Start with what has not changed
A management system is a chain of evidence: something happened, somebody recorded it, somebody assessed it, somebody acted, and a document can be produced for each link. None of that is a writing problem, and that is why the arrival of a tool that writes extremely well changes less than it first appears to.
What it does change is the cost of the parts that were always clerical — restructuring, summarising, translating, drafting the first version of something you were going to rewrite anyway. That is a real saving and it is worth taking. It is not the same as the system getting better, and the distinction matters because an auditor will ask for the evidence, not for the prose.
Where it genuinely helps
Turning a long document into a structure. Give it a procedure somebody wrote over ten years and ask what the actual steps are, who owns each, and what is missing against a clause you name. It is good at seeing shape in prose, and shape is what a badly written procedure hides.
Moving between Arabic and English without losing a technical term. This is the one task where it is reliably better than the alternative most teams use, and bilingual operations lose a great deal of time here. Check the terms that carry legal or standard-specific meaning; the rest will be fine.
First-pass sorting of a backlog: a thousand observations that nobody has read, grouped into themes so a human decides what matters instead of reading a thousand lines to find out there are nine themes. Treat the grouping as a draft, not a finding.
Arguing with your own thinking. Ask it to attack a risk assessment you wrote, or to say what an auditor would ask about a control you are confident in. It is a cheap and patient second reader, and it does not mind being asked the same thing four ways.
What it must not be allowed to decide
Whether something is legally required. A model will answer a question about a regulation in the same confident register whether it knows the instrument or is reconstructing a plausible one, and regulations are exactly the kind of text it has seen a great deal of and remembers imprecisely. Legal requirements come from the instrument itself, cited, every time.
Whether a control is adequate. Adequacy is a judgement about this plant, these people, this shift pattern and what actually happens at three in the morning. A model has none of that and will not tell you it has none of it — it will produce a reasonable-sounding assessment built from the average of everywhere.
Why an incident happened. Causation is established from evidence, sequence and testimony, and the failure mode here is specific: a model is very good at producing a causal story, and a plausible causal story is the single most dangerous output in investigation work because it ends the search.
Anything where being confidently wrong is worse than being slow. That is the test, and it is a better one than any list: if the cost of a wrong answer that looks right is higher than the cost of doing it yourself, do it yourself.
The data rule, before anything else
Incident records are personal data. They carry names, injuries, medical detail and sometimes disciplinary outcomes, and that is among the most sensitive material an operation holds about its people. Pasting one into a public tool is a disclosure, whatever the intention was.
In Saudi Arabia that engages the Personal Data Protection Law, and the practical consequence is simple enough to apply without a lawyer: strip identifiers before the text leaves your control. A model can restructure an investigation report perfectly well when the injured person is "the operator" and the plant is "Site A".
Check what your employer has actually approved before any of this. Many operations have a position on these tools and most people have not read it; a technically sound workflow that breaches your own IT policy is still a breach.
How to tell a useful answer from a plausible one
Ask it for the source and then open the source. Not "is this right" — it will tell you it is — but the clause number, the document, the page. The question that exposes most errors is the dullest one: where exactly does this come from.
Notice when it agrees with you. These tools are built to be agreeable, so agreement carries almost no information. If you push back on a correct answer and it folds, you have learnt something about the tool rather than about the answer.
Give it the document rather than relying on its memory. Nearly every bad answer in this field comes from asking it to recall a standard instead of handing it the clause and asking what the clause says.
The part auditors will ask about
A management system is auditable because each step has an owner and a date. An output with no author and no provenance is not evidence, and "the AI produced it" is not a competent person reviewing it — which is what a clause requiring competence actually asks for.
The practical answer is unglamorous: whoever signs it owns it. Use the tool to produce the draft, then review it as the competent person and sign as that person. Record that a drafting tool was used where your own policy requires it. Nothing in that is new — it is how organisations have always handled work produced by one person and signed off by another.
If you are starting this week
Pick one task that is clerical, repeated, and harmless if it comes out wrong. Translating toolbox-talk material. Turning a handwritten inspection backlog into a table. Summarising a month of observations into themes for a meeting. Do it for a month and keep a note of where it saved time and where it cost you time in checking.
That note is the thing worth having. It is specific to your operation, it will not match anybody else’s, and it is a far better basis for deciding where this belongs than any general claim about the technology — including this page.