KPMG Put 100 Agents on a Real Audit. They Asked Better Questions
Audit is the textbook case of work that is huge, repetitive and regulated, which makes it the perfect stress test for multi-agent systems. KPMG UK just ran that test. According to the Financial Times on Nov 18 (readable relay: https://europeanbusinessmagazine.com/ai-agents-better-questions-kpmg/), the firm's Project Exponential put more than 100 AI agents on the financial statements of an unnamed FTSE 250 company, running in parallel with a conventional human audit team.
The agents did not just fetch documents. They collected information, analysed and documented data, drafted conclusions and observations, suggested next steps, and split the work among themselves. The system took about four months to build. KPMG UK head of audit Cath Burnet said that in several instances the agents went considerably deeper into issues than the regular team and asked questions the humans had not raised.
Now the fine print, which is the more useful part. This was not a statutory audit: an audit partner signed off on the experiment, but its findings were not used to approve the accounts. A human auditor had to step in when one agent failed to match regulatory requirements to the company's documentation, exactly the kind of mapping error that looks fine in a summary. KPMG plans to extend the approach to as many as 20 larger audits and says the technology is not mature enough for thousands of engagements. The bottleneck it names is not the model but data quality: plenty of companies don't have clean, structured information for agents to work on.
Read it next to this week's papers on agents that say done when they aren't, and the KPMG result looks right-sized. Agents asking deeper questions is a real gain, because breadth is cheap for them and expensive for juniors. But the one stated failure was a quiet mismatch caught by a human, and the partner's signature is still what makes it an audit. The Big Four's pyramid of junior staff is what's under pressure here, not the partner.
← Back to all articles
The agents did not just fetch documents. They collected information, analysed and documented data, drafted conclusions and observations, suggested next steps, and split the work among themselves. The system took about four months to build. KPMG UK head of audit Cath Burnet said that in several instances the agents went considerably deeper into issues than the regular team and asked questions the humans had not raised.
Now the fine print, which is the more useful part. This was not a statutory audit: an audit partner signed off on the experiment, but its findings were not used to approve the accounts. A human auditor had to step in when one agent failed to match regulatory requirements to the company's documentation, exactly the kind of mapping error that looks fine in a summary. KPMG plans to extend the approach to as many as 20 larger audits and says the technology is not mature enough for thousands of engagements. The bottleneck it names is not the model but data quality: plenty of companies don't have clean, structured information for agents to work on.
Read it next to this week's papers on agents that say done when they aren't, and the KPMG result looks right-sized. Agents asking deeper questions is a real gain, because breadth is cheap for them and expensive for juniors. But the one stated failure was a quiet mismatch caught by a human, and the partner's signature is still what makes it an audit. The Big Four's pyramid of junior staff is what's under pressure here, not the partner.
Comments