AI-Assisted SWMS and Risk Assessments, Safely

AI is genuinely useful for safe work method statements and risk assessments, in one specific way: it beats the blank page. Point a good model at a task and it returns a structured first draft in seconds, suggests hazards you might have missed, and rewrites dense control measures into something a crew will actually read. It becomes dangerous the moment you let it set the controls or stand in for the people who do the work, because it cannot see your site, know your crew, or carry the duty of care.

I have written, reviewed and rejected more SWMS than I can count, across enterprise construction and social-infrastructure work, and I have built AI tools grounded in the WHS legislation itself. So this is not a pitch for a generator. It is how I would, and would not, use AI to write one.

What is a SWMS, and where does AI fit?

A safe work method statement is the document that says how a high-risk construction task will be done safely. Under the model WHS Regulations a SWMS is required for high-risk construction work, the named categories that carry the most serious risk: work at height with a risk of a fall, work near energised electrical services, demolition, asbestos, confined spaces, work near mobile plant or traffic, and the rest. It has to identify the hazards, set the control measures using the hierarchy of controls, and describe how those controls are put in place, monitored and reviewed. Safe Work Australia is blunt about one thing in particular: a SWMS must be site-specific, and a generic statement may not meet the WHS laws unless it is reviewed and amended for the actual workplace (Safe Work Australia interactive SWMS guidance).

Two things about a SWMS make it a poor fit for full automation. It has to be developed in consultation with the workers who actually do the task, and a competent person has to be satisfied it reflects reality before work starts. AI sits cleanly at the drafting layer underneath both of those. It does not sit at the deciding layer, and the moment it does, you have a document problem wearing the mask of a safety control.

Where does AI genuinely help with a SWMS or risk assessment?

Use AI where a wrong first draft costs minutes, not exposure. In early 2026, around 82% of EHS leaders already used AI to some degree, with about one in five using it extensively (NSC and Wolters Kluwer Enablon, The Safety Shift: EHS Readiness in 2026, retrieved 5 June 2026). That survey is US based, but the direction of travel here is the same. The safe uses cluster in four places.

Beating the blank page. A clear prompt turns a task description into a structured draft SWMS or risk assessment: a sequence of work, candidate hazards, and a control column mapped to the hierarchy. You are editing, not starting from nothing, and editing is where your judgement adds value.

A second set of eyes on hazards. Ask a model to list every hazard a task could present and it will surface things a tired supervisor at 6am might skip: the overhead service, the manual-handling load, the interface with the public. You confirm or discard each one. It is a prompt, not a decision.

Plain English for the crew. SWMS are routinely written in a register no one on the tools reads. AI is good at rewriting a control into a short, direct instruction, which is the difference between a document that sits in a folder and one that changes what people do.

Consistency across many statements. If you run dozens of SWMS across a project, AI helps keep format, terminology and structure consistent, and it can compress a long risk register into the handful of controls that matter for today's work.

Where AI quietly makes a SWMS more dangerous

The failure mode is not a model writing nonsense. It is a model writing something plausible, generic and tidy that gets signed because it looks finished. In the same 2026 survey, 90% of EHS leaders held at least one concern about AI and 65% named overreliance the top one (NSC and Wolters Kluwer Enablon). In SWMS work, overreliance shows up in five ways.

Generic controls that do not match the site. A model trained on the average of the internet writes the average control. It does not know your sequence changed, that the crane now swings over a live road, or that the only access is through an occupied building. The hazard that hurts someone is almost always the site-specific one the generic draft left out. There is a subtler version of this. Practitioners comparing AI and template output keep finding the same skew. The controls bunch up at the weak end of the hierarchy of controls, heavy on PPE and procedure, light on the elimination, substitution and engineering controls that actually protect people (OneClickSWMS). So challenge an AI draft on whether it reached for a higher-order control before it settled for a sign and a toolbox talk.

Confident, wrong references. Ask a general model for the regulation or the standard and it will sometimes quote a clause that does not say what it claims. In a SWMS that is not a typo, it is a control built on a fiction. Ground the model in the actual instrument, and verify what it cites.

Consultation skipped. AI can simulate a hazard list. It cannot consult. The workers who do the task hold the work-as-done knowledge that the procedure and the model both miss, and consultation is a legal requirement, not a nicety.

