OpinionThe Global Eye 24 view

Audit firms should use AI to widen the questions asked, not to shrink the people answering

KPMG's pilot suggests AI agents can interrogate accounts better than people, but the human check is still needed, and regulators should insist it stays.

Watch in 47 seconds

Key points

  • AI agents asked sharper questions than accountants in the pilot, but people still had to intervene
  • A bad question costs minutes to discard, while a bad audit conclusion can misstate accounts, so AI belongs upstream of judgement
  • Regulators should require disclosure of AI use, published error rates and named human sign-off

The KPMG pilot deserves to be taken seriously, but not as a prelude to replacing auditors. We read it as evidence that AI is most useful in audit when it widens the questions put to management, while people remain accountable for the answers and the conclusions. The Big Four firm used AI agents to review the financial statements of a FTSE 250 company, and the agents asked better questions than the accountants did. Human intervention was still required. Both halves of that result matter, and policy should reflect both.

Audit exists to test what a company says about itself. A reviewer that reads every note, cross-checks every figure and never tires has an obvious advantage over a stretched team working to a deadline. Critics of the profession also argue that junior staff have reasons to avoid awkward queries with a paying client. Software has no career to protect, and that is an argument for using it to generate lines of inquiry, not a reason to trust it with the verdict.

The limits are real, and the pilot's details are unclear. We do not know how many of the agents' questions led anywhere, how many were wrong, or what the human intervention involved. The account also notes that few companies currently hold data clean enough for such agents to work with. One pilot at one company is a promising result, not a method.

The strongest objection is about incentives. If a machine can draft the questions, firms will be tempted to cut the teams who used to do the work, keep a handful of signatories and sell cheaper audits of lower quality. Even a small error rate in AI analysis means everything must be checked, and a model that asserts something confidently and wrongly is more dangerous than one that is visibly unsure. There is also a pipeline problem: junior auditors learn the trade by doing the work now handed to software.

These risks are serious, but they argue for how the tool is deployed, not for leaving it unused. Used to generate questions, AI makes its own mistakes cheap: a poor question costs a few minutes to discard, whereas a poor conclusion can leave misstated accounts in front of investors. The control failure reported at Citigroup, where trades at a UK securities unit were recorded using an internal model regulators had not approved and risk-weighted assets were understated by about $35bn, is the kind of gap a systematic questioner could be pointed at. We do not know how that error was found, and we are not claiming AI would have caught it. The point is that tireless, unembarrassed questioning is worth having, provided someone qualified decides what the answers mean.

Three things should follow. First, the audit regulator should require firms to tell audit committees when AI agents have been used in an audit, and what humans did with their output. Second, firms should publish the results of their pilots, including how often the agents were wrong, so that claims of better questions can be tested rather than assumed. Third, a named, qualified human must keep full responsibility for the opinion, with the firm's liability unchanged by the tooling it chooses.

Companies have their own part to play. Audit committees should ask their auditors how AI is being used and press for any savings to be reflected in fees, and finance teams should start cleaning the data these tools need. Firms, for their part, should protect training places for junior auditors, because the people who will one day sign opinions have to learn what a good question looks like from somewhere.

This editorial is the view of Global Eye 24, written with AI assistance. Spot an error? Write to corrections@globaleye24.com.