Are AI-generated RAMS HSE-compliant? An honest 2026 assessment

AI-generated RAMS are now widely available and getting cheaper. But are they actually HSE-compliant? Will they pass main contractor pre-start checks? Are the legislation references correct? Will the controls actually work on a real site? The honest answer in 2026 is "it depends on the AI" and this guide explains what to check, what to verify, and how to use AI-generated RAMS safely.

The legitimate concern

Every UK trades business owner who has tried an AI tool has had moments of unease. The output looks confident and well-written. But how do you know the legislation references are accurate? How do you know the controls actually address the hazards? How do you know the AI did not just make something up that sounds plausible?

This concern is correct. AI tools can produce confident-sounding but incorrect output, especially when asked to cite specific legislation or technical standards. For RAMS specifically, getting the legislation wrong is a problem because the document loses credibility if a main contractor spots an error.

Two categories of AI RAMS tool

AI tools producing RAMS fall into two categories.

Generic AI

Asking ChatGPT, Claude or Gemini to "write me a RAMS for a roofing job" produces a generic output. The structure is roughly correct. The hazards listed are reasonable. The legislation references are usually correct for the major regulations (CDM 2015, Work at Height Regulations) but can be wrong on more specific items (HSG numbers, BS standards, specific HSE bulletins). The output reads like a competent but inexperienced person wrote it.

Generic AI is fast and cheap but the quality varies and the user needs to manually verify legislation references. If you do not know the right answer, you cannot tell when the AI is wrong.

Domain-trained AI with hazard libraries

Specialist tools like Complys use AI on top of pre-built hazard libraries that were curated by humans for each trade. The AI does not invent hazards; it picks from a verified library. The AI does not invent legislation; it references the standards in the library. The output quality is more consistent because the underlying knowledge is fixed.

This category is more reliable but only as good as the underlying library. If the library is sparse or inaccurate, the output will be too.

What HSE compliance actually means for RAMS

Strictly, HSE does not 'approve' RAMS documents. There is no certification scheme. The HSE expects:

A suitable and sufficient risk assessment per MHSWR 1999
Safe systems of work documented per HSAWA 1974
Compliance with CDM 2015 for construction work
Compliance with trade-specific regulations (Work at Height, COSHH, CLAW, GSIUR, etc.)

An HSE-compliant RAMS therefore needs to show:

Hazards have been systematically identified for the specific work
Risk has been assessed using a recognised method (severity x likelihood)
Control measures are appropriate to the hazards (in HSE hierarchy of control)
The method of work eliminates or reduces residual risk to acceptable levels
Personnel are competent for the activities described
Plant and equipment are properly inspected and used
Emergency procedures are in place
Document is reviewed when conditions change

Any RAMS - AI-generated or human-written - that addresses these eight points is HSE-compliant. The question becomes: does the AI tool produce output that addresses all eight?

The reliability question

For AI RAMS to be reliably HSE-compliant, three things need to be true.

1. The hazard library must be comprehensive and accurate

Generic AI produces hazards from its training data, which is variable. Domain-trained AI uses a fixed library curated by humans. The library should reference the correct HSE guidance documents, the correct British Standards, and the correct legislation for the trade. Verify before relying.

2. The legislation references must be current

UK construction legislation changes regularly. CDM 2007 was replaced by CDM 2015. The Work at Height Regulations were amended in 2007. RIDDOR was significantly updated in 2013. Building Safety Act 2022 introduced new duties. An AI tool that references outdated legislation produces non-compliant RAMS.

3. The output must be project-specific

Generic boilerplate fails. The AI must take inputs about the actual project (scope, location, hazards, plant, sequence) and produce output that reflects those inputs. If you can swap one project's RAMS for another without changing anything, the AI is not doing the job.

Where Complys sits in this

The Complys AI RAMS builder is in the second category - domain-trained AI on top of curated hazard libraries. The libraries were built by working trade professionals (the company was founded by a working scaffolder, DDC Scaffolding) and reviewed for HSE alignment.

