AI in ESG reporting: assistant, not author

Walk any sustainability software stand this year and you will hear the same promise: artificial intelligence will read your invoices, draft your disclosures, and hand you a finished sustainability statement. After a few cycles of reporting across dozens of countries, I have learned to listen to that promise with one hand on my wallet. AI is genuinely useful in ESG work — I use it — but it earns its place as an assistant, never as the author of record.

The distinction matters more now that the EU AI Act is phasing into force and auditors are starting to ask not just for your numbers, but for the provenance of the tools that produced them. So it is worth being precise about where these systems help and where they quietly create risk.

Where AI genuinely helps

The honest wins are the unglamorous ones. ESG reporting is, in large part, a data-wrangling problem, and that is exactly the terrain where modern models shine:

• Extraction. Pulling consumption figures out of thousands of utility bills and supplier PDFs that arrive in twenty different layouts — tedious for a person, fast for a model.
• First-pass classification. Mapping a messy expense line to the right emissions category, or flagging which ESRS datapoints a given document might touch.
• Drafting and translation. Turning a validated set of figures into a readable narrative, or rendering a methodology note into another working language for a local team.
• Anomaly spotting. Surfacing the site whose energy use halved overnight, so a human knows where to look first.

Notice the pattern: in each case AI compresses effort on tasks that a person can still check. The model proposes; a named human disposes. That is the whole game — it speeds up the work without taking away the one thing that makes a disclosure trustworthy, which is a person willing to put their name to it.

Where it quietly hurts

The danger is not that AI is useless. It is that it is fluent, and fluency reads as confidence. A language model will give you a plausible emission factor with the same calm tone whether it is correct or invented. In a disclosure that an auditor will test and a regulator may sanction, a confident wrong answer is far more dangerous than an obvious gap.

An auditor cannot cross-examine a model. They can only cross-examine you.

Three failure modes recur. The first is the broken audit trail: if a figure was reshaped by a tool nobody can explain, you have a number you cannot defend. The second is silent drift — a model that quietly recategorises data differently this quarter than last, breaking the period-over-period comparability that reporting depends on. The third is plausible fabrication, where a tidy figure is simply wrong, and tidy figures are precisely the ones that escape scrutiny. The common thread is that each failure is invisible until someone goes looking — which is exactly why the looking cannot be optional.

A practical rule

The working rule is simple: AI can touch the inputs, but a human owns the output. Concretely, that means a few non-negotiables.

• Every AI-assisted step is logged — which tool, which version, on which data — so the trail survives staff turnover and the next audit.
• Any figure heading for a disclosure is reconciled against source data by a person who can explain it, not just approve it.
• Models assist with reconciliation and validation; they do not get the final word on a published number.

None of this is anti-technology. The teams that will struggle are the ones that treat the model as an oracle. The teams that will pull ahead are the ones that treat it as a very fast, very tireless junior colleague: enormously helpful, occasionally confidently wrong, and never the one whose name goes on the report.

The point of accountability

ESG disclosure exists to let stakeholders trust what an organisation says about its impact. That trust rests on accountability — on someone being able to stand behind every number. A model cannot stand behind anything; it has no stake, no signature, and no memory of why it chose what it chose. So by all means let AI carry the load it is good at. Just keep a human hand on the pen. In sustainability reporting, the assistant can be artificial, but the author has to be accountable.

Working on something in this space? I'd be glad to compare notes — get in touch.