer ones. Four sentences of specifics beats a page of context every time.
A worked example, and what is wrong with it
Here is the kind of output to expect. The illustrative example published on our free AI SOP generator is an espresso machine opening procedure — switch on and allow the machine to reach operating temperature, flush each group head, check the water filter indicator and tell the duty manager if it shows red, wipe and purge the steam wands, pull a test shot and check it against the recipe card.
As a draft, that is genuinely good. The sequence is right and nothing important is missing. As a procedure you hand to a new barista on Saturday, it is not finished. Here is the edit pass:
| What the draft says | Why it is not usable yet | What you replace it with |
|---|---|---|
| Allow the machine to reach operating temperature, typically 15 to 20 minutes | A plausible general number, not a measurement of your machine | Your machine by name, your warm-up time, timed once and written down |
| If the filter indicator shows red, tell the duty manager | Assumes a role that may not be on site at opening | The named role on your rota, plus what to do if that person is not in |
| Check the test shot against the recipe card | Refers to an artefact you may not actually have | Your target weight and time, or a step that creates the recipe card first |
| Wipe the steam wands with the dedicated cloth | Says what to do, not what good looks like or why it matters | The standard, where the cloth lives, and the consequence of skipping it |
Four edits, six minutes. Notice that none of them required writing prose. The AI did the writing; the human did the knowing.
The two things a human must always add
Everything above is correction. These two are additions, and no generator will produce them because no generator has access to them.
1. The decision rules.What happens when a step fails. If the temperature is out of range, if the count does not match, if the customer says no — what does the person do, who do they tell, and are they allowed to carry on? A procedure that only covers the happy path is a description, not a procedure. This is also the part that stops the “I did not know if I should keep going” conversation you have had four times this year.
2. The evidence and the callouts. What proof each step leaves behind — a photo, a signature, a log entry — and the site-specific warnings that exist because of something that happened here. If the third socket on the prep bench trips the circuit, that belongs in the procedure, and only your team can put it there.
A useful test of whether you have added enough: read the draft and ask what your most experienced person knows that the document does not. Write that down. That gap is the whole value of documenting anything, and it is precisely what writing an SOP your team will actually follow turns on.
How do you test the draft on someone who did not write it?
Two passes, ten minutes each, and they must be different people.
The read-back test. Hand it to someone who has never done the task and watch them attempt it. The rule is that you do not answer questions. Every point where they ask, guess, hesitate or look up at you is a defect — mark it and move on. Six questions means six edits. This test is brutal and it takes ten minutes, and it finds more problems than any amount of re-reading by the author, because the author cannot see their own assumptions.
The expert pass.Then give it to the person who does the task daily and ask a narrow question: not “is this good?” but “what in here is wrong?” The first question gets you a shrug. The second gets you three corrections, because experienced people are much better at spotting errors than at judging documents.
If nobody asked anything and nobody corrected anything, be suspicious. It usually means the procedure is too vague to be wrong.
When is a generator the wrong starting point?
Three cases, and recognising them saves time.
A good template already exists. For standard, well-defined processes, starting from a written template beats starting from a generated draft, because the template already contains the sections people forget. Our staff training SOP and the rest of the free UK and US template library are readable in full with a PDF download and no signup — start there and adapt, rather than generating from nothing. The same applies to routine operational documents like an opening and closing procedure.
Somebody already recorded it. If a walkthrough video exists, use it. The transcript carries the specifics an AI cannot invent — the real names, the real thresholds, the aside about the sticky valve — which is why turning a video into an SOP produces a far more accurate draft than describing the same task from memory.
The process is not agreed. If two shift leads do it differently and nobody has decided which is correct, a generator will produce a confident document that papers over the disagreement, and the team will ignore it. Settle the argument first. If you are not sure what belongs in a procedure at all before you start, the guide to what an SOP is covers the anatomy.
What to do on Monday
Pick the one process you explain out loud most often — the one that generates the same question every fortnight. Write four sentences: who, when, start and end, and the failure you are preventing. Generate a draft. Spend fifteen minutes deleting every invented specific and adding the decision rules and the evidence. Then hand it to someone who has never done the task and shut up while they try.
That is one procedure, in under an hour, that is actually correct. Do it once a week for a quarter and you have thirteen. If you want to try the loop without signing up for anything, the generator gives you three instructions free with a PDF export, and TrainedTeam is free for up to 5 users when you want the finished procedures to live somewhere your team can be assigned them.