AI for Workplace Safety: What Works Now, and What Does Not

The same question reaches me most weeks, in one form or another. What can this actually do for safety, and where does it fall over. I answer it as the person whose team ships the thing, which makes for a narrower answer than a vendor’s and a more useful one.

The honest answer has three parts, and the third one is new. There is a short list of things AI does reliably today. There is a shorter list of things it should never be handed. And since July 2026 there is a set of Australian obligations that apply the moment you put it anywhere near safety work.

What AI Is Genuinely Good At in Safety Work.

Start with the regulator rather than a vendor. Safe Work Australia’s guidance on AI and digital technologies sets out opportunities beside risks, and the opportunities it names include automating routine tasks, identifying hazards such as fatigue from historical data, and helping identify trends in data to improve WHS strategies.¹ That is a fair description of where the technology is strong.

Here is the same ground from the building side, with what each one looks like on an Australian site.

Reading a Long Document Against Your Criteria.

This is the one that pays for itself. A safety management plan or a method statement for high-risk construction work runs to 40 pages, and the person reading it is checking for perhaps 20 things. A model reads the whole document against those 20 criteria and says what is present, what is partly present, and what is missing.

In ComplyFlow you build that agent yourself. You pick the document type, and a wizard drafts review criteria you can keep, edit, weight, or replace with your own; before you save it you run it against a real sample document and read the actual result.² Once it is live, an agent sits on a form question and reviews the answer and its attachment as the contractor submits it.³

What it does not do is approve anything. It produces a grade and the reasons behind it. A person decides.

Reading a Certificate and Pulling Out the Details.

Contractor compliance runs on documents with dates on them: a white card, a high-risk work licence, a certificate of currency, a plant registration. Reading the number, the class, the name, and the expiry off a scanned certificate is narrow, repetitive extraction, which is exactly what models are dependable at.

It also removes the most common error in any contractor register, which is a date typed in wrongly or not typed in at all. File uploads are among the question types an agent can review,³ so a certificate is read against criteria in the same way a method statement is. The expiry the model reads still lands in an approver’s queue, and the approver still decides.

Finding the Pattern Across Many Records.

A person reading one incident report at a time cannot see that three sites account for most of the near misses on night shift, or that one contractor’s crews fail the same inspection item every quarter. Seeing that requires reading everything at once, which is the one thing a machine does better than a good manager.

Your incident register already charts by type, tag, or question over any period, and inspections already turn every failed item into an action with an owner and a due date. What AI adds is being able to ask the question in plain English instead of building a report: which sites reported repeated near misses last quarter, and whose actions are overdue.⁴

Drafting the First Version of Something.

A draft is not a decision, which is why this use is safe. A method statement for the work in front of you, an inspection checklist that fits the site, an induction drafted from your own procedures: every one of them is faster to edit than to start. The judgement stays in the editing, and the person doing the editing is still the author.

Answering a Question From Your Own Record.

This is the newest of the five and, for a safety manager, the most useful. ComplyFlow is one of the first compliance platforms with an MCP server, the open standard that lets an assistant such as Claude or GitHub Copilot read another system. Ask which contractors at a site have expired insurance documents, or for a summary of the incidents reported at your Western Sydney sites last month, and the answer comes back from the live record.⁴

Two things make that safe rather than alarming. Access is read-only, so the assistant can look up and summarise but cannot create, change, or delete anything.⁴ And it runs on a token you scope to particular modules and set to expire, which inherits your own permissions and site assignments and can never see more than you can.⁵ The connection itself is documented in the Help Centre.

What AI Should Not Be Trusted With.

Four things, and the reason is the same each time. The model has no stake in being right.

Jurisdiction-specific judgement. Australia has nine work health and safety jurisdictions and they are not identical. Ask a general-purpose model what a SWMS must contain, or when an incident becomes notifiable, and you get a confident answer assembled from whichever jurisdiction dominated its training data. Confident and wrong is the worst combination available in compliance work.

