Enterprise & Work

KPMG Ran 100 AI Agents Through a Full FTSE 250 Audit — and the Bots Dug Deeper

KPMG ran 100+ AI agents through every phase of a FTSE 250 audit. The bots dug deeper than humans in several areas — and regulators have new rules ready.

KPMG's "Project Exponential" put AI agents alongside a real audit team and found the bots dug deeper on individual quest
KPMG's "Project Exponential" put AI agents alongside a real audit team and found the bots dug deeper on individual questMohamed_hassan / pixabay
By Rebecca Stone6 min read

Why it matters

  • KPMG deployed more than 100 AI agents across every phase of a parallel audit of a FTSE 250 company in 2025.
  • UK audit head Cath Burnet said AI was able to go so much deeper than the regular audit team in several cases.
  • KPMG plans to expand Project Exponential to as many as 20 larger audits.
  • The FRC published new guidance applying auditor-independence rules to AI use days before the experiment became public.
  • Two KPMG entities were fined £21 million after the 2018 Carillion collapse.

KPMG deployed more than 100 AI agents across every phase of an annual audit of a FTSE 250 company for the first time — and the firm's UK head of audit says that in several cases the agents went deeper than the human audit team did.

The experiment, which KPMG runs internally under the name "Project Exponential," placed the agents alongside a conventional audit of a UK-listed company conducted this year. The agents did not sign anything. The official audit opinion rested entirely on the regular audit performed by the human team. The Big Four firm described the setup as a full parallel audit, the first of its kind for KPMG, according to a Financial Times report.

The results are the strongest signal yet that AI agents can perform substantive audit work at scale — and the clearest illustration of why regulators and the industry's own fee structure are not ready for it.

What did the AI agents actually do?

The agents worked independently through the core mechanics of an audit. According to KPMG, they collected data, analyzed it, documented their findings, and drafted audit conclusions — the labor-intensive work normally distributed across teams of junior auditors and senior managers.

In one area of the audit, a human had to intervene because an AI agent could not correctly match a set of regulations against the company's documents. The failure is instructive: mapping abstract regulatory requirements to specific corporate records is exactly the kind of judgment-laden task auditors spend years training to do.

In several cases, however, the technology outperformed the humans on depth. "AI was able to go so much deeper" than the regular audit team, said Cath Burnet, KPMG's UK head of audit. In some situations, she said, the agents asked "better questions" than the firm's own auditors.

The distinction matters for the economics of auditing. Traditional auditors pull samples — they examine a fraction of transactions and extrapolate. Agents that can read everything change that calculus. Where a human team samples, an AI pipeline can in principle analyze far more transactions, a point the UK's audit regulator itself flagged in new guidance this month.

How much human oversight was required?

KPMG is explicit that humans remained "critical" throughout the experiment. Staff set goals for the agents, provided context, reviewed their work periodically, and challenged conclusions before the audit opinion was finalized. The human side of the parallel audit included two partners and a director working alongside senior managers and junior staff.

The firm also claims the system provides "full visibility and full traceability." It records every agent interaction with documents down to individual characters and timestamps — an audit trail for the auditors' machines. That capability addresses one of the central objections regulators have raised to AI in assurance work: if a machine reaches a conclusion, a human must still be able to show the evidence trail behind it.

KPMG plans to expand Project Exponential to as many as 20 larger audits. But Burnet cautioned that the technology is not yet available at the scale the firm would need to run it across thousands of audits. One bottleneck is data quality: few companies currently have clean enough data for AI agents to work with reliably, she said.

What does the FRC's new guidance say?

Days before the experiment became public, the UK's Financial Reporting Council published guidance applying existing auditor-independence rules to the use of AI. The core question the FRC addressed is when auditors may share findings from AI-powered audits with their clients without compromising their independence.

The problem is structural. Strict limits or outright bans already apply to certain non-audit services for audit clients. A firm that advises a company may later end up reviewing its own recommendations, with a built-in incentive not to upset the client.

AI agents intensify that conflict, the FRC argues, precisely because of the depth Project Exponential demonstrated. Because agents dig deep into a company's data during an audit and can analyze far more transactions than sampled methods, they surface findings that go beyond the audit itself: inefficient processes, potential cost savings, and patterns that might never have come up in a conventional audit.

The FRC sees a risk that such findings could come across as advertising for extra services or serve as a sales pitch for consulting work. Under certain circumstances, the regulator says, sharing those findings could pose a "distinct threat" to auditor independence.

The FRC drew one clear boundary. "Whatever technology a firm uses, the auditor remains accountable," said Mark Babington, the FRC's executive director of regulatory standards. Auditors must be able to explain how they used AI, challenge its results, and stand behind the evidence, he said. Certain findings — such as major weaknesses in internal controls — still must be reported to clients, and the FRC says these mandatory disclosures do not count as non-audit services.

Why does this matter for KPMG specifically?

Auditor independence is a particularly sensitive issue in the UK, and KPMG sits at the center of that sensitivity. After the construction and outsourcing group Carillion collapsed in 2018, a parliamentary inquiry found that KPMG, as the company's longtime auditor, had failed to challenge management's "increasingly fantastical figures" for 19 years. Two KPMG entities were fined a total of £21 million as a result.

The Carillion failure drove reforms to audit regulation in the UK and made the question of audit quality — whether auditors challenge management hard enough — a matter of parliamentary record. An experiment in which machines ask "better questions" than auditors reads differently against that backdrop than it would for a firm without that history.

What happens to audit fees and the Big Four business model?

The experiment raises a more fundamental commercial question. If AI agents can take over the labor-intensive work of large teams of junior auditors, what happens to the industry's fee structure? KPMG sees "clear time-saving opportunities," according to the FT, but will not say whether those savings would lead to lower fees for clients.

The pressure is already visible elsewhere in the firm. KPMG International has negotiated lower fees with its own auditor, arguing that AI makes the work cheaper, the FT reported. Rival PwC has already run several "parallel audit" projects of its own.

At Deloitte, the same pressure surfaced inside the firm. At an internal town hall in May, according to the Wall Street Journal, management presented a chart showing the share of hourly consulting revenue shrinking to a thin sliver of the growing overall market by 2035. One Deloitte consultant summed up the message: "They heavily implied our model is toast. We're basically getting replaced by robots."

KPMG's next step — scaling Project Exponential to as many as 20 larger audits — will test whether parallel audits can move from experiment to routine practice while regulators, and the firms' own pricing committees, decide who captures the savings.

Original: ft.com

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

728 articles

Related articles

  1. KPMG Puts AI Agents Inside U.S. Audit Workflows With Clara
  2. AI at Gilbert + Tobin: A Law Firm CEO's $62 Annual Review
  3. Over 40% of Agentic AI Projects Face Cancellation by 2027
  4. OpenAI audit exposes major gaps in AI agent monitoring systems
  5. Partnership on AI: Agent Frameworks Can't Show What Agents Did

« Previous article