Automation complacency. The better the draft looks, the less people check it. A polished SWMS invites a signature it has not earned.

Data confidentiality. Job detail, site layouts and people end up in the prompt. Treat that as you would any other sensitive disclosure.

Does the new NSW AI law change this?

In New South Wales it now does, in principle. In early 2026 NSW passed the Work Health and Safety Amendment (Digital Work Systems) Act, the first law in the country to put algorithms, artificial intelligence and automation squarely inside the WHS duty of care (Moore Australia). The duty sits with the PCBU whether the system was built in-house or bought off the shelf (Norton Rose Fulbright).

It is NSW first, not yet fully commenced, and Safe Work Australia is weighing whether the national model laws should follow. So this is direction, not a settled national rule yet. The direction still matters for how you use AI on a SWMS. If a digital tool shapes how work is planned or controlled, the safety risk of that tool is now your risk to manage, not the vendor's. It makes the point the callout just made even harder to argue with: a competent person owns the document, whatever drafted it.

A safe workflow: AI-assisted SWMS, step by step

Here is the division of labour I would actually run. AI drafts and reformats. A competent person decides, consults and signs. The table makes the line explicit.

The order matters. The model produces the draft early, so your time goes into steps 3 to 6, which is where SWMS succeed or fail. Steps 4 and 5 are the ones the law cares about most. Working through the controls and consulting the workers who do the task are duties under the model Code of Practice for managing WHS risks, not nice-to-haves you can let a draft skip. A useful drafting prompt looks like this:

You are helping me draft a SWMS for review by a competent person.
It is not the final statement.
 
Task: [describe the task, plant, site, crew, sequence]
Jurisdiction: [e.g. NSW, model WHS Regulations]
 
Produce:
1. A step-by-step work sequence.
2. For each step, the foreseeable hazards.
3. For each hazard, candidate controls ordered by the hierarchy of controls
   (eliminate, substitute, isolate, engineering, administrative, PPE).
4. A list of the assumptions you made, and the site-specific things you
   cannot know and I must verify on site and in consultation with the crew.
 
Do not claim the SWMS is compliant. Flag anything that needs a competent
person's judgement.

The last instruction is the important one. A model that lists its own assumptions and gaps is telling you exactly where your judgement has to go. That is AI used well: it shows its working and hands you the decisions.

What about AI SWMS generator tools?

They are fine for what they are, and risky when you forget what they are. An AI SWMS generator is a drafting and template tool. It is genuinely useful for structure, formatting and a starting hazard list. It becomes a problem when the marketing slides from "draft" to "compliant", because no tool can know your site or carry your duty. Treat "a compliant SWMS in minutes" as a claim about speed, not about safety.

Before you buy one or roll it out, ask four questions:

1. Where does our data go? Job and site detail is sensitive. Look for enterprise or on-device handling and a clear data policy.
2. Does it tie controls to the hierarchy of controls, or just list them? A flat list of controls with no hierarchy is a red flag.
3. Does the workflow force consultation and competent review? Or does it let a draft be signed straight out of the box?
4. Can it ground its references in the actual regulation and your own documents, rather than a model's memory?

At an organisation level, the cleaner way to think about this is to treat AI as a managed system rather than a clever assistant. Safety teams are starting to govern it the way they govern everything else that touches risk. That means sitting it under ISO 45001 alongside the AI management standard ISO/IEC 42001, with light rules for brainstorming and strict review for anything that becomes a control (Safetysure). Plausibility is not verification, however confidently the output reads.

If you want a fuller way to weigh up a tool, I wrote a separate guide to picking the best AI tool for you. The short version: a tool that makes the human review easy is worth more than one that promises to remove it.

How do you adopt AI here without losing rigour?

AI does not lower the bar on a SWMS. It moves your effort to the parts that need a person: the site walk, the conversation with the crew, the control that actually fits the task. Let it draft, reformat and brainstorm. Keep the deciding, the consulting and the signing human, and ground anything it says about the law in the law itself.

If this is useful, the field guide sets out the wider principle, that AI belongs around the safety decision and not on it, and the piece on AI for incident investigation applies the same test to ICAM. For grounding a model in primary legislation rather than its own memory, see how I encoded the WHS Act into an AI skill. If you are working out where AI fits in your own safety system, reach out, I am always happy to compare notes.