Back to all posts

KPMG Ran 100 Audit Agents in the Shadows First. Copy That Part.

2026-11-21
luodaInt
KPMG Ran 100 Audit Agents in the Shadows First. Copy That Part.

KPMG Ran 100 Audit Agents in the Shadows First. Copy That Part.

This week KPMG UK said it had run an audit of a UK-listed company with AI agents. The headline was that the agents sometimes asked better questions than the accountants did. That line is what spread, including to the front page of Hacker News.

The more useful detail for a business is how the trial was set up. The agents did not replace the audit. They ran next to it.

What KPMG actually did

According to the Financial Times report, as summarised by the Irish Times, The Decoder and others:

  • KPMG deployed more than 100 AI agents to review the financial statements of an unnamed FTSE 250 company.
  • The agents worked across the phases of the audit. They gathered, analysed and documented data, drafted findings, and suggested next steps.
  • The work ran in parallel with a conventional audit. The official audit opinion was based on the human team's work, not the agents'.
  • In several cases, said Cath Burnet, KPMG's UK head of audit, "the AI was able to go so much deeper" than the regular team, and in some situations asked "better questions".
  • In one area a human had to step in, after an agent failed to match a set of regulations to the company's documents.
  • KPMG plans to extend the approach to up to 20 larger audits, and says the technology is not ready to run across thousands of engagements. Patchy client data is one of the obstacles it named.
  • KPMG acknowledged "clear time-saving opportunities" but declined to say whether audit fees would fall.

That is one firm describing its own pilot to the press. There is no published methodology and no independent evaluation, so treat "better questions" as KPMG's assessment, not a measured result.

The shadow run is the real lesson

Most companies trying agents on a serious process do one of two things. Either they run a demo on clean sample data, or they switch the agents on in production and hope someone notices the mistakes. KPMG did neither.

Parallel, not replacement

The agents worked on a live engagement with real documents, but nothing they produced carried weight on its own. That gives you a direct comparison: same inputs, same deadline, human output next to agent output.

Failures become findings

When an agent could not match regulations to the company's documents, a human caught it and it became a known gap. In production, the same miss would have been a quality problem.

Scale up in steps

One engagement, then up to 20 larger ones, while the firm says it is not ready for thousands. The rollout depends on what the pilot showed, not on the vendor roadmap.

You can run the same pattern on an invoice approval flow, a claims triage queue, a contract review step or a monthly close. Let the agents produce their version of the output. Keep shipping the human version. Compare the two, case by case.

What to compare in your own shadow run

A shadow run only helps if you decide in advance what you are looking for. Based on what KPMG reported, four things are worth tracking:

  1. Where the agent went further. These are the cases KPMG highlighted. Write them down, because they are your business case.
  2. Where it missed or misread a rule. The regulation-matching failure is the classic one: an agent that is confident about documents it did not fully reconcile against the rules.
  3. What it could not do because of the data. KPMG named client data quality as a blocker. Many agent projects that stall are really data projects.
  4. How much human review time each case needed. Time saved on preparation can come back as time spent checking. Measure both.

Accountability does not move

The UK Financial Reporting Council's March 2026 guidance on generative and agentic AI in audit says it plainly: the human auditor is always accountable, and the guidance does not change that. Your sector may not have a rule like this yet. Design as if it does.

The regulator is already writing the next rules

The FRC published that guidance on generative and agentic AI tools in March 2026, describing it as the first from any audit regulator. On 16 November it added guidance on keeping auditors independent when they use technology, including AI. That guidance covers the case where insights produced by the tools are passed on to the audited company's management.

The point for everyone else: once agents start producing useful observations, people will want to use them beyond the original task. Decide early who sees agent output, in what form, and who signs it.

Where this leaves a smaller business

You do not need 100 agents or a Big Four budget to copy the method. A practical version:

  • Pick one process with a clear, checkable output and a person who already owns it.
  • Run the agent beside that person for a fixed period, often a few weeks or one reporting cycle.
  • Log every disagreement and decide who was right.
  • Expand only to the cases where the agent was consistently right and cheap to check.

The trial does not show that agents can audit a company on their own. KPMG says as much. It does show a sensible way to find out what they can do before your customers find out for you.

If you want help setting up a shadow run for one of your own processes, get in touch.

Sources