Home
Tax Technology/Europe

18 Nov 2026

KPMG AI Audit Trial: 100 Agents Out-Questioned Auditors

KPMG AI Audit Trial: 100 Agents Out-Questioned Auditors

**Last updated: November 2026 · Status: Pilot **

The KPMG AI audit experiment is the clearest signal yet that agentic AI is moving from the back office into the audit itself. In a pilot reported by the Financial Times on 18 November 2026, KPMG UK set more than 100 AI agents to work on the full annual audit of an unnamed FTSE 250 company. Its UK head of audit, Cath Burnet, said that in several cases "the AI was able to go so much deeper" than the human team, and sometimes asked better questions than the firm's own accountants.

Newsletter

The TaxSpoc Brief

Free weekly brief: the tax stories that matter, for advisors and in-house tax teams.

Free. Unsubscribe anytime. See our Privacy Policy.

If you are a CFO, finance director, audit committee member or tax lead, this affects you directly: the way your books are examined is changing, and the quality of your data will decide how fast.

KPMG AI audit trial: at a glance

AttributeDetail
FirmKPMG UK
Internal nameProject Exponential
Audited entityUnnamed FTSE 250 (UK-listed) company
AI agents deployed100+
ScopeEvery phase of the annual audit, run in parallel with the conventional audit
Build timeAbout four months
Human teamTwo partners and a director, plus senior managers and junior staff
Statutory opinionBased on the regular human-led audit, not the AI work
Next stepExtension to up to 20 larger audits
RegulatorFinancial Reporting Council (FRC)

What did KPMG's AI agents actually do?

KPMG's agents worked across the whole audit rather than on isolated admin tasks. They gathered evidence, analysed and documented data, drafted conclusions and observations, suggested next steps and divided responsibilities among themselves. Humans set their goals, supplied context and challenged their conclusions.

That is what separates this trial from earlier AI-in-audit headlines. Most audit tools so far have been assistants: a model that summarises board minutes or vouches a sample of expenses. Project Exponential instead used many specialised agents coordinating with each other. That is the agentic model the FRC describes as AI that can orchestrate and carry out multiple tasks toward a goal with some degree of autonomy.

KPMG says the system also keeps its own audit trail, logging every agent interaction with a document down to individual characters and timestamps. For a profession built on evidence, that traceability matters as much as speed.

Did AI agents really ask better questions than human auditors?

In some cases, yes, according to KPMG itself. Burnet said the agents dug deeper into certain issues, flagged areas that needed more investigation and raised questions the human team had not necessarily asked. This is partner testimony rather than a published benchmark, so treat it as a strong qualitative signal, not a scored result.

Why would software ask better questions? There are three structural reasons:

  1. No sampling ceiling. Human auditors test samples because time is limited. Agents can examine far more transactions, so anomalies that would never land in a sample get surfaced.
  2. No fatigue or anchoring. An agent reviewing the 4,000th journal entry is as attentive as on the first, and less tempted to accept last year's explanation.
  3. Parallel specialisation. With 100+ agents, each can go deep on a narrow area, such as revenue cut-off, related parties or leases, at the same time.

The trial also showed the limits. A human had to step in when one agent failed to match a set of regulations correctly to the company's documents. Interpreting rules against messy real-world records is still the hard part, and it is where professional judgement lives.

Who signed off: the AI or the auditor?

The human auditor. An audit partner signed off on the experiment, but the official audit opinion was based on the conventional audit, after human review of the agents' conclusions. KPMG calls human involvement "critical" and has given the FRC a high-level overview of the project.

That matches the regulator's position. In its Generative and Agentic AI Guidance of March 2026, the FRC states that the human auditor is always accountable and that the guidance does not change that.

The FRC's new guidance on AI insights and auditor independence

Two days before the KPMG news, on 16 November 2026, the FRC published guidance on auditor independence when using technology, including AI. It covers when auditors may pass technology-derived insights to management or audit committees, and when giving clients access to the tools themselves could compromise integrity, objectivity or independence.

The tension is real. Because agents can analyse whole populations rather than samples, they surface things beyond the audit's scope: inefficient processes, cost-saving opportunities, unusual patterns. Shared carelessly, those findings start to look like a pitch for consulting work, which the FRC's Ethical Standard strictly limits for audit clients. Significant control weaknesses must still be reported, and the FRC is clear that those disclosures are not non-audit services.

The UK now has the most developed rulebook for AI in audit: the FRC describes its March guidance as the first of its kind from any audit regulator.

Is the rest of the Big Four doing the same?

Yes, though KPMG is the first to publicise a full parallel audit of a listed company. PwC's global assurance leader, Andy Hammond, has said the firm has run several parallel audits to see how AI performs without relying on it. Deloitte and EY are also investing heavily in agentic tools.

KPMG has been building toward this for some time. Its global audit platform, KPMG Clara, began adding AI agents for routine substantive procedures in 2025, and in April 2026 the firm said it would pilot orchestration agents that direct other AI tools.

What the KPMG AI audit means for finance and tax teams

Your audit will increasingly be run by software reading all of your data, not a sample. The cleanliness and structure of that data will decide audit cost, speed and how many awkward questions you get.

KPMG itself named poor data quality as a key barrier to scaling the approach: few companies have records clean enough for agents to work with reliably. That is why it is moving to 20 larger audits, not thousands.

Practical steps for finance leaders:

  1. Standardise your source data. Structured, machine-readable transactions are far easier for agents to test. Mandatory e-invoicing is already pushing businesses this way; see our global e-invoicing roadmap for the mandates arriving in your jurisdictions.
  2. Expect more questions, not fewer. Full-population testing surfaces outliers that sampling missed. Prepare explanations for unusual journals, manual adjustments and related-party flows.
  3. Ask your auditor how AI is used. Audit committees should ask which procedures are agent-run, how outputs are reviewed and how client data is secured.
  4. Separate insight from advice. If your auditor's AI produces efficiency findings, check how they are shared under the FRC's independence guidance.
  5. Watch the fee conversation. KPMG acknowledges clear time savings but has not said whether audit fees will fall.

Need help getting your data audit-ready? Find tax technology and e-invoicing advisers in the TaxSpoc Directory.

The bottom line

The KPMG AI audit pilot does not replace the auditor: the partner still signs, and a human still had to intervene. But it shows that agents can already cover a full audit cycle and, at times, probe harder than the people supervising them. The audit firms will move first, and businesses with clean, structured data will feel the benefits soonest.

Frequently asked questions

What is KPMG's Project Exponential?

Project Exponential is KPMG UK's experiment in running more than 100 AI agents through every phase of an annual audit of an unnamed FTSE 250 company, in parallel with a conventional human audit. The agents gathered, analysed and documented evidence and drafted conclusions. KPMG built the system in about four months.

Did AI replace the human auditors in the KPMG trial?

No. The statutory audit opinion was based on the regular human-led audit. An audit partner signed off on the experiment, humans set the agents' goals and reviewed their conclusions, and a human had to intervene when one agent failed to match regulations to the company's documents.

Is AI allowed in UK statutory audits?

Yes. The FRC supports appropriate use of AI and has published guidance on generative and agentic AI in audit. The human auditor and the audit firm remain fully accountable for audit quality and the opinion, whatever technology they use.

Will AI make audits cheaper?

Possibly, but it is not confirmed. KPMG says the technology offers clear time savings but has declined to say whether audit fees will fall. Data quality is a major constraint, so companies with messy or unstructured records may see smaller savings.

How should companies prepare for AI-driven audits?

Focus on clean, structured, machine-readable financial data, including e-invoices, and document the reasons for manual journals and unusual transactions. Audit committees should ask auditors which procedures use AI and how the outputs are reviewed and secured.

Published by
TAXSPOC News Desk

TaxSpoc is a free platform sharing tax news, insights, and practical content with a growing Global Tax Community. Our goal is simple: democratize tax knowledge and make high-quality tax content accessible to everyone—no paywalls, no fluff.

For any questions, clarifications, contributions, or concerns related to this article, don’t hesitate to contact us at hello@taxspoc.com.