An auditor reviews printed financial records beside an AI analysis screen

KPMG Says Its AI Auditors Are Nearly as Good as Humans — But Humans Still Sign

Key Takeaways

  • KPMG has piloted its first full artificial intelligence-based audit, running AI agents through every phase of an audit of an unnamed FTSE 250 company, in parallel with its conventional human audit.
  • The firm says the AI-based audit was "in most cases" as good as the work produced by its own accountants, and potentially better in some areas. Its UK head of audit, Cath Burnet, said the AI went deeper into some issues and asked "better questions" than the human team.
  • The experiment, codenamed Project Exponential, used more than 100 AI agents, according to reporting based on the Financial Times. The Times, which also covered the story, described the number as "dozens". The two accounts differ, and the exact figure is unconfirmed.
  • The AI audit was not used to sign the company's accounts. The official audit opinion came from the conventional audit, although a KPMG partner signed off on the experimental exercise itself.
  • Human judgement remained essential: in one area, an auditor had to intervene when an AI agent failed to match regulations to the company's documents.
  • KPMG plans to expand the project to as many as 20 larger audits, and eventually to offer AI-based audits to clients. But the firm says the technology is not yet ready to run at the scale of thousands of audits, and that few companies have clean enough data for it to work reliably.
  • The UK's audit regulator, the Financial Reporting Council, published guidance on 16 November on maintaining auditor independence when using technology including AI.

What KPMG Actually Did

The headline is easy to misread, so it is worth stating precisely what happened before drawing any conclusions from it.

Earlier this year, KPMG UK built and deployed AI agents to examine the financial statements of an unnamed FTSE 250 company — a company listed on the London Stock Exchange's FTSE 250 index of mid-sized listed businesses. The agents worked through every phase of the annual audit, alongside KPMG's conventional audit team, which was conducting the same audit in the ordinary way. The two processes ran in parallel, on the same company, at the same time, so that the results could be compared.

This is what makes the experiment significant. KPMG has used AI in auditing before, as have the other Big Four firms — typically for discrete tasks such as reading contracts, sampling transactions or flagging anomalies. What is new here is the scope: an end-to-end audit, from gathering data through to drafting conclusions, carried out by autonomous agents rather than by people using AI tools for individual jobs. The Financial Times, which first reported the story, described it as the first time KPMG had set an automated system loose on an entire audit of a UK-listed company. The Times' own headline put the claim more plainly: "KPMG says its AI auditors are nearly as good as humans."

The project was codenamed Project Exponential. According to reporting based on the Financial Times, it deployed more than 100 AI agents. The Times, covering the same experiment, said KPMG had built and deployed "dozens" of AI agents. The discrepancy is not trivial — "dozens" and "more than 100" are different claims — and it has not been reconciled in the reporting we reviewed. We use the Financial Times-based figure of more than 100 in this article, and flag the difference, because readers deserve to know that the number is not settled.

The system was built over about four months by a team of two partners, a director, senior managers and junior staff. The agents independently gathered, analysed and documented data across every audit phase, then drafted their own conclusions and observations. The audit partner signed off on the experimental exercise — but, crucially, its findings were not relied on for the statutory audit opinion. The official opinion on the company's accounts came from the human-led audit. That boundary is the most important fact in the story, and it is the one most likely to be lost in the headline.

What the AI Did Well — and Where a Human Had to Step In

KPMG's assessment of the experiment is positive, but it is specific, and the specifics matter.

Cath Burnet, KPMG UK's head of audit, told the Times that "in most cases, [the AI-based audit] was aligning to what we were doing as part of the normal audit" — that is, the agents' work agreed with the human team's work most of the time — and that it was potentially better in some areas. The AI agents were much quicker at digesting large datasets, she said, and were therefore able to provide "deeper insights" than the human teams.

Speaking to the Financial Times, Burnet went further. "In several cases the AI was able to go so much deeper" than the regular audit team, she said, and in some situations the technology asked "better questions" — questions the human auditors had not necessarily raised, which identified areas requiring further investigation. That is a striking claim, and it is the one that has travelled furthest in the coverage. It is worth pausing on what it means: not that the AI was right more often, but that it probed individual issues more deeply and, in places, asked questions the humans had not thought to ask. In an audit, where the value often lies in the question that uncovers the problem, that is a meaningful capability.

But the experiment also found a clear limit, and KPMG has not hidden it. In one area, a human auditor had to intervene because an AI agent could not correctly match a set of regulations against the company's documents. The agent's failure was not a small error; it was the kind of judgement task — reading a regulation, understanding what it requires, and applying it to a specific set of company records — that sits at the centre of what auditors do. The intervention is reported in both the Times and the Financial Times-based accounts, and it is the single most informative detail in the story, because it shows where the boundary between machine capability and human judgement currently falls.

