TL;DR
KPMG has tested AI across an entire audit of a listed British company for the first time, the Financial Times reported on 18 November. This year a fleet of over 100 AI agents examined the financial statements of a FTSE 250 business, which was not named, while KPMG’s accountants carried out the usual audit alongside. Cath Burnet, who leads audit for KPMG in the UK, said that in several cases “the AI was able to go so much deeper” than the human team.
How the pilot worked
The project, codenamed Project Exponential, used agents across every phase of the audit rather than for isolated tasks. Agents collected and analysed data, documented it, drafted conclusions and suggested next steps, working with one another without step-by-step direction. According to the FT, earlier AI use by auditors had been limited to specific tasks, and no large UK audit had used the technology end to end.
The experiment did not replace the real audit. The agents took four months to build, and a partner signed off the AI version, yet the firm set it aside when it formally approved the accounts. Staff set the agents’ goals, briefed them on the business and on previous audits, reviewed their output and challenged their findings. In one area, Burnet said, staff stepped in after an agent could not line up the relevant regulations with the client’s paperwork, and she said human involvement remained “critical”.
KPMG now intends to apply the approach to as many as 20 bigger audits and, in time, to sell AI-based audits to clients. Burnet cautioned that the technology cannot yet run thousands of audits and that few companies have data clean enough for the agents.
The fee question
KPMG said there were “clear time-saving opportunities” but would not say whether fees would fall. PwC has run its own parallel audit projects, its global assurance leader Andy Hammond told the FT.
The timing is pointed. This week the Financial Reporting Council issued guidance warning that AI-derived insights can threaten auditor independence. The FT also cites a Centre for Public Interest Audit study, published in October, in which 23% of audit committee chairs said they had “limited” confidence in AI.
Looking forward
In our view, the economics are the story. Audit has long been sold on the hours of large junior teams, and a credible agent-led audit challenges that model directly. Expect audit committees to start asking for the traceability evidence Burnet describes, which records each agent’s work on documents to the character, with timestamps.