The Risk of Trusting an AI-Generated Emergency Plan as Finished

“We ran it through AI and had a full emergency plan by that afternoon.”

 

That sentence should worry you, not reassure you. An AI-generated emergency plan isn’t a drafting task that can be handed to a general-purpose tool and checked later. It’s a judgment task, built entirely on situational awareness: a real, current, physically verified understanding of one specific site. AI doesn’t have that. It can’t get it from a form, and it can’t tell you when it’s missing it. For work this specific, the WLSS team uses people who have stood in the building, not a tool that has only ever read about one like it.

 

The short answer: an AI-generated emergency plan can look like a finished document, but it cannot verify a single fact about the site it claims to cover. Until a qualified person checks it against the real building, it’s an unverified draft, not a plan.

 

The duty, and what the standard actually requires

 

Under section 19 of the Work Health and Safety Act 2012 (SA), every person conducting a business or undertaking (PCBU) carries a primary duty of care: to ensure, so far as is reasonably practicable, that the health and safety of workers and others at the workplace is not put at risk. That duty is the law under WHS (OHS) legislation, and it exists regardless of how a business goes about meeting it.

 

Regulation 43 of the Work Health and Safety Regulations 2012 (SA) turns that duty into a specific, prescriptive requirement. A PCBU must prepare, maintain and implement an emergency plan that provides for, among other things: emergency procedures, including an effective response to each type of emergency; evacuation procedures; notifying emergency services at the earliest opportunity; medical treatment and first aid; effective communication between the person coordinating the emergency response and everyone at the workplace; testing of the emergency procedures; and information, training and instruction for workers. Workers and their health and safety representatives must be consulted when the plan is made and reviewed, under the WHS Act’s general consultation duty.

 

AS 3745:2010 is the recognised, practical process for actually building a plan that satisfies Regulation 43, but it’s a method, not the law itself. Its process is deliberately site-specific: assessing this building, this occupancy, these hazards. A plan that covers the right section headings without going through that site-specific assessment hasn’t met the regulation just because it mentions the right things. And several of Regulation 43’s requirements (testing, training, maintenance, consultation) are ongoing legal obligations. They can’t be satisfied by a document that was generated once and filed away.

 

Where AI breaks: situational awareness

 

“Situational awareness” isn’t a buzzword here. It’s the specific, physical knowledge an emergency plan actually runs on. Which fire door sticks in humid weather. Which supervisor is reliably on site and which one is on the road three days a week. Whether the assembly point printed on last year’s plan is now a delivery bay. Whether the two people listed as fire wardens actually know each other, let alone their roles. A tool only ever has what it’s told. It cannot walk the building, cannot watch how people actually move through it under pressure, and cannot notice the gap between what a form says and what’s true on the day. A questionnaire answered accurately still describes the site secondhand, and an emergency plan built entirely secondhand is a liability with a professional-looking cover page.

 

The usual warning about AI is garbage in, garbage out: feed it bad information, get a bad result. That undersells the actual problem. You can feed a tool good information, answered honestly and in detail, and still end up with a plan that isn’t right, because nothing about the process verifies it. It can’t check whether the assembly point in the answer still exists. It can’t check whether the “clear” fire exit is currently stacked with pallets. It can’t sense that the person named as deputy fire warden left the business four months ago and nobody updated the form. None of it is actually right, not 80 percent right, not mostly right, until a person with the relevant experience has checked it against the real site and confirmed it. Until that happens, it’s an unverified draft wearing the shape of a finished plan.

 

Every part of a genuine emergency plan depends on judgment a document generator cannot supply:

 

What the plan depends on	Why it needs a person on site, not a document generator
Deciding which scenarios are genuinely credible for this site	Requires judgment about this occupancy’s actual hazards, not a generic list
Confirming the physical layout: doors, exits, assembly points	Only confirmed by walking the building, on the day
Structuring the emergency control organisation	Requires knowing who’s actually reliable, present, and capable of leading under pressure
Testing procedures and training workers	A recurring legal obligation under Regulation 43, not a one-off output
Final review and sign-off	Requires someone qualified to be accountable for the plan, documented as such

 

Every row in that table is judgment, and judgment needs a person who has actually seen the site and knows what they’re looking at.

