Skip to content

Knoxville, TN ET Director of AI

Essay // 6 min read

Agents Did the Audit. A Partner Still Signs.

Noel DeLisle Director of AI, GO2 Partners

In February I wrote that questions were becoming the human advantage. Answers were getting cheap. The scarce thing was knowing what to ask.

This week the Financial Times reported that in some situations KPMG's AI agents asked better questions than its accountants.

So I owe that essay a second look.

Here's what happened. KPMG ran an experiment code-named Project Exponential. It deployed more than 100 AI agents to review the financial statements of an unnamed FTSE 250 company and ran it alongside a conventional audit. A team built the agents in four months. Cath Burnet, KPMG's UK head of audit, said that in several cases "the AI was able to go so much deeper" than the regular audit team. In some situations it had asked "better questions".

More than 100AI agents deployed to review the financial statements of an unnamed FTSE 250 company

Key figure. KPMG deployed more than 100 AI agents in Project Exponential, running alongside a conventional audit. Source: Financial Times

More than 100 agents. A real audit running beside them. Deeper work from the machines.

That lands right on something I argued.

I'd rather refine the idea than defend it by looking away from the news.

In that essay I split questions into two kinds. Task level questions ask how to do the work. Structural questions ask whether the work should look this way at all. I said the second kind was where people would matter most.

Read the KPMG account closely and the line moves a little but it holds. According to the FT the staff set the agents' objectives. They checked the agents' work periodically and challenged their conclusions before the audit opinion was finalised. In one area a human had to intervene after an agent failed to match a set of regulations to the company's documents.

So the agents asked better questions inside an audit that people had framed. They pushed harder on the evidence. People decided what the audit was for and when to step in and what counted as finished.

I gave the machines too little credit in February. The sharper question about the evidence can come from an agent now.

That makes my claim narrower. I said the future would belong to whoever was willing to ask better questions. I'd now say it belongs to whoever decides which questions the system gets to ask and what happens with the answers.

Pull quote, repeated from the essay. The author's revised claim about where the human advantage lies. Source: This essay

The FRC handed me a good example of that earlier this week. Its new guidance on auditor independence says that sharing insights from technology with management or those charged with governance can create threats to integrity, objectivity or independence in some circumstances. Especially where the work behind the insight isn't part of the audit.

Agents that dig deeper will surface more insights. Deciding which ones belong in a conversation with the client is a judgment call about the frame. That call sits with a person.

Two weeks before the questions essay I wrote that human in the loop is a design claim.

My argument was that approval doesn't equal control. Putting a person at the end of a process makes it defensible more than it makes it safe. I laid out three models. Over the loop, where people design the decision architecture. On the loop, where they monitor and step in on drift. Out of the loop, for low risk work where review adds noise.

An audit signature is the cleanest public test of that idea I've come across.

The UK regulator has been plain about where accountability sits. In March the Financial Reporting Council published guidance on generative and agentic AI in audit. It describes it as the first such guidance from any audit regulator in the world. It says regulatory accountability for AI tools and the quality of audit outputs is unchanged and that "the human auditor is always accountable".

So a person signs. The open question is where that person sits.

In the KPMG pilot an audit partner signed off the experimental audit. KPMG didn't rely on it when it formally signed off the company's accounts. The conventional audit was running alongside it.

I think that's the right way to do this. The partner could hold the agents' work up against a full human audit of the same company. The pilot signature had something solid next to it. That's a strong system testing a new tool carefully.

The harder part comes later. KPMG plans to expand the project to up to 20 larger audits. If it keeps growing there won't always be a second audit running in parallel and the signature will have to rest on the design itself.

Here's what I think it takes for a signing partner to be over the loop and not just near it.

The questions I asked in February still apply. In an audit they get more specific.

Did the partner shape what the agents were told to look for or inherit it?

Can the partner follow any agent's conclusion back to the evidence?

Who notices when an agent misses, the way a person had to step in on the regulations match?

When the partner's judgment and the agents' findings disagree, which one moves?

Which insights stay inside the audit and who decides?

KPMG's answer on the second question sounds encouraging. Burnet said the system gave full visibility into the agents' work and full traceability of it. That matters a lot. Traceability is what lets a partner interrogate the work at any point instead of only reviewing it at the end.

She also said human involvement stayed critical. I'd want to know what that involvement looks like across 20 larger audits and then across hundreds.

I still think the signature matters. What it vouches for is changing.

More and more it'll vouch for a system the signer understood well enough to trust and challenge. That's a different skill from doing the work and I don't think we've named it well yet.

Burnet was candid about the limits. She said few companies currently have clean enough data for agents to work with and the technology isn't yet available at the scale needed to run thousands of audits. So there's time.

And the regulator restated the accountability rule before the scale arrived. I'd take that order every time.

Most professional work still ends with a person's name on it. Over the next few years more of the work behind that name will run through systems like this one, and the useful thing to know is where your own judgment actually sits inside them.

Sources

AI asked 'better questions' than KPMG accountants in pilot audit of effectiveness. Ellesheva Kissin, Financial Times, syndicated by The Irish Times, November 18 2026. https://www.irishtimes.com/business/2026/11/18/ai-asked-better-questions-than-kpmg-accountants-in-pilot-audit-of-effectiveness/

Innovative new guidance supports audit firm adoption of emerging AI technologies. Financial Reporting Council, March 30 2026. https://www.frc.org.uk/news-and-events/news/2026/03/innovative-new-guidance-supports-audit-firm-adoption-of-emerging-ai-technologies/

New guidance supports auditor independence when using technology, including AI. Financial Reporting Council, November 16 2026. https://www.frc.org.uk/news-and-events/news/2026/11/new-guidance-supports-auditor-independence-when-using-technology-including-ai/