An audit is built on questions: Which transactions look unusual? Does the evidence support management’s explanation? Have the correct accounting rules been applied? KPMG’s experiment with more than 100 AI agents in auditing produced a provocative result. According to the firm’s UK audit chief, the agents sometimes asked better questions and investigated issues more deeply than human accountants.
That does not mean artificial intelligence successfully replaced an audit team. The system, developed under KPMG Project Exponential, ran a parallel audit of an unnamed FTSE 250 company’s financial statements. Its findings were not used to approve the accounts, and a human auditor had to intervene when one agent failed to match regulations correctly with company documents. No comparative accuracy rates were published.
The experiment nevertheless offers one of the clearest previews yet of agentic AI in professional services: software that does not merely answer prompts, but gathers information, coordinates tasks, follows evidence and proposes conclusions across a complex workflow.
What KPMG Project Exponential actually tested
KPMG spent about four months developing Project Exponential before deploying more than 100 specialized autonomous agents across different phases of a real company audit, according to Financial Times reporting and KPMG’s public account of the work. The client was identified only as a FTSE 250 company, protecting the confidentiality expected around sensitive financial records.
Rather than creating one general-purpose AI auditor, KPMG divided the work among agents with narrower responsibilities. Across the parallel engagement, they could:
- Gather financial information and locate relevant source material.
- Read company documents, policies and supporting records.
- Analyze evidence for inconsistencies, gaps or unusual patterns.
- Document the evidence considered and work performed.
- Pass findings to other agents responsible for related procedures.
- Suggest follow-up questions and additional investigations.
- Prepare draft observations and conclusions for human review.
This multi-agent structure resembles the way an audit team distributes work among specialists, junior staff, managers and partners. An agent examining a balance could refer a concern to another agent focused on contracts or regulatory requirements. A coordinating layer could then track the issue through further analysis.
KPMG said the technology could generate significant time savings. It has also discussed expanding the experiment to as many as 20 larger audits. That planned expansion should not be mistaken for full autonomous deployment: the initial engagement remained a controlled, parallel exercise rather than part of the formal audit opinion.
Why this was more than a chatbot reading spreadsheets
Conventional generative AI tools generally wait for a person to submit a prompt. They might summarize a spreadsheet, explain a variance or draft an email, but the human user decides what to ask next and moves information between tasks.
Project Exponential was designed as an end-to-end professional workflow. Its agents could pursue assigned objectives, use available tools, preserve working context, communicate with other agents and recommend another step. This is the distinction between a chatbot and agentic AI: one generates a response, while the other participates in a sequence of actions.
That difference matters in artificial intelligence auditing. A financial statement audit is not a single calculation. It requires planning, risk assessment, evidence collection, regulatory interpretation, documentation and repeated challenges to management’s assertions. Multi-agent AI systems may be able to examine far more documents than a person and revisit a question without fatigue or pressure from a deadline.
KPMG’s observation that agents sometimes probed more deeply is therefore plausible. Software can keep asking what supports a claim, search connected records and explore alternate explanations. It is not constrained by hierarchy or worried that another question will appear inconvenient. But persistence is not the same as judgment, and a deeper investigation can still reach the wrong conclusion.
The regulatory failure reveals the central weakness
One incident in the trial illustrates why AI audit automation remains dependent on experienced professionals. A human auditor had to step in after an agent failed to match relevant regulations with the company’s documents.
This was not a trivial formatting error. Auditors must understand which rule applies to a particular entity, transaction, reporting period and set of circumstances. Similar terminology may carry different meanings across accounting standards, legislation, regulatory guidance and company policies. The correct answer often depends on context that is implicit rather than neatly stated in a document.
The episode highlights several unresolved risks:
- Contextual understanding: An agent may retrieve a relevant-looking rule without understanding whether it applies to the facts.
- Regulatory interpretation: Accounting requirements frequently demand judgment rather than simple keyword matching.
- AI hallucinations in finance: A fluent explanation may cite unsupported facts, merge separate rules or conceal uncertainty.
- Professional skepticism: An audit requires disciplined doubt, including recognition that evidence supplied by management may be incomplete.
- Autonomous reasoning limits: A chain of individually plausible steps can produce a materially incorrect conclusion.
These weaknesses help explain why the trial does not prove that AI auditors outperform professionals overall. KPMG reported encouraging behavior, including better questions in some instances, but it did not publish independently verified accuracy results, error rates or a controlled comparison against human teams. The claim describes selected performance within an experiment, not a general superiority finding.
Human oversight must be designed into the audit
The UK’s Financial Reporting Council has emphasized issues such as auditor independence, integrity, accountability and the danger of over-reliance as AI enters audit work. An audit firm cannot delegate its legal and professional responsibilities to software, even if that software appears highly capable.
Human oversight in AI must therefore involve more than approving an automatically generated conclusion. Auditors need to challenge an agent’s reasoning, inspect its supporting evidence, test contradictory explanations and determine whether the work meets applicable standards. A qualified professional remains responsible for the opinion and must be able to explain why the conclusion is justified.
