Incident sheet · /enterprise-ai

Filed
Read
3 min · 611 words
Severity
Routine

KPMG conducts first full audit using more than 100 AI agents

KPMG used more than 100 AI agents in a parallel FTSE 250 audit test. Agents asked deeper questions in some cases, but human intervention was still required.

AI asked better questions than auditors in a large-scale KPMG test
FeedAI asked better questions than auditors in a large-scale KPMG test — Tumisu / pixabay

Event log

  1. More than 100 AI agents reviewed an FTSE 250 company's annual report in KPMG's first full audit using AI.
  2. The trial ran in parallel with a traditional audit and did not support formal approval.
  3. A human auditor intervened on one occasion after an agent failed to link correct regulations.
  4. KPMG plans to expand the test to up to 20 major audits.
  5. Commercial availability of AI-based audits remains KPMG's long-term goal.

By Rebecca Stone

The Financial Times reported that audit firm KPMG tested more than 100 AI agents in its first full audit using the technology. The project reviewed the annual report of an anonymous British publicly traded company in the FTSE 250, according to the report. The FTSE 250 is an index of smaller UK-listed companies.

What did the agents review?

Each AI agent (an autonomous software program that gathers information and proposes actions) independently collected and analyzed information. KPMG said the agents documented their findings and formulated preliminary conclusions, meaning working findings subject to audit review. They also collaborated with one another and suggested the next steps in the audit.

The reported workflow covered information gathering, analysis, documentation, preliminary conclusions, collaboration, and workflow suggestions. Human auditors worked in parallel, but the available account does not assign each step exclusively to the agents. The account also gives no client name, sector, audit duration, or broader deployment figure.

Did the AI replace human approval?

No. The Financial Times reported that KPMG conducted the AI audit in parallel with a traditional audit, meaning a second review ran alongside the conventional process. KPMG did not use the AI results as the basis for its formal approval of the company's financial statements. The distinction is material: “full audit” described KPMG's use of the system, not a transfer of approval authority to software.

The trial therefore tested agent performance without making autonomous output the source of KPMG's audit opinion, its formal conclusion on the financial statements. KPMG retained the conventional review as the formal basis for approval. The available account identifies no separate AI-generated audit opinion and no final accounting judgment delegated to an agent.

Did the agents outperform the human reviewers?

KPMG reported that the system succeeded in several instances in “asking better questions than its human colleagues.” The firm also said the agents went significantly deeper into the review in those cases. That comparison came from KPMG's own account, and no independent evaluator was identified.

KPMG reported that the technology still made errors. On one occasion, KPMG said a human auditor intervened when an AI agent failed to link the correct regulations to the company's documents. Regulations here mean the reporting rules applicable to the company and its audit work. The failure required human correction rather than autonomous resolution by the agent.

The available account gave no total count of interventions, no measured error rate, and no time or cost comparison. Because the AI work did not support formal approval, KPMG's account links the documented failure to intervention rather than to a reported opinion error. That preserves the distinction between test performance and an audit conclusion.

What happens after the trial?

The Financial Times reported that KPMG plans to expand the test to include up to 20 major audits. KPMG's long-term goal, as reported, is to offer AI-based audits commercially. The 20-audit step would extend the exercise beyond the single FTSE 250 company, although the available report gives no start date or selection criteria.

The commercial goal would move AI-based audit work from an internal trial to a client service. However, the announcement specifies no launch date, pricing, service commitment, or client-availability timetable. It also gives no model name, AI vendor, software version, project cost, time saving, or error rate for the first test.

The disclosed next step is therefore a larger test with human approval still intact. Commercial availability remains a stated long-term objective. KPMG did not say that the first trial replaced traditional audit work, removed human sign-off, or changed the basis for its formal approval.

via ft.com (Original)

Filed under

  • kpmg
  • ai-agents
  • audit
  • audit-automation
  • human-oversight
Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Senior reporter covering money, funding and market shifts around enterprise information technology and the leaders who run at CIO Monitor.

202 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous article