KPMG Project Exponential: 100+ AI Agents Audited a FTSE 250 Company in Parallel KPMG UK ran Project Exponential, deploying more than 100 AI agents across every stage of an experimental audit of an unnamed FTSE 250 company in 2026 alongside a conventional team, with agents built over four months and humans retaining formal sign-off, according to Financial Times reporting. KPMG UK head of audit Cath Burnet said the agents sometimes asked "better questions" and dug deeper, though one agent failed to match regulations to company documents and required human override. The UK Financial Reporting Council published guidance on November 16, 2026 on auditor independence when using technology, and a Centre for Public Interest Audit report cited by the FT found 23% of audit committee chairs had limited confidence in how AI is used in their company's audit. KPMG UK deployed more than 100 AI agents across every stage of an experimental audit on an unnamed FTSE 250 company in 2026 , running Project Exponential beside a conventional team. Financial Times reporting https://www.ft.com/content/5168b3e9-75d0-4be0-9ccd-d327d6b16c05?ref=explainx , carried by Business Standard https://www.business-standard.com/industry/news/kpmg-finds-ai-agents-ask-better-questions-than-human-auditors-in-trial-126111800471 1.html?ref=explainx , quotes Cath Burnet , UK head of audit, saying agents sometimes asked "better questions" and dug deeper — while humans kept the formal sign-off . For agent builders, the pilot is a production-shaped workload : multi-document reasoning , regulatory mapping , agent-to-agent handoffs , and human challenge before an external opinion ships. What Project Exponential did | Element | FT-attributed detail | |---|---| | Scale | 100+ agents across entire audit process | | Target | Unnamed FTSE 250 UK-listed company | | Build time | Four months to develop agents | | Human shape | Two partners, a director , senior managers, juniors supervising agents | | Agent duties | Collect data, analyze, document, draft conclusions, propose next steps | | Formal opinion | Not based on agent output for published accounts | Burnet said the experimental team had a "different shape and size" but human involvement stayed "critical." One cited failure mode: an agent could not match regulations to company documents , requiring human override . Why "better questions" matters for agent design Audit is not autocomplete. It is adversarial questioning of management assertions. Burnet's "better questions" line implies agents surfaced follow-ups junior staff might skip under time pressure — similar themes to higher-scoring coding models and merge rates https://explainx.ai/blog/higher-scoring-coding-models-swe-bench-merge-rate-natural-experiment-2026 where tooling changes behavior , not only scores . Architectural lessons: 1. Explicit challenge loops before conclusions finalize. 2. Regulation corpora versioned per jurisdiction — the UK FRC Ethical Standard updates matter. 3. Provenance logs — KPMG told the FRC it records character-level interactions with timestamps firm claim, not independently verified . Independence and the FRC November 2026 guidance The UK's Financial Reporting Council published guidance November 16, 2026 on maintaining auditor independence when using technology. Mark Babington said "Whatever technology a firm uses, the auditor remains accountable." Risk: sharing AI-generated insights with audit clients outside the financial statement audit can threaten objectivity . That limits how eagerly firms productize "free AI insights" as client perks . A Centre for Public Interest Audit https://www.ft.com/content/5168b3e9-75d0-4be0-9ccd-d327d6b16c05?ref=explainx report cited in FT coverage found 23% of audit committee chairs had limited confidence in how AI is used in their company's audit — a demand signal for transparent agent runbooks . Fee pressure versus pyramid economics Big Four audits historically bill large junior teams doing routine testing under partner supervision. Agents that draft conclusions attack that pyramid . KPMG acknowledged "clear time-saving opportunities" to the FT but would not say fees would fall . Irony: KPMG International has argued its own auditor should charge less because AI makes work cheaper — a tension founders selling agent SRE tools will recognize when buyers demand price cuts but vendors need margin . PwC and the parallel audit pattern Andy Hammond , PwC global assurance leader, told the FT the firm ran parallel audits "to see how AI performs across the whole process without relying on it." That shadow mode is how enterprises should deploy coding agents too: Claude code review risks https://explainx.ai/blog/claude-code-team-code-review-process-ai-reviewer-risk-tiers-reverts-down-40-percent-2026 discuss human gates before merge. KPMG's Clara intelligence platform marketing already describes agentic execution with auditor oversight — Project Exponential is the named UK experiment behind that narrative. Limits Burnet acknowledged - Technology not at scale for thousands of audits yet. - Many companies lack clean enough data for agents. - Some clients may decline AI-assisted audits. Those constraints rhyme with Revolut's fincrime GPT story https://explainx.ai/blog/revolut-openai-codex-fincrime-gpt-5-5-enterprise-story-2026 : models work when pipelines are curated . How this differs from consumer agent hype No Felony Bench link — this is regulated professional work , not third-party compromise . Security teams should still ask whether agents can leak PII from trial balances to external model APIs ; the FT piece does not detail data residency — a gap for CISO questionnaires . Financial document review workflow echoing KPMG agentic audit parallel testing Image: Library reference — audit agents need instruction