In June, KPMG pulled a flagship report on the promise of agentic AI after it emerged that AI had fabricated much of it. According to reporting in the Financial Times and elsewhere, an analysis of the report found that of its 45 citations, only five pointed to real, uncorrupted sources. Forty of the forty-five were fake. Named organizations, including UBS and the UK’s National Health Service, told the FT that the report’s claims about their own AI use were untrue or misleading.

Sit with who this happened to. Not a startup, not an intern. One of the four firms whose entire business is verification, publishing a report on how well AI works that AI itself had corrupted. If the house that sells assurance can be embarrassed by its own chatbot, the comfortable assumption that a serious firm will always catch the machine’s mistakes just took a very public hit.

And KPMG is not even the first of its peers to be caught this way. Late last year, Deloitte’s Australian arm agreed to partially refund the government over a A$440,000 report that cited academic papers which did not exist and a quote fabricated from a federal court ruling. EY pulled a report of its own after analysts found it riddled with invented citations. Three of the four firms that sell the world its assurance have now shipped AI-generated work that the AI made up. That is not a run of bad luck.

This is the risk everyone else is walking into, usually less visibly. In the legal world, where the problem has been tracked most carefully, aggregated trackers have logged hundreds of cases of AI-fabricated citations across more than two dozen jurisdictions. The trouble in accounting is that a made-up number buried in a set of books does not raise its hand the way a fake case citation does in front of a judge. It sits there quietly until someone with judgment finds it, or until an auditor does.

Which points at the sharper issue, and it is not really accuracy. It is defensibility. Writing in Accounting Today, Graeme Chard, an executive at an accounts-payable software vendor and so not a neutral party, made the cleanest version of the argument: most of the AI shipping into finance in 2026 would fail a serious audit, not because it gets the answer wrong, but because it cannot show its work. Ask it to walk you through the payment it approved on the 14th, which model ran, what data informed it, which policy version applied, who signed off, and it hands you a confident sentence rather than a record. “The AI approved it” is not an audit trail. If you cannot reconstruct a single AI decision on demand, neither can the person auditing you.

And the rules are behind the tools. Writing in Forbes, Yvonne Hinson argues that AI is reshaping audit risk faster than the standards can keep up, and that the profession has not yet settled who is responsible when the AI is wrong. Until it does, the answer falls back to the oldest one in accounting. It lands on the person who signed.

None of this is an argument against using AI, and the KPMG report is not a reason to unplug anything. It is a reason to treat AI the way the profession has always treated a powerful tool, with a control wrapped around it. The boring discipline the vendors leave out of the demo is the whole edge now: a defined verification step before anyone relies on AI output, a reviewer who knows enough to catch the error, a logged trail of how the number was produced, and a named human who owns the result. We made the same case when AI first started touching client data, and it has not aged a day.

The lesson of the KPMG episode is not that AI cannot be trusted. It is that trust in accounting was never supposed to be automatic. AI can do the work. It cannot sign the work. Keep a human who can, and who actually did.

— Footnote

Footnote is an independent publication, with no affiliate links and no vendor paying for placement. It is not professional accounting, tax, or legal advice. Details of the KPMG, EY, and Deloitte reports are drawn from Financial Times, Fortune, and other established coverage; the accounts-payable audit argument is the stated opinion of a named software-company executive, flagged as such; and the hallucination-tracking figures come from legal-sector datasets, not accounting-specific ones. Details are current as of July 2026.