Each trade has its own hazard library:

Scaffolding: 28 hazards referencing TG20:21, BS EN 12811, SG4:22, SG6:23, HSG150
Roofing: ~20 hazards referencing HSG33, NFRC technical bulletins, BS 5534, CLAW 2002, INDG327
Electrical: 12 hazards referencing BS 7671, HSG85, HSG107, EAW 1989, BS 7430
Plumbing: 12 hazards referencing WRAS, BS 6700, BS EN 806, COSHH 2002
Gas engineering: 12 hazards referencing GSIUR 1998, BS 6172, IGEM standards, HSG274
Groundworks: 12 hazards referencing HSG47, HSG185, CIRIA C760, BS 6031
Carpentry: 12 hazards referencing INDG401, COSHH for wood dust
Decorating: 12 hazards referencing CLAW 2002, COSHH 2002, INDG136
Plant: 12 hazards referencing PUWER 1998, LOLER 1998
Vehicle: 12 hazards referencing PUWER, Driver Hours, Road Traffic Act
Arboriculture: 12 hazards referencing HSE Forestry Industry Advisory Committee guidance
Contractor: 12 hazards covering principal contractor scope under CDM 2015

The AI generates RAMS from these libraries based on user inputs about the specific job. The output references the right legislation because the libraries reference the right legislation. Every section is editable so the user can adjust for specifics the library does not cover.

What to verify before relying on AI RAMS

Even with a domain-trained tool, you should verify before sending RAMS to a main contractor:

Are the legislation references current? Look for CDM 2015 (not 2007), the latest BS standards (e.g. BS EN 12811 not BS 5973), and the most recent HSG numbers.
Are the hazards specific to YOUR scope? Generic boilerplate gets rejected. Your inputs should produce outputs that match your job.
Are the controls in HSE hierarchy? Engineering controls first, then administrative, then PPE. PPE-only solutions where engineering controls are practical fail review.
Is the method statement realistic? Read it through. Would a competent person actually do the job in this sequence? If not, edit.
Is competency referenced? Specify the cards or qualifications required, not just "trained personnel".
Are emergency procedures site-specific? "Call 999" is not enough. Reference the site arrangements.

The honest conclusion

Yes, AI-generated RAMS can be HSE-compliant when the AI is built on a verified hazard library and the user reviews the output before submission. Complys is built specifically for this purpose. Generic AI tools (ChatGPT, Claude, Gemini) can also produce useful first drafts but require more verification.

The pattern that works in 2026 is: AI generates the document fast (30 seconds), the trade specialist reviews and edits for project-specifics (5-10 minutes), the document goes to the main contractor for pre-start review. Total time: 15 minutes vs 1-2 hours from scratch.

See the Complys RAMS builder or start the 90-day free trial to test the output for your trade.

FAQ

Has any HSE inspector ever criticised an AI-generated RAMS?

Not specifically that we have heard of. HSE inspectors look at content, not authorship. A well-structured RAMS from any source passes. A poor RAMS from any source fails. The author tool does not enter the assessment.

What if the AI gets a hazard wrong?

You edit it. Every section of a Complys RAMS is editable before download. The AI gets you to 95% in 30 seconds; you handle the 5% the AI cannot know.

Can I use AI RAMS for high-risk jobs?

Yes, but with more care. High-risk work (deep excavation, asbestos, working at height over 3m, hot works in confined spaces) deserves more verification. The AI gets you a starting structure; the trade specialist verifies the specifics.

Are AI RAMS accepted by SSIP schemes (CHAS, SMAS, SafeContractor)?

SSIP schemes assess whether the RAMS process is competent, not whether it was AI-generated. As long as the output is project-specific, addresses the hazards properly, and references the correct legislation, the source does not matter to SSIP.

Related guides

What is a RAMS document and do you need one? - the plain-English explainer
RAMS template free UK 2026 - what makes a good template
Method statement examples UK 2026 - real examples for several trades
RAMS vs risk assessment - the difference and when each is needed