 

 

This isn’t hypothetical

 

Australian regulators are already treating this as a live WHS (OHS) risk, not a future one. Safe Work Australia’s guidance on AI and digital technologies is explicit: when a PCBU introduces or uses AI in the workplace, including to produce safety documentation, it must manage the resulting risks the same way it manages any other WHS risk, by eliminating them, or minimising them so far as is reasonably practicable. An AI-generated emergency plan that hasn’t been verified against the site is exactly the kind of unmanaged risk that guidance is describing.

 

The same underlying failure has already drawn formal regulatory action overseas. In April 2026, the American regulator the FDA (the United States’ Food and Drug Administration) issued a warning letter to a cosmetics manufacturer after an inspection found the company had used AI to generate drug specifications and procedures without proper human review, with the FDA specifically documenting the company’s “overreliance on AI” as a finding. It’s a US case, not an Australian one, but the failure it describes isn’t country-specific: AI-generated compliance material used without qualified human review, treated as finished because it read professionally. That’s precisely the scenario Safe Work Australia’s own guidance now expects Australian PCBUs to be managing as a risk in its own right.

 

And it’s worth being honest about how often that failure actually happens, rather than assuming AI is right most of the time and wrong at the margins. A Stanford University study (US research) testing AI tools against real legal research queries found general-purpose AI tools returned incorrect or fabricated answers in the majority of complex queries tested, and even AI tools built specifically for professional research still returned errors in a meaningful share of cases. The same research noted a gap between the error rates AI vendors advertise and the error rates their tools actually produce on complex, real-world queries. The practical lesson isn’t “AI is usually right.” It’s that nobody using it can tell, from the outside, which answer is the wrong one, which is the whole problem with trusting it on a document that has to be right the first time.

 

What to actually do, and the WLSS view

 

Five steps matter most.

 

1. Have a qualified person conduct site discovery in person: walking the building, not filling in a form about it.
2. Build the plan through AS 3745:2010’s site-specific process, confirming every scenario, procedure and named role against the actual site, not a generic template.
3. Test procedures and deliver training as Regulation 43 requires, and keep the records that prove it happened.
4. Consult workers and their health and safety representatives when the plan is made and reviewed: a legal requirement, not a courtesy.
5. Review and re-verify the plan on a genuine schedule and after any trigger event (a fit-out, a change of use, a change in key personnel), not on a fixed-and-forgotten cycle.

 

The real risk in all of this isn’t only that AI lacks situational awareness. It’s that most businesses have no reliable way to tell when an AI-generated plan is wrong, because catching the error requires exactly the field experience the business doesn’t have in-house. That’s the gap. AI produces a confident-sounding document either way, right or wrong, it reads the same. Without someone who has done this work across enough real sites to know what’s missing, there’s no way to tell the difference between a sound plan and a plausible one.

 

Frequently asked questions

 

 

Why WLSS

 

That’s what the WLSS team’s cross-industry experience is actually for. Not to work faster with AI, but to know what to check and what’s wrong before it costs someone something. The WLSS team doesn’t use AI to build emergency management plans, not because doing it without AI is slower, but because nobody, including WLSS, can reliably validate an AI-generated plan against a site nobody has walked. So the WLSS team walks it, every time, and puts every recommendation in writing.

 

If someone hands you an AI-generated emergency plan and calls it finished, don’t ask how fast they produced it. Ask how they checked it was right, and who is qualified enough to have known if it wasn’t.

 

• Triple ISO certified: ISO 9001, ISO 45001, ISO 14001
• 80+ years combined team experience
• 500+ SA businesses supported

 

Book a Free Advisory Call
Or call (08) 8322 2279