files and evidence trails like any enterprise agent. Expansion roadmap Burnet plans up to 20 larger audits next. Watch for: - Client consent clauses in engagement letters. - FRC enforcement if client-facing AI insights blur independence . - Competitive responses from EY, Deloitte, PwC already piloting private company agent tools per Bloomberg Tax on KPMG private audits https://news.bloombergtax.com/financial-accounting/kpmg-tests-advanced-ai-on-private-company-audit-clients-first?ref=explainx . What the FT does not prove - Which foundation models or agent frameworks KPMG used. - Accuracy rates versus humans on same samples . - Fee quotes for AI-assisted engagements. Treat "better questions" as qualitative partner testimony , not a benchmark leaderboard . Agent orchestration patterns implied by the pilot Project Exponential is not one chat thread . The FT describes agents that work autonomously with one another and suggest subsequent steps — a multi-agent workflow closer to Microsoft Agent 365 https://explainx.ai/blog/microsoft-ignite-2026-complete-coverage-every-announcement-november-2026 narratives than to single-shot Copilot prompts . Implied control plane requirements: | Control | Why auditors care | |---|---| | Objective scoping | Staff set engagement-specific goals so agents do not roam into unrelated subsidiaries . | | Periodic human review | Partners challenge conclusions before they enter workpapers . | | Versioned regulations | UK standards change; agents must cite rule versions agents used. | | Immutable logs | FRC scrutiny expects defensible evidence chains if a filing is questioned years later. | If you sell agent observability , this pilot is a reference sale for finance verticals — even though KPMG has not named vendors. Comparison with coding-agent benchmarks Developers measure agents with SWE-bench and merge rates . Auditors measure material misstatement risk and regulatory defensibility . The FT quote about better questions is closer to "did the agent find the right anomaly to test?" than "did it pass unit tests?" That distinction matters when finance teams import coding agent hype . A repo-wide coding agent can ship a broken feature ; an audit agent that misses a fraud pointer can destroy licenses . Higher-scoring coding models https://explainx.ai/blog/higher-scoring-coding-models-swe-bench-merge-rate-natural-experiment-2026 show scores do not always improve outcomes — audit may face the same Goodhart problem if firms optimize for agent speed not detection quality . Data cleanliness as the hidden gate Burnet said few companies have sufficiently clean data for agents. In audits, clean means consistent chart of accounts , documented controls , and machine-readable prior workpapers — rare outside large listed filers . Builders pitching RAG over PDFs should expect months of normalization before agent autonomy is allowed near sign-off , echoing enterprise Copilot rollouts https://explainx.ai/blog/microsoft-copilot-autopilot-m365-frontier-rollout-november-2026 where Frontier gates precede GA . Investor and committee confidence gap The FT-cited Centre for Public Interest Audit numbers — 23% of audit committee chairs with limited confidence in audit AI use, 15% of investors similar — mirror enterprise buyers asking "show me the eval harness" before production agent access . Product marketing should offer committee-ready summaries : what models ran, what data never left the tenant, what humans signed. When AI-assisted audits might ship to clients Burnet's 20-audit expansion and eventual client offering signal a multi-year ramp , not next quarter GA . Watch engagement letters for: - Client opt-in to agent testing. - Confidentiality clauses for model providers . - Liability allocation if an agent misses a material misstatement during pilot phases. None of those terms are public yet; FT reporting stops at partner quotes . Readers who need contract-level detail should wait for KPMG or the FRC to publish more. Security posture for finance agents This story is not a breach, and the FT report describes no security incident. Still, audit agents touch payroll, revenue, and tax positions — prime PII and MNPI . CISOs should ask KPMG-style vendors the same questions we pose for MCP misbinding https://explainx.ai/blog/mcp-tool-provider-misbinding-ai-agents-security-2026 : - Can an agent call tools outside the engagement folder ? - Are prompts logged with retention matching regulatory holds ? - Is inference on-prem or API with zero-retention contracts? Hacker News context The FT story hit Hacker News https://news.ycombinator.com/item?id=50542658&ref=explainx with modest points but ties into "will AI replace jobs" threads. The defensible answer from primary reporting : agents augmented questioning on one pilot , did not sign accounts , and regulators warned on independence the same week. Practical checklist if your CFO asks for audit agents 1. Map every data class agents may read payroll, contracts, tax . 2. Require human sign-off on any output that could enter regulatory filings . 3. Log tool calls with retention matching SOX holds. 4. Run shadow mode — PwC's parallel audit pattern — for at least one close cycle before client-facing promises . 5. Train audit committees with FRC-style independence scenarios, not only demo wow moments . KPMG's pilot suggests the Big Four will sell agent-assisted assurance as premium diligence , not cheap automation , until regulators and committees gain confidence. Expect 2027 engagements to mention agent logs in audit planning memos the way SOC 2 evidences appear today, especially for listed filers with complex revenue recognition . Related reading Project Exponential details come from Financial Times reporting cited by Business Standard unless KPMG publishes a primary technical paper.