KPMG ran 100 AI agents through a real audit. One needed a human

A Big Four firm put more than 100 AI agents on the annual audit of a UK-listed company, alongside the human team and on the same client. The audit opinion the company filed did not rely on any of it. That second sentence is the one worth reading twice.
What actually ran
KPMG deployed more than 100 agents across every phase of the audit of an unnamed FTSE 250 company this year, running the experiment in parallel with its conventional audit, the Financial Times reported. The agents gathered data, analysed and documented it, drafted conclusions and observations, and worked autonomously with one another — suggesting next steps rather than waiting to be told.
The build took four months. Two partners and a director worked on it with senior managers and junior staff; KPMG declined to say how large the team was, only that Project Exponential had a "different shape and size" to a normal audit. An audit partner signed off the experimental version before it was set aside.
Cath Burnet, KPMG's UK head of audit, said that in several cases "the AI was able to go so much deeper" than the regular team, and that in some situations the agents asked "better questions" than her own auditors. KPMG gave the Financial Reporting Council a high-level overview of the work. It did not hand the regulator a file.
Where it broke
One area of the audit needed a human because an agent could not match a set of regulations against the company's documents. Burnet offered that detail as a caveat, and it is the most informative sentence in the story.
The failure was not arithmetic and it was not coverage. It was the step where you read a rulebook and a contract and decide which clause governs this arrangement — the judgment layer that audit partners are paid for and that juniors spend three years learning. Agents are strong exactly where the traditional audit is weakest: looking at everything, consistently, at the same standard, at three in the morning. They are weak at the moment the question stops being "what does this document say" and becomes "which standard applies, and does this fact pattern satisfy it".
KPMG's answer to the trust problem is a logging claim: full visibility and full traceability, with agent interactions recorded down to individual characters and timestamps. That is the right shape of answer. It is also a claim nobody outside the firm has tested, because no one outside the firm has seen the logs.
The regulator moved two days earlier
On November 16 the FRC published guidance on auditor independence when technology is used, and it is a narrower document than the headlines suggested. It sets no new requirements; it explains how the existing Ethical Standard applies when an audit firm passes insights from AI work to a client's management, or hands over the tool so management can generate those insights itself, as Financial News summarised the guidance.
The FRC names four threats — management, self-review, self-interest, familiarity — and two of them are new in practice. Self-review: a model auditing information that earlier outputs of a similar model helped shape may lean toward corroboration rather than challenge. Self-interest: an insight that comes from proprietary technology the client does not have can read as a demonstration, a "basis for winning future work".
That is the regulator worrying about something other than bad numbers. A sampling audit surfaces little beyond its sample. An audit that reads every transaction surfaces inefficiencies, cost leaks and control gaps that were never in scope — which is a consulting pipeline wearing an audit's clothes. Mandatory disclosures, such as significant findings or deficiencies in internal control, sit outside the concern, because telling the audit committee what you found is the job. "Whatever technology a firm uses, the auditor remains accountable," Mark Babington, the FRC's executive director of regulatory standards, said in the FRC's announcement.
The UK is the market where this lands hardest. After Carillion collapsed in 2018, a parliamentary inquiry found KPMG, its auditor of 19 years, had failed to challenge management's "increasingly fantastical figures", and two KPMG entities were later fined a combined £21 million. Independence rules written after that episode are not academic.
Who pays for the time saved
KPMG says the project offers "clear time-saving opportunities" and will not say whether clients will see lower fees. Its own behaviour is a hint: KPMG International has already negotiated lower fees from its own auditor on the argument that AI makes audit work cheaper, the FT reported. PwC has run parallel audits of its own.
The structural pressure is on the pyramid. Audit is the last large white-collar business still staffed as a pyramid of juniors doing vouching, reconciliation and testing — the parts agents do best. That is the same arithmetic that hit junior hiring at the labs, where Anthropic's junior tech hiring fell to 9% as agents outworked staff, and it raises the question the profession has not answered: if nobody vouches for four years, who becomes the partner who can spot the regulation mismatch the agent missed?
At Deloitte, a May town hall presented a chart showing hourly consulting revenue shrinking to a sliver of the market by 2035, according to the Wall Street Journal; one consultant's summary was that management "heavily implied our model is toast".
The skeptic's read
A parallel audit that no opinion relies on is a demonstration with the client's name removed. KPMG disclosed no error rate, no count of what the agents found that the humans missed, no false-positive figure, and no cost per audit. "Asked better questions" is a claim about process, not a result. One engagement, one unnamed client, and a caveat that few companies have data clean enough for agents to work with reliably — that is a pilot, not evidence of a changed profession.
There is a deeper point about what an AI audit would even mean. Auditing standards are built around scarcity of attention: sample sizes, materiality thresholds, risk-based rotation of procedures. If coverage becomes cheap, the binding constraint moves to judgment, and the standards have not been rewritten for a world where evidence gathering costs almost nothing and deciding what the evidence means costs everything. KPMG's single regulation-matching failure is a small illustration of a large hole: the part that failed is the part the standard is actually about.
The agents were also not autonomous in the way the framing implies. Staff set their objectives, briefed them on the business and prior audits, reviewed output periodically and challenged conclusions. Four months of build for one audit is a cost that does not appear anywhere in a "time savings" story.
What to watch
The 20 larger audits KPMG plans to run, and whether any of them produces a published coverage or error metric. The first FRC inspection of an AI-assisted file, since the regulator expects the volume of technology-derived insights to rise. Whether an audit opinion is ever signed with agent output in the workpapers under the firm's name — that is where liability stops being a slide. And the fees: if AI audits are sold at a discount, the independence question the FRC just wrote about stops being hypothetical.
If an AI agent does the vouching, should the partner who signs the opinion have to disclose which parts a human actually checked? Write to us via our contact page.







