.png)
Ask a general-purpose AI tool to draft the governance section of your AASB S2 disclosure and you will get something good. Fluent, well structured, confidently compliant. That is exactly the problem.
AI rarely fails ASRS reporting by being obviously wrong. It fails in five predictable ways: it loses the trail back to source, it describes things that never happened, it sounds certain when it is guessing, it does not know your business, and it cannot know what your auditor will ask for. Each one is fixable, and none of them is a reason to stop using AI.
We see these patterns across more than 40 AASB S2 engagements and in the assurance feedback that follows them. If you want the broader view of what you can run yourself, start with our guide to using AI for climate disclosure. This post goes one level down: the failure modes themselves, how to spot each one in your own draft, and what to do about it.
Your auditor does not test your reported total. They pick individual line items and follow each one back to the document it came from. A number can be completely correct and still fail, because the path from source to disclosure was never recorded.
A chat window is not a working paper. Paste twelve months of electricity bills into a general assistant, ask for a total, and you may well get the right figure. What you will not get is anything an auditor can open, check and recalculate. The conversation is gone the moment the tab closes, and nobody can show which bill produced which line.
The provenance gap is a broken chain. The evidence gap is worse: there is nothing at the end of the chain at all.
Generative AI is designed to fill gaps with plausible text. Ask it for a governance section that "complies with AASB S2" and it will describe the oversight a compliant company would have: regular board consideration of climate risk, director training, climate factored into capital allocation. If those things did not happen at your company, you now have a well-written description of a company that is not yours.
This matters most in governance, which is assured in full from year one and is the one area where you generally cannot fix things next year. Once the reporting period closes, the window to create the evidence closes with it.
General AI tools sound exactly as certain when they are guessing as when they are right. In most work that is an annoyance. In climate reporting it is a liability, because the output looks finished.
Scenario analysis is where this bites hardest. The scenario sets everyone cites, from NGFS to IPCC pathways to the IEA, are revised periodically. A model can cite a superseded version, or produce scenario names and parameters that sound right and do not exist, in the same assured tone. The better the prose, the less likely a busy reviewer is to question it.
"Do not confuse rigour with complexity. A simple methodology that is consistently applied, documented and understood is infinitely more valuable than an overly complex model nobody really understands."
Tom Nisevic, CFO, Metro Trains Sydney
A general assistant has read a great deal about climate risk. It has never seen your asset register, your risk appetite statement or the risks you already track. It also does not have the licensed, current datasets that serious climate work depends on, from up-to-date emission factors to location-specific physical risk data.
Without that context, AI produces the climate risk list for your sector, not for your company. It cannot tell you that a flood risk matters more to you than a carbon price because three of your largest sites sit in the same catchment. The same problem runs through the climate-related metrics, where deciding which assets are vulnerable to transition or physical risk is entity judgement with no external methodology to check against.
Get this right and the speed gains are real. With AI tooling working on its own data, SEE Group cut its climate risk assessment from 13 weeks to 2.
This is the hardest gap to close, because the information does not exist anywhere AI could find it. What counts as acceptable proof of director training, how much documentation an excluded risk needs, how emissions working papers should be laid out: these vary between assurance firms and often between engagement teams within the same firm.
Auditors arguably have more influence over how ASRS works in practice than anyone else, and their interpretive differences are often what separates a smooth assurance process from a painful one. No model can tell you what your particular team will push on.
All five gaps fail the same way: a claim that reads well with nothing a person can stand behind. So test it the way your auditor will. Pick five claims from your draft at random, and ask the person who owns each one to walk you from the sentence back to the source document. If they can, that claim is ready, whatever mix of AI and people produced it. If they cannot, you have found a gap while there is still time to fix it.
For the full breakdown of which parts of AASB S2 to own, automate or get help with, including the internal hours each resourcing model really takes, download our guide ASRS: To DIY or not to DIY? And where to use AI?
If you want a second pair of eyes on your draft before your auditor's first one, book a 30-minute call with the Trace team and we will run that sample with you.
There are five: provenance gaps, where a correct number has no recorded path back to source; evidence gaps, where AI describes activity that never happened; the plausibility trap, where confident output is unverified; context gaps, where AI lacks your company data and licensed datasets; and audit expectations gaps, where it cannot know what your specific auditor wants.
Not on their own. Auditors sample individual line items and trace each back to a source document such as an invoice or fuel card statement. A figure produced in a chat session has no persistent trail for them to follow. AI can extract and check data, but the calculation and the link to source need to live somewhere an auditor can open and recalculate.
It can help research and draft, but it should not supply the scenarios or parameters. Scenario sets such as NGFS, IPCC pathways and the IEA are revised periodically, and general AI tools cannot reliably tell you whether they are citing the current version. They can also produce plausible parameters that do not exist. Verify every source independently and keep the judgement with your team.
Ask them directly, and early. Book a walkthrough before you start drafting and bring specific questions about governance evidence, how judgements should be documented and how emissions workings should be presented. Expectations vary between firms and engagement teams, and none of this is published, so the conversation is the only reliable source.
Generally, no. Governance is assured in full from year one, and evidence that was not created at the time usually cannot be recreated afterwards. If climate oversight is not yet on your board agenda and minuted, the time to fix that is before the period closes, not when the auditor asks.
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)