.png)
You've been handed the climate disclosure, the budget is thin, and the first-wave reports you've read don't look that complicated. So you've opened an AI model, pasted in AASB S2, added some context about the business, and asked for the governance section. It came back in seconds and it reads well.
You can do real parts of this yourself, and AI will make you faster at them. The catch is that a disclosure that reads well and a disclosure that survives assurance aren't the same document, and a general-purpose model can't tell the difference. It won't flag which of its numbers it invented, and it can't produce the evidence your auditor will ask for.
The stakes are specific. Across first-wave Australian companies that disclosed their assurance fee, the median was $89,000. One assurance partner put the gap between a well-prepared and a poorly-prepared client at close to half of that. Preparation isn't overhead on this job, it's the line item.
We've spoken with 100+ Australian businesses on ASRS and with the assurance teams reviewing them, and we build AI tooling for this work, so weigh that as you read. Here's what to own yourself, where AI genuinely earns its place, and where you need a person.
Use it hard, and use it early, on the mechanical work:
One rule of thumb covers most of it. AI is reliable when it's working on material you've given it. It gets risky the moment you ask it to supply material you don't have.
That's the line to hold. Everything below is about the second half of it.
Reporters tell us the volume of evidence surprised them, and that collecting it was the most cumbersome part of the job. Auditors describe the same thing from the other side: missing evidence is where first-year reports fall down.
Auditors don't test your total. They take a sample of line items and trace each one back through the invoice, the fuel card statement, the asset register, whatever sits behind it, to the number that made it into the disclosure. Break the chain anywhere and the total doesn't help you.
A construction business with several hundred vehicles needed a refrigerant figure fast. Someone loaded fleet hours into an AI tool and used what came out. The number looked reasonable and may well have been right. Their own director caught it first, and his email is the cleanest description of this risk we've read:
"That isn't audit evidence. There's no source document, no published reference and no way for the auditor to reperform it."
The same test applies to every statement, not just emissions. Say you meet quarterly on climate risks and opportunities, and you'll be asked for the agenda, the attendee list and minutes from at least one of those meetings.
The practical fix is cheap. As you build the report, write the source document next to every number and claim. Anything you can't name a source for isn't finished, however well it reads.
Physical risk means regional climate data mapped to your actual asset locations under specific warming scenarios. Transition risk means sector and geography-specific pathway data, and a view on which of it is credible. Emissions means knowing which published factor applies to which jurisdiction in which year, and whether an older one needs adjusting.
Ask a general-purpose model any of those and you'll get a confident answer. You won't get a flag on the ones it made up. Push back on a specific figure and it will usually concede the point, which tells you the check works, but you'd have to run it on every sentence and every number in the report. That's not a realistic use of a finance team's week.
So flip the order of work. Get the factor, the scenario data and the asset list from the published source yourself, hand them to the model, and let it do the structuring, drafting and arithmetic you can verify. Facts in, formatting out.
This is also why "we used AI" tells you very little about anyone's disclosure. Two organisations can run the same underlying model and land in completely different places, because one has encoded what good looks like into how it's used: what to ask, what to check, which outputs to distrust, what a specific auditor will query before they query it. The model is the commodity. The judgement around it isn't.
We've spent six months building our climate risk model and dozens of custom skills to get accurate output, and our climate experts still find things to amend. Directed properly, that tooling took SEE Group's climate risk assessment from 13 weeks to two. If you have that expertise in-house, you're in good shape. If you don't, a general-purpose model on its own is thin cover.
In year one, only governance, your identified climate risks, and Scope 1 and 2 emissions are formally assured. Financial effects and scenario analysis aren't, and auditors told us those two generate most of their comments anyway.
The reason is simple. Your auditor reads the whole report and checks it against the parts that are assured. If the strategy narrative doesn't line up with the risk table, you'll hear about it. One assurance lead said her biggest concern on first-year reports was whether the organisation had identified its own climate risks correctly in the first place. Get that wrong and everything downstream points the wrong way.
This is the part AI can't do for you, because it needs to know your business rather than the standard. A model will write a technically correct paragraph on transition risk. It can't tell you whether transition risk is material to your revenue.
Getting that right means your finance and operations people in a room, arguing about which exposures are real. That argument is the capability the legislation was after. A well-written document that skips it leaves you compliant on paper and none the wiser about your own exposure.
Assurance teams tell us the biggest driver of your fee isn't your size, it's how many times they have to re-read your draft. A well-run engagement is two rounds: they review one strong draft, give feedback, and version two is close to final. What they see far more often is five or six versions, and each one is a full re-review rather than a skim.
One client took on their own climate risk assessment and scenario analysis. Two months from their deadline, their auditor said plainly it wasn't good enough to audit. We're rebuilding it now as an urgent project, at a peak-season price nobody would have chosen with three more months on the clock. Money saved by skipping preparation tends to come back with interest, on an invoice from a firm you have far less room to negotiate with, in the July to October window when every other company on your financial year wants the same team's attention.
Whoever does the work, including us, ask the same question of it: is there a named human who has read the output, understands why every claim is there, can point to the source behind it, and will be in the room when your auditor asks? AI inside a process like that is a productivity tool. AI instead of that process is a well-formatted guess.
Then ask what you'll be left holding. A good engagement leaves you with a documented boundary assessment, a risk register with the reasoning still in it, and a team who could walk an auditor through both. If it leaves you with a finished PDF and nothing else, you've bought a dependency. Ask directly what your team will be able to do themselves in year three that they can't today.
And before anything goes near a board pack, whether it came from a model, a consultant or your finance lead, run the one-minute test on every material claim: who in this building can explain out loud where that came from, to an auditor who's never met them? If the answer is nobody, it doesn't matter how well it reads.
If you want a second read on where your own line sits, bring your boundary decisions and evidence gaps to a 30-minute call and we'll tell you which parts you can run yourselves: book a time here.
Parts of it, yes. AI is reliable for structuring the report, drafting narrative from facts you supply, tightening language and locating requirements in the standard. It can't produce audit evidence, decide what's material to your business, or tell you which of its own numbers it invented. Those parts stay with your team or an adviser.
Traceability, not the reported total. Assurance teams sample individual line items and follow each one back to its source: the invoice, the fuel card statement, the asset register, the meeting minutes. If the chain breaks at any point, the number fails, regardless of how the surrounding narrative reads.
Across first-wave Australian companies that disclosed their fee, the median assurance cost was $89,000. One assurance partner estimated the difference between a well-prepared and a poorly-prepared client at close to half the total fee, driven almost entirely by how many rounds of redrafting the audit team has to review.
Governance, your identified climate risks, and Scope 1 and 2 emissions. Financial effects and scenario analysis aren't formally assured in year one, but your auditor reads the full report and checks it for contradictions, so those sections still generate most of the review comments.
Undirected use does, because a general-purpose model produces confident output with no source trail. Used the other way round, with your data and published factors going in and drafting coming out, it lowers risk by cutting the redrafting rounds that drive your fee. The deciding factor is whether an expert reviews the output and signs their name to it.
Trace is a climate reporting platform specialising in ISSB and AASB standards, helping businesses navigate mandatory climate disclosure with clarity and confidence.
London - Sydney
.png)
.png)
.png)
.png)