Burnet's own summary of the experiment reflects that balance. "What the project really helped to show us was that the human element remains really important," she told the Times. And on the question that will dominate the commentary — whether AI will eventually replace auditors — she was direct: "I don't think AI will ever run an audit. My belief is that you need human judgment. I call it 'AI-enabled, human assured', because without the human element we wouldn't be able to [reach a conclusion]."

The firm also built traceability into the system: according to the Daily Pioneer, which reported on the Financial Times story, the system logs every character-level interaction with timestamps, so that each step the agents take can be traced. That matters for audit, where the ability to show how a conclusion was reached is almost as important as the conclusion itself. An audit opinion is, in the end, a statement that a professional is willing to stand behind; a log of what an AI agent did is a step towards making that possible, but it is not the same thing as a human taking responsibility.

Why This Matters for Auditors

For the audit profession, Project Exponential lands in the middle of a long-running argument about what AI will do to professional work — and it gives both sides of that argument something to point to.

The case for optimism is straightforward. Auditing is one of the most labour-intensive areas of professional services: it involves reading enormous volumes of documents, checking transactions, testing controls and documenting everything. AI agents that can digest large datasets quickly, go deeper into individual questions and ask questions humans missed are, on the face of it, a genuine improvement in audit quality. Burnet herself said the agents "absolutely will help us with audit quality", according to the Times. If AI can do the heavy reading and the deep probing, human auditors can spend their time on the judgement calls — the areas where, as this very experiment showed, machines still need help.

The case for caution is equally visible in the same experiment. The agent that failed to match regulations to documents failed at a task that requires understanding, not just processing. Audit is full of such tasks: deciding whether a judgement made by management is reasonable, assessing whether a risk has been properly disclosed, weighing evidence that points in different directions. These are not problems of volume, and more processing power does not solve them. The survey evidence in this area is instructive. The Centre for Public Interest Audit, a policy and research group for large audit firms, published its 2026 Audit Trust Index in October. Censuswide surveyed 257 senior finance professionals between 3 and 25 August, including 57 audit committee chairs. The survey found that for reviewing large volumes of data or contracts, 67% had more confidence in AI than in auditors, against 14% for auditors — but that the ranking reversed for judgement tasks. In other words, the people who buy audits already draw the same line the KPMG experiment drew: machines for volume, humans for judgement.

There is also a question about what happens to the profession's pipeline. Audit has traditionally been a training ground: junior staff do the detailed checking, learn the business, and gradually take on judgement work. If agents take over the detailed checking, the path by which juniors learn the judgement work becomes less clear. KPMG's team for Project Exponential included junior staff alongside partners and senior managers, which suggests the firm is thinking about that pipeline, but the industry has not answered the question yet. It is worth noting, too, that the experiment was built by a small team over four months — a reminder that the capability, for now, is a crafted project rather than a product that can be switched on.

What the experiment does not do is settle the question of jobs. Burnet's statement that AI will never run an audit is a statement about the nature of the work, not about headcount, and KPMG has not said what the project means for staffing. Readers should be careful not to read a reassurance about the survival of the audit opinion as a reassurance about the survival of every audit job in its current form. The history of automation in professional services suggests the work changes shape before it disappears — and the shape here is already visible: fewer hours on reading and checking, more on judgement, challenge and sign-off.

Why This Matters for Clients

For the companies being audited, the immediate question is the most practical one: will this make audits cheaper, faster or better? KPMG's answers, so far, are partial.

On quality, the firm's claim is that AI agents can provide deeper insights and ask better questions — and the experiment, as reported, supports that claim at least in part. A deeper audit that asks questions a human team might not have raised is a better audit, and clients should welcome it. On speed, the agents were "much quicker at digesting large datasets", which should, in principle, shorten parts of the audit timetable. But KPMG has not published timings, and no client has yet been through a full AI-based audit as the real thing, so the practical benefit is still unproven.

On fees, KPMG declined to say whether AI would reduce what clients pay, according to reporting on the Financial Times story. That is a notable silence. If AI makes audits faster and cheaper to deliver, clients will reasonably ask why the price has not fallen; if the answer is that the savings are being reinvested in deeper analysis and better quality, that is a defensible position — but it is one the firm will have to make explicitly. For now, the honest answer is that nobody knows what an AI audit will cost, because nobody has yet bought one as a service.

There is also a question of readiness. Burnet cautioned that few companies currently have data clean enough for AI agents to work with reliably, and that the technology is not yet available at the scale needed for thousands of audits. That is a significant caveat. An AI audit depends on structured, accessible, consistent data; many companies' financial data is none of those things, spread across systems that do not talk to each other, in formats that require human interpretation before any analysis can begin. Some companies, Burnet also noted, may prefer not to use the service at all. The result is that AI auditing, for the foreseeable future, is likely to be a service for a subset of well-organised clients, not a universal replacement for the conventional audit.

