Skip to content
Week 47News, and what it means for you

AI agents

KPMG let AI agents audit a listed company, and they sometimes asked better questions

In a pilot at a UK-listed company, more than a hundred AI agents worked alongside the regular audit team. In places they dug deeper than the auditors themselves.

KPMG in the UK had more than a hundred AI agents audit the accounts of a UK-listed company, alongside a regular team of auditors. In several places, the firm says, the agents went considerably deeper than the team.

The pilot was called Project Exponential and ran earlier this year at an unnamed company in the FTSE 250, the index of mid-sized firms on the London Stock Exchange. Two partners and a director developed the agents in about four months, together with managers and junior staff. An AI agent is an AI program that carries out a task on its own. Here the agents gathered data, analysed and documented it, drafted conclusions and suggested next steps. Agents working on different parts of the audit agreed among themselves who would do what. The Financial Times broke the story.

Cath Burnet, KPMG UK’s head of audit, told the paper that in several places the agents went considerably deeper into issues than the regular team. They flagged areas that needed further work and asked questions the auditors had not. In one area an agent failed to match a set of regulations to the right company documents, and an auditor stepped in. An audit partner signed off the pilot, but KPMG did not rely on it when it formally signed off the accounts.

KPMG plans to extend the approach to up to twenty larger audits and hopes eventually to offer AI-based audits to clients. According to Burnet, the technology is not yet ready for thousands of engagements, and few companies keep data clean enough for agents to work with. KPMG has not published figures on time saved or accuracy.

The AI sometimes asked better questions than the human team.

Cath Burnet, head of audit at KPMG UK, to the Financial Times, summarised

How an audit works

An auditor checks whether a company’s accounts give a true and fair view. Checking every entry of a large company is impossible, so auditors sample: some items are picked by risk, such as large or unusual ones, the rest at random, so that the auditor can draw conclusions about the whole. The UK regulator, the FRC, wants firms to record those choices better.

Some tests are always required. An auditor must, for instance, test journal entries, the manual postings through which figures can be adjusted after the fact. Software has picked out the striking entries for years: those around the year end, outside office hours or posted by top management. According to the ICAEW, the accountants’ professional body, it is not enough that the software flagged something: the auditor must understand how that choice was made and record it.

What the firms already do with AI

In 2024 the Big Four carried out nine in ten audits of the UK’s 350 largest listed companies, and they have been investing heavily in AI for years. In 2025 KPMG built AI agents into Clara, the audit platform for its more than 95,000 auditors worldwide, for tasks such as vouching expenses and searching for liabilities not yet on the books. PwC, according to its global assurance leader Andy Hammond, has run several parallel audits in which AI works alongside without the firm relying on it.

What the regulator asks

The Financial Reporting Council issued its first guidance on AI in audit in June 2025 and, in March 2026, became the first audit regulator anywhere to publish guidance on generative and agentic AI. The core message is always the same: the auditor stays accountable, whatever tool they use.

On 16 November, two days before the KPMG news, came guidance on independence. An auditor may not advise a client they audit, and insights from AI can blur that line. The FRC’s Mark Babington told the Financial Times that auditors must be able to explain how they used the technology, challenge its outputs and stand behind the evidence.

Our analysis

Agents dig, people decide, and that is how it should be

The most interesting thing about this pilot is not that the agents found things, but how it was set up. People gave the agents their objectives and information about the company and earlier audits, checked their work along the way and challenged their conclusions. What the agents did in documents was logged, KPMG says, down to the character and the second. When an agent matched regulations wrongly, an auditor saw it and stepped in. That is not a failure but a design that works: the agents widen the search, the people keep the judgement.

What the coverage glosses over: ‘deeper’ is, for now, KPMG’s own assessment, with no figures on errors found, time saved or cost. And the agents worked alongside the team, not instead of it. That is exactly the right order: side by side first, then compare, and only then trust.

Our expectation: this pattern fits far more checking work than the audit, such as reviewing contracts, checking invoices or testing whether a process follows the rules. The bottleneck KPMG names, data that is not in good order, applies to most organisations. Whoever keeps their records well structured will get the most out of it, with their own auditor too.

What stands out

  • The gain was in the questions According to KPMG, the agents flagged areas that deserved more work and asked questions the team had not.
  • The work can be traced According to KPMG, the system recorded what the agents did in documents, with timestamps and down to the character. So you can see where a conclusion came from.
  • The agents divided the work themselves Agents on different parts of the audit agreed among themselves who would pick up what, much as a team of people would.
  • What it will cost is still open KPMG sees clear time savings but will not say whether audits will get cheaper. With its own auditor, KPMG International did negotiate lower fees, citing AI.

What it means for you

AI agents can make checking work broader and deeper, as long as people keep the judgement and every step can be traced.

  • Let agents work alongside your people first Start in parallel with the existing work and compare the results before you rely on anything.
  • Log what every agent does Only then can a colleague check a conclusion and explain it to an auditor or regulator.
  • Get your data in order Agents work best with clean, structured data. That also helps straight away when your own auditor comes in.

Where we come in · Solutions

We develop agents that search with you, while a colleague decides

Flaire develops tailored AI agents for the checking work in your own administration: pulling invoice details from PDFs and checking them, labelling emails or checking orders before they reach your ERP. We set them up so that every step can be traced and a colleague has the final say.

Which checking work would you like to have looked into more deeply?

Tell us where you now take samples or check by hand. We’ll think it through with you and always reply within one working day.

More from week 47

All news

Search our services, the news and the rest of the site.