On November 18, 2026, KPMG announced that it had completed the first end-to-end real audit by AI agents in the history of the Big Four accounting firms: more than 100 AI agents ran through every stage of the annual audit of a UK FTSE 250 listed company. Internally called “Project Exponential,” the project ran in parallel with a real audit team; the AI agents independently collected and analyzed data, drafted conclusions, and proposed follow-up recommendations. Cath Burnet, KPMG UK’s head of audit, said that in several stages AI “can dig much deeper than a conventional audit team” and even raised “better questions than auditors.” The final statutory audit opinion was still issued by the conventional team and signed by a partner, and the AI conclusions were not used to approve the formal financial statements.
Only one failure was recorded throughout the project: an AI agent could not match regulatory documents with company documents and required human intervention. The task required understanding regulatory intent, judging scope of application, and identifying exception clauses, pointing to the boundaries of AI in semantic understanding and contextual reasoning. Such interpretive work accounts for a high proportion of auditing.
KPMG built the entire system in four months, developed jointly by two partners, a director, senior managers, and junior staff. The system provided full visibility and traceability, recording each agent’s interaction with documents down to character and timestamp granularity.
Structural Reasons Why AI “Asks Better Questions”
Traditional audits rely on sampling; auditors are limited by time and manpower, and the number of documents they can read closely each day is finite. AI agents can traverse the full dataset in the same amount of time, systematically search every transaction and note, identify anomalous patterns likely to be missed by sampling, and raise questions. This stems from AI’s coverage radius, not a higher level of intelligence.
Data Quality: The Bottleneck Hit Earlier Than AI Capability
Cath Burnet said the technology is not yet ready to support thousands of audits because “very few companies have data clean enough for AI agents to work reliably.” KPMG plans to first expand to as many as 20 larger audits before considering roll-out to clients. Audit inputs include ledgers, contracts, bank statements, and internal control records; these data are often spread across systems, inconsistently formatted, and burdened by legacy issues. For AI agents to work autonomously, they need a highly structured and clearly defined data environment, which most companies’ real-world IT conditions do not yet provide. At present, the bottleneck to scaling is the maturity of client-side data governance, not model capability.
A Signal from Regulators: Issued Two Days Earlier, Pointing Directly at Independence
The Project Exponential results were announced on November 18; two days earlier, the UK audit regulator, the Financial Reporting Council (FRC), published guidance on maintaining auditor independence when using AI. AI’s full-population analysis capability may uncover large numbers of “anomalies” that would not originally fall within the scope of a statutory audit; if used to sell consulting services to clients, this would constitute a threat to independence. Traditional sampling naturally limits the number of such findings, while AI’s full coverage removes that limit. The FRC warned that AI may produce a “confirmation bias” and that traditional independence safeguards may not necessarily be effective in an AI environment.
The Unresolved Question of Fee Structure
KPMG confirmed that there are “clear time-saving opportunities” but declined to say whether it would reduce client fees. KPMG International has already argued to its own auditors that “AI should lower costs” and secured lower fees. As an external audit service provider, it has not yet said whether it will apply the same logic to client pricing. Audit pricing has long been tied to labor hours; as AI agents reduce labor demand, the basis for fees faces pressure to be restructured.
Assessment: A Quantified Red Line, Not a Technical Showcase
Project Exponential is the first set of end-to-end boundary data on AI-agent audits from a real business scenario, covering 100+ agents, a FTSE 250 company, a complete audit cycle, and one documented failure. The data show three things: AI agents have surpassed humans in quantitative coverage, with practical evidence for full-population traversal and anomaly identification; AI agents still have gaps in interpretive regulatory reasoning; and institutional obstacles are no lower than technical ones—data quality, regulatory independence rules, and fee models determine whether the approach can scale from 20 audits to 2,000. The current state of commercial deployment for AI agents is this: in some dimensions they are better than humans, in others they are not yet ready, and the hardest problems are often not in the model.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接