RetroDeckRETRODECK

KPMG's AI Audit Trial: Agents Asked Better Questions, but Grounding Still Needed Humans

In a trial reported by the Financial Times on 18 November 2026, KPMG UK ran more than 100 AI agents on a parallel audit of an unnamed FTSE 250 company. Its audit head said the agents sometimes asked better questions. No metrics were published, and one agent failed to match regulations to documents.

What was reported, and by whom

On 18 November 2026 the Financial Times reported that KPMG UK had run its first AI-based audit of a UK-listed company, an unnamed FTSE 250 business, under the code name Project Exponential. The FT story was syndicated the same day by The Irish Times, and picked up by The Decoder, European Business Magazine and others. No KPMG press release, technical report or research paper accompanied it; the FT interview with Cath Burnet, KPMG UK's head of audit, is the primary source.

The facts as reported: more than 100 AI agents ran alongside a conventional audit team, covering every phase of the audit rather than isolated tasks. A team of two partners and a director, working with senior managers and junior staff, built the agents in about four months. The agents gathered and analysed data, documented their work, drafted conclusions and observations, and proposed next steps. Staff set objectives, supplied business context and prior-year audit information, checked the work periodically, and challenged conclusions before the opinion was finalised. An audit partner signed off the experimental audit, but KPMG did not rely on it when formally signing the company's accounts.

What better questions means, and what it does not

The headline rests on two quotes from Burnet in the FT: that in several cases the AI was able to go so much deeper than the regular team, and that in some situations it asked better questions. Neither is defined further in the coverage. There is no count of questions, no rate of issues found, no error rate, no time or cost figure. KPMG acknowledged clear time-saving opportunities but declined to say whether audit fees would fall.

So the trial did not measure performance in any sense a benchmark would recognise. It was a single parallel run on a single client, assessed by the firm that built the system, with the human team as the comparison. That is a reasonable first step for a regulated profession, and it is also exactly the kind of evidence that should not be generalised. A fair reading is that experienced auditors saw the agents' lines of inquiry and judged some of them sharper than their own.

The failure that matters most

The one specific failure reported is the most informative detail in the story. In one area of the audit, an agent failed to correctly match a set of regulations to the company's documentation, and a human had to step in.

That is a retrieval and grounding task, not a reasoning one. Mapping a requirement to evidence means finding the right clause in a regulatory corpus, finding the right passage in a client's records, and deciding whether one satisfies the other. Similarity search alone often fails here: a regulation and the document that satisfies it can use different vocabulary, and the evidence may be split across several files. When the match is wrong, every conclusion built on it inherits the error, however fluent the write-up.

Burnet's other caveat points the same way. She said few companies currently have data clean enough for AI agents to work with. Agents can only be as grounded as the corpus they retrieve from.

Traceability is the real product

The claim KPMG leaned on most for trust was not accuracy but traceability. Burnet described full visibility and full traceability of the agents' work, with interactions with documents logged down to individual characters and timestamps.

For an auditor, that is the point. An opinion has to be defensible, and a conclusion that cannot be traced to a specific passage in a specific document is not audit evidence. It also explains why the agents could plausibly ask better questions: coverage reporting notes that agents can examine far more transactions than traditional sampling, and a system that reads everything and cites exactly where it looked surfaces anomalies a sampled review can miss.

The lesson travels beyond audit. Any agent working in a regulated domain needs answers that carry their sources, at the span level, and a human step that challenges conclusions before they are acted on.

The regulatory backdrop and what comes next

Two days earlier, on 16 November 2026, the UK Financial Reporting Council published guidance on auditor independence when firms use technology, including AI. It warns that sharing technology-derived insights with clients, especially beyond the audit itself, can threaten independence. The FRC's Mark Babington said auditor independence must remain at the heart of audits that serve the public interest, and the FT quoted him adding that whatever technology a firm uses, the auditor remains accountable.

The FT also cited a Centre for Public Interest Audit finding that 23 percent of audit committee chairs had limited confidence in AI, and reported that PwC has run several parallel audit projects of its own. KPMG plans to expand Project Exponential to up to 20 larger audits, and says the technology is not yet ready for the thousands of audits it performs.

What builders should take from it

The trial supports three practical rules for agents working over private, high-stakes documents. Ground every claim in a retrievable source and keep the citation. Treat requirement-to-evidence matching as its own retrieval problem, with evaluation, rather than trusting a general search. And keep a human challenge step where the cost of a wrong conclusion is high.

RDK is built around the first rule. Private chunk content stays on the user's node, encrypted with the vault key, while the central index holds only titles and embeddings for routing, and agents receive a few hundred tokens of grounded context instead of whole documents.

Frequently asked questions

What was KPMG's Project Exponential?
It was KPMG UK's first AI-based audit of a UK-listed company, reported by the Financial Times on 18 November 2026. More than 100 AI agents ran a parallel audit of an unnamed FTSE 250 company alongside a conventional team. An audit partner signed it off, but the formal opinion relied on the conventional audit.
Did the AI agents outperform the human auditors?
Not in any measured sense. KPMG's head of audit, Cath Burnet, said the agents went deeper in several cases and sometimes asked better questions. No metrics on accuracy, issues found, time or cost were published, and one agent failed to match regulations to company documents, requiring a human to step in.
Is there a KPMG report or paper on the trial?
None was found as of 20 November 2026. The primary source is the Financial Times article and its interview with Cath Burnet, syndicated by The Irish Times. Other outlets, including European Business Magazine and The Decoder, relied on the FT's reporting rather than a KPMG publication.
What does the trial show about AI agent grounding?
Its one reported failure was a grounding task: matching regulatory requirements to the right evidence in company documents. Its strongest trust claim was traceability down to characters and timestamps. Together they suggest agents in regulated work succeed or fail on retrieval quality, clean data and cited evidence.