Controls needed for AI financial reporting and audit work
Any production system handling client records will need controls comparable to, and in some cases stronger than, those governing human audit teams:
- Traceability: Reviewers should be able to reconstruct which records, rules, prompts, models and tools informed a conclusion.
- Complete audit trails: The system must preserve actions, handoffs, overrides and changes rather than only its final answer.
- Access controls: Each agent should receive only the financial data and tools necessary for its role.
- Data quality checks: Incorrect, incomplete or poorly classified inputs can undermine even a technically capable model.
- Explainability: Auditors need intelligible reasons and evidence, not an unsupported risk score.
- Escalation rules: Ambiguous regulations, conflicting evidence and high-risk judgments should trigger mandatory human review.
Model monitoring and version control also matter. If an AI accounting software provider updates a model during an engagement, the firm’s documentation must show which version performed each procedure and whether the change affected reproducibility.
KPMG and PwC are testing a Big Four operating model
KPMG is not alone. Financial Times reporting has described PwC running parallel AI audit projects as Big Four accounting firms explore how autonomous systems can support professional work. Approaches differ, but the common strategy is cautious: test agents alongside established teams, compare outputs and keep their conclusions outside the formal opinion until governance and reliability are demonstrated.
The KPMG UK experiment stands out because of its scale and coordination across more than 100 agents. PwC’s work reinforces the broader point that the industry is moving beyond isolated generative AI assistants toward systems that can participate in complete processes.
These parallel trials also limit risk. Firms can assess whether agents miss evidence, duplicate work, misinterpret standards or produce convincing but false explanations without making the technology the basis of a signed opinion. Regulators and audit committees will expect evidence from repeated engagements, not a single successful demonstration.
What AI agents could mean for accountants and fees
Much of today’s junior audit work involves collecting documents, reconciling figures, completing templates and documenting procedures. Automated financial audits could reduce the time required for those tasks. Future junior accountants may spend less time moving data and more time investigating exceptions, validating AI output and understanding the client’s business.
That transition creates a training challenge. Repetitive procedures have traditionally helped new accountants learn how transactions flow through an organization. If agents absorb that work, firms will need deliberate ways to teach accounting fundamentals, professional skepticism, interviewing skills and regulatory judgment. Knowing how to supervise AI cannot substitute for knowing when its answer is wrong.
Staffing models may also change. Audit teams could become smaller, more specialized and more technology-intensive, combining accountants with data engineers, model-risk experts and cybersecurity professionals. Claims about AI agents replacing accountants are therefore too simplistic. Specific tasks and entry-level roles may shrink, but accountability and complex judgment become more important, not less.
There is also an unresolved commercial question. KPMG expects meaningful time savings but has not confirmed that clients will receive lower audit fees. Firms may argue that savings will fund better coverage, larger samples, deeper investigation and expensive technology controls. Clients may instead expect productivity gains to reduce prices.
That tension reaches the traditional billable-hours model. If AI financial analysis completes days of work in minutes, charging primarily for staff time becomes harder to justify. Professional services firms may move toward fixed, outcome-based or technology-inclusive pricing. For now, the cost of building, securing, validating and governing these systems complicates any simple equation between fewer hours and lower fees.
The likely future: AI-enabled and human-assured
KPMG Project Exponential shows why AI agents in accounting deserve serious attention. Coordinated systems can search extensive evidence, maintain parallel lines of inquiry and identify follow-up work at a scale human teams cannot easily match. They may also ask useful questions without fatigue, hierarchy or deadline pressure.
But the trial equally demonstrates why autonomous agents cannot be treated as independent professionals. One failed connection between regulations and company documents can alter an audit conclusion. Without reliable context, traceable evidence and skeptical human review, greater speed simply allows errors to travel faster.
The most credible future is therefore AI-enabled and human-assured. Agents will perform large-scale analysis, prepare documentation and surface anomalies. Experienced auditors will test those findings, interpret rules, challenge management and accept responsibility for the final opinion. Project Exponential is an important experiment—not proof that machines have replaced auditors, but evidence that the structure of audit work is beginning to change.
Frequently asked questions
Did KPMG use AI to approve the FTSE 250 company’s accounts?
No. The agents conducted an experimental parallel audit. Their findings were not used to formally approve the unnamed company’s financial statements, and the official audit remained the responsibility of human professionals.
Did KPMG prove that AI agents are more accurate than accountants?
No. KPMG said agents sometimes asked better questions and investigated issues more deeply, but it did not publish comparative accuracy rates or independent evidence that AI outperformed professional auditors overall.
What went wrong during Project Exponential?
A human auditor intervened when an agent failed to match regulations correctly with company documents. The incident exposed limitations in contextual understanding and regulatory interpretation.
Will AI agents replace junior accountants?
They are likely to automate portions of evidence gathering, reconciliation, analysis and documentation. However, junior staff will still need to verify evidence, investigate exceptions and develop the judgment required to supervise AI-generated work.