For the trades
Yes, and it will do it in two minutes, and it will read better than the one you wrote last year. That is the problem. A RAMS that reads well and says nothing about your site is the document that gets a firm into trouble — and AI produces exactly that, with total confidence, unless somebody who knows the job goes through it. Here’s how to use it without that happening.
AI can draft a RAMS or method statement, and nothing in UK law says it can’t. But the law requires a suitable and sufficient risk assessment for the specific job and site, produced by people with the skills, knowledge and experience to do it. An AI draft is a generic starting point; a competent person at your firm has to make it site-specific and take responsibility for it. The firm, not the tool, is accountable for what goes to site.
Published
Use it for the draft. Never for the thinking. An AI tool is a fast typist who has read a thousand method statements and has never seen your site. It will give you a structure, the standard hazards for the type of work, and tidy wording for the control measures you tell it about. Everything specific — the access, the services, the neighbours, the sequence, the weather on a Cornish cliff in February — has to come from a person who has stood on the job.
Then that person signs it, and that signature means what it always meant: I have assessed this work and I stand behind this document. The HSE will never ask whether AI typed it. It will ask whether it was right.
Three things, and none of them mention who holds the pen.
The Management of Health and Safety at Work Regulations 1999 require every employer to assess the risks of the work. The HSE’s own words are “suitable and sufficient” — which it explains as proportionate, specific to the work, and covering the significant hazards. A generic assessment that could apply to any site is the textbook example of one that isn’t.
CDM 2015 requires that anyone appointing or carrying out construction work has, or employs people who have, the skills, knowledge, experience and — for organisations — the capability to do it safely. That competence sits in a person. A tool that has read a lot of RAMS has none of it.
For demolition, dismantling and structural alteration, CDM 2015 requires the arrangements to be recorded in writing before work starts. For everything else the HSE says a method statement isn’t required by law but is prepared for most construction work anyway, because it’s how you plan the job step by step and brief the people doing it — and most principal contractors insist on one. Its whole value is that it describes this job in this order on this site. A method statement for “a job like this” has no value at all.
Responsibility doesn’t move. If an AI-drafted RAMS misses the overhead line and someone is hurt, the firm that issued the document is the one answering for it. “The software wrote it” is the same defence as “the lad wrote it”, and it has the same chance.
Plenty, and it would be dishonest to pretend otherwise. Used as a drafting tool, it will:
That last question — what have I missed — is the most useful thing you can ask it, and the one almost nobody does. Used that way it makes a competent person’s RAMS better. Used to replace the competent person, it makes it worse and more convincing at the same time.
Ask for a RAMS for “replacing a flat roof” and you get a RAMS for the average flat roof. Yours has a fragile skylight, a neighbour’s conservatory under the eaves and no rear access. None of that appears unless you put it there.
“Edge protection will be installed to all open edges.” Will it? By whom, with what? A RAMS that promises a control you haven’t arranged is worse than one that says nothing, because the lad on the roof thinks it’s coming.
The overhead line across the yard. The asbestos soffit nobody mentioned. The school run at 3pm past the only gate. These are the hazards that hurt people, and they are precisely the ones a tool that has never seen the site cannot know.
AI writes in the confident, finished voice of a document that has been checked. That’s the trap: the better it reads, the less anyone reads it. A rough draft gets questioned. A polished one gets signed.
By someone who has been to the site, or is going to be running it. Not the office, and not the person who typed the prompt.
Access and egress — the real ones. Where the van parks, how materials get in, where people get out if it goes wrong. If the draft says “site access via main entrance” and the site is a terraced house, start again.
Services. Overhead lines, buried cables, gas, water, drainage. Has someone actually checked, and does the document say what was found and how it’s being dealt with?
The sequence. Read the method statement as if you were doing it tomorrow. Does step four depend on something that doesn’t happen until step six? AI gets order wrong more often than it gets hazards wrong.
Every control measure is real. For each one: who is providing it, when, and is it on the order? Delete anything you can’t answer that for, and deal with the hazard another way.
Who’s actually on the job. The draft assumes a trained, experienced crew. If the labourer started Monday, the supervision and the briefing need to say so.
The thing only you know. Every site has one — the neighbour, the dog, the slope, the tide, the fact that the client’s father-in-law will be “helping”. If it isn’t in the RAMS, the RAMS isn’t for this site.
A RAMS is full of personal data once it’s finished: names of operatives, their tickets and card numbers, the client’s name and address, an emergency contact. None of that is needed to draft it, and all of it should stay out of a general AI tool — the same rules as using AI for quotes and emails.
Site visit first, with notes. The person who’ll run the job writes down what they saw — access, services, neighbours, the awkward bit. Five bullet points on a phone is enough.
Draft with AI from those notes, plus your standard wording for the type of work. Roles, not names.
Ask it what you’ve missed. Then decide, yourself, whether each suggestion applies to this site.
The competent person goes through the six checks and changes what needs changing. This is the step that can’t be skipped and can’t be delegated to the tool.
Brief it to the crew, and sign. A RAMS nobody has been talked through is a filing exercise. The briefing is where it starts protecting people.
Done this way, AI takes an hour of typing out of every RAMS and puts nothing at risk. Done the other way — prompt, print, sign — it takes the hour out and puts the firm on the hook for a document nobody read. The tool is the same in both cases. The difference is whether a competent person was ever in the loop.
This is plain-English guidance for small businesses, not formal legal advice. Where something genuinely needs a solicitor or a data-protection specialist, I'll tell you straight.
Free template. What staff may and may not put into AI tools, which tools are approved, who checks the output. Prevents most incidents before they happen.
Names and addresses, key-safe codes, photos of someone’s kitchen. What a trade can safely paste into an AI tool, what it can’t, and the ten-minute fix.
Written by someone who spent forty years on the tools. What genuinely helps a working trade, and what is sales patter.
Kelvo lets your staff ask about your RAMS templates, site rules and procedures and get answers only from what you’ve approved — and brings you the questions nobody has answered yet. For Cornish contractors with staff.
Kelvo for contractors