The regulatory picture is moving at the same time. The UK's Financial Reporting Council published guidance on 16 November on maintaining auditor independence when using technology, including AI — just days before the KPMG story was reported. The FRC warned that AI could threaten auditor independence if insights generated by the technology are used to advise audit clients, because an auditor that advises a client on the basis of its own AI's findings risks compromising the independence that the audit opinion depends on. That guidance frames the regulatory context in which KPMG's expansion will happen: the technology will be judged not only on whether it works, but on whether it can be used without undermining the independence of the audit itself.

The trust picture is mixed. The Audit Trust Index survey found that 79% of respondents were happy with how firms deploy AI, down from 87% in 2025 — a decline, but from a high base. At the same time, 23% of audit committee chairs reported limited confidence in how AI is used in their company's audit, and over-reliance weakening professional scepticism is now seen as the top risk, cited by 46%, while concern about hallucinations fell to 34% from 45%. The pattern is consistent: confidence in the technology is broadly holding, but the specific worry has shifted from AI making things up to humans trusting it too much. That is precisely the risk the KPMG experiment illustrates — an agent that fails on a judgement task is only dangerous if nobody notices — and it is the risk the FRC's guidance is aimed at.

What Happens Next

KPMG plans to expand Project Exponential to as many as 20 larger audits, and eventually to offer AI-based audits to clients. That is a significant commitment, and it means the experiment will now be tested on bigger, more complex companies, where the data is messier, the judgements are harder, and the stakes of a missed issue are higher. How the agents perform there will tell us much more than a single FTSE 250 pilot can.

Three things are worth watching as the expansion proceeds.

The first is whether the "better questions" finding holds up. In a single pilot, a handful of deeper probes is encouraging; across 20 larger audits, the question becomes whether the agents consistently add insight, or whether the pilot's successes were partly a function of a well-chosen test case and a carefully built team. The second is the failure mode: the regulation-matching failure in the pilot is the kind of error that, in a real audit, could matter a great deal, and the question is whether the next round of pilots finds more such failures, or whether the system has been improved to handle them. The third is the regulatory response: the FRC's guidance is recent, and how it is applied to real AI audits — what counts as a threat to independence, and what safeguards are expected — will shape what KPMG and its competitors are actually allowed to do.

It is also worth watching what the other Big Four firms do. KPMG is the first to report a full parallel audit of this kind, but the other large firms have their own AI programmes, and the competitive pressure to match or exceed this claim will be considerable. The risk is a race in which each firm announces a more automated audit than the last, and the discipline of the human sign-off is quietly eroded. The safeguard against that is the one Burnet herself named: "AI-enabled, human assured" — a phrase that is only meaningful if the human assurance is real, and if the humans doing it have the time, the scepticism and the authority to overrule the machine.

For readers trying to place this story in the wider picture of AI at work, our explainers on what AI agents are and how they work and on AI in accounting in 2026 provide useful background: the first on the agent technology behind Project Exponential, the second on how AI is already changing the accounting work that surrounds the audit itself.

What We Could Not Verify

  • The number of AI agents deployed is reported differently by different outlets. Reporting based on the Financial Times says more than 100 agents were used; The Times, covering the same experiment, says KPMG built and deployed "dozens" of AI agents. We could not reconcile the two figures, and we have used the Financial Times-based figure while flagging the difference.
  • We could not read the Financial Times' original report directly, and the Times article is paywalled. Our account of the experiment relies on the Times' own summary as returned in search results, and on secondary reporting of the Financial Times story by the Daily Pioneer, The Bharat, European Business Magazine, Streamline Feed and others. Direct quotations from Cath Burnet are as carried in those accounts.
  • The date of the Financial Times report is given as 17 November in some secondary accounts and 18 November in others. The discrepancy is minor and does not affect the substance, but we could not resolve it.
  • The identity of the FTSE 250 company audited has not been disclosed, and KPMG has not named it. Any claim about which company it was would be speculation, and we have not repeated any.
  • The claim that the AI audit was "in most cases" as good as the human audit, and "potentially better in some areas", is KPMG's own assessment of its own experiment, as reported by the Times. It is not an independent evaluation, and no independent comparison of the two audits has been published. We have kept the claim attributed to KPMG throughout.
  • The Audit Trust Index figures (257 senior finance professionals surveyed 3–25 August 2026, including 57 audit committee chairs) come from secondary reporting of the Centre for Public Interest Audit's survey. We did not read the survey itself, and the percentages are as reported.
  • KPMG has not disclosed the cost of the project, the time saved, or whether client fees will change. Any figure on those points would be invented, and we have not given one.
  • Whether the AI agents' "better questions" led to any finding that changed the audit outcome has not been reported, and we do not know whether any such finding existed. The pilot's findings were not relied on for the statutory opinion, so the question is, for now, unanswerable.

Similar Posts