The Shift Report · Issue 9 · 26 July 2026
AI-generated procedures in operations: who reviews, who approves, who owns the mistake
Drafting a procedure got cheap in about 18 months. Approval hasn’t moved at all. Every AI-assisted document still needs a named source, a competent reviewer, a controlled version, and evidence that the instruction works at the point of use.
A supervisor asks an AI tool to write a safe procedure for clearing a conveyor jam.
The answer looks respectable. A purpose, a list of personal protective equipment, numbered steps, a warning about stored energy.
It may still miss the isolator fitted at your second site. It may assume a guard design you replaced 3 years ago. It may describe the normal job perfectly and say nothing about the moment the belt moves after isolation.
This stopped being hypothetical in April. The FDA issued a warning letter to Purolea Cosmetics Lab, and among the findings was that the company had used AI to create specifications, procedures and master production and control records, then put them to use without reviewing them for accuracy and compliance. The regulator’s point was narrow and old: someone competent reviews the document before the work runs on it.
Drafting got cheap in about 18 months. Approval hasn’t moved at all.
That gap matters because a document becomes official very quickly. Add the company logo, put it in the right folder, assign it through the training system, collect 40 electronic signatures. The record looks tidy. The work may still be exposed.
The useful jobs for AI here are real. Rough notes into a first draft. Translation. Comparing versions. Finding the section nobody wrote. Turning a 14-page procedure into a toolbox talk. That’s hours of administration gone, and it’s worth having.
Approval stays with someone who knows the law, the equipment, the site and how the task goes when the shift is running late. The operator who actually does the job needs a real chance to challenge it. Then training has to show that people can use the procedure, including the stop conditions and the exceptions.
That gives you a standard you can apply on Monday morning. Every AI-assisted document gets a named source, a competent reviewer, a controlled version, and evidence that the instruction works at the point of use.
The tool can produce the words. Your operation puts its name on them.
Three things worth knowing
1. The first enforcement action naming AI in the document chain has already landed.
An EHS Today article published on 20 July sets out the controls AI-generated policies, procedures and training need: an inventory of where AI is used, competent review, source verification, version history and regular audits. It also separates recording course completion from proving competence. Its anchor is the FDA’s April warning letter to Purolea Cosmetics Lab, where AI-generated specifications, procedures and production records went into use unreviewed, against 21 CFR 211.22(c).
My take: that’s a pharma regulation and most of you aren’t in pharma. The question behind it is the one every auditor asks in every sector: who approved this, and on what basis? Keep one approval route for every controlled document, whatever produced the first draft, and add 4 fields to the record. Who supplied the source material. Who checked the technical content. Who checked it against the real task. How competence was verified. Generation should make the preparation faster and the approval stricter.
2. When AI takes the easy jobs, the shift can get harder.
Safe Work Australia’s new guidance warns that automating routine tasks can leave workers with a greater proportion of complex, cognitively demanding work. The result can be more intense workloads and fatigue, even when the system removes time from the process.
That risk is easy to miss in a project measured through transactions per hour.
My take: assess the whole shift before automating a task. Look at the mix of routine work, exceptions, decisions and recovery time. Then map what remains after implementation. If 2 hours of straightforward work disappear, decide how people will rotate, take breaks and escalate difficult cases. Time saved at one process step can become fatigue somewhere else.
3. The robot demo should include something going wrong.
A TCS survey of 300 manufacturing leaders, published on 22 July, found that 77% expect physical AI to have a significant or transformational impact on warehouse operations. Yet 68% remain at the experimental stage or have not deployed it.
The survey comes from a company selling physical AI services, so treat the figures as direction rather than evidence of operating performance.
My take: put failure into the acceptance test. Block a sensor. Present the wrong load. Interrupt the connection. Have someone enter the working area. Watch what stops, what keeps moving, who receives the alert and what evidence is required before restart. A successful supplier demonstration shows the machine can work. A useful one shows your team can recover when it does not.
One thing to try this fortnight
Pick one procedure that has been drafted, rewritten, translated or summarised with AI. If you don’t think you have one, ask. You have one.
Put 3 people around it:
- the person who owns the process;
- an experienced person who does the task;
- someone who has never done it.
Ask them to mark 4 things:
- a step that depends on local equipment or layout;
- a point where an experienced person uses judgement;
- a condition that should stop the job;
- a part a new starter could follow incorrectly.
Then change the document and run the task, or a safe simulation of it. Record who approved the final version and what you used to check that people understood it.
No new system needed. Twenty minutes with the right 3 people will tell you whether the procedure describes the work or only resembles it.
Final thought
AI will put far more procedures into circulation, because it takes most of the effort out of writing them. Approval, worker review and competence checks have to keep pace, and right now they are the slowest part of the chain in almost every operation I see.
That’s close to the problem I’m working on with SOPwise: turning operational know-how into instructions and training people can use, while keeping the source, the review and the decision points visible.
An approved procedure has to survive the shift.
SOPwise turns a Standard Operating Procedure into a training package: task-based steps, critical mistakes, and scenario questions. £35 per SOP, no subscription. Built for UK manufacturing and chemical distribution.