Anything that needed a person in the room. A model can tell you a traffic management plan names a spotter. It cannot tell you the spotter could not see the reversing line that morning because a container had been moved overnight. Documents describe controls. People verify them.

Deciding whether a person may be on site. That decision carries a duty, and a person holds the duty. A system can put everything relevant in front of you, including what is expired, what is missing, and what the worker was inducted in and when. The decision still belongs to somebody who can be asked afterwards why they made it.

Any output that goes out without a named human check. This is where the research matters, because the failure mode is not the model’s, it is ours. Automation bias is the tendency to over-rely on automation, and the evidence on it is blunt: a decision aid raises overall performance and, at the same time, stops the people using it noticing the errors the aid itself introduces.⁶

The finding that should bother an experienced manager comes from Parasuraman and Manzey in Human Factors, June 2010. Automation bias, they concluded, “occurs in both naive and expert participants, cannot be prevented by training or instructions, and can affect decision making in individuals as well as in teams.”⁷ Knowing about the trap is not protection from it. A named reviewer and a recorded check are.

Safe Work Australia says the same thing in its own language. Its guidance lists over-reliance on AI to detect hazards as a risk in its own right, with the example of relying on technology to spot fatigue while ignoring a worker’s own indicators.¹ It also warns that when more of a job becomes reviewing what a machine produced, that shift is itself a hazard.⁸ Whoever holds the safety role in your business has to design against that, not only against the obvious risks.

What the Law Now Asks of You.

This is the part that changed, and it is the reason this post needed rewriting rather than tidying.

Safe Work Australia has put AI on the hazard list. Guidance published on 23 July 2026 says that if you are a person conducting a business or undertaking, you must manage health and safety risks from AI and digital technologies as much as you reasonably can, just as you would with any other hazard or risk.⁹ There is no separate AI process to learn: you identify the hazards, assess the risks, control them, and monitor and review the controls, in consultation with workers and their representatives.⁸ The consultation duty bites earliest, because it applies before you introduce a digital technology that may affect health and safety, not after.⁹

New South Wales has legislated. The Work Health and Safety Amendment (Digital Work Systems) Act 2026 received assent on 18 February 2026 and inserts a new section 21A into the NSW Act. It requires a business that uses a digital system to allocate work to ensure that allocation does not create risks, and names excessive workloads, unreasonable performance metrics, excessive monitoring, and discriminatory practices among them. A digital work system is defined broadly as an algorithm, artificial intelligence, automation, or an online platform. The Act has not commenced; the primary duty and the work allocation provisions start on proclamation.¹⁰ Passed and in force are two different facts, and you need both of them.

Australia has no AI Act, and none is proposed. The live guidance is the National AI Centre’s Guidance for AI Adoption, whose implementation guidance was published on 5 May 2026 and sets out six essential practices: decide who is accountable, understand impacts and plan accordingly, measure and manage risks, share essential information, test and monitor, and maintain human control.¹¹ The National AI Plan of 2 December 2025 settled the direction. The government’s approach, it says, will “continue to build on Australia’s robust existing legal and regulatory frameworks”, with agencies and regulators keeping responsibility for AI harms inside their own domains.¹²

Read those six practices next to a safety management system and they are the same document with the nouns changed. If you want the skills side of this rather than the technology side, the five AI skills a WHS manager should be building answers which habits are worth building and how to build them without a course.

What a Safety Team Should Do Next.

Take one page to your next WHS committee meeting. List every place AI already touches your safety work, including the tools nobody approved. Against each one write three things: who is accountable if the output is wrong, what the named human check is and where it is recorded, and whether the output could change a decision about a person’s right to be at work. Everything in that last category comes back to a person, permanently. Then add AI to your risk register as a hazard with controls, because that is now what the regulator expects of you, and get the consultation on the record before the next tool arrives rather than after it has.

That is a morning’s work. The teams I watch get the most out of this are not the technical ones. They are the ones who decided early, and in writing, which questions they were never going to let a machine answer, starting with whether the person standing at the gate is cleared to walk through it.