AI Audit Agents: What KPMG’s Trial Means for Firms
KPMG’s AI audit agents examined a FTSE 250 company and challenged human auditors. Learn what the trial reveals about audit costs, risks, oversight and adoption.

When AI Agents Challenge Human Auditors: What KPMG's Trial Means
KPMG's UK audit team says that when AI agents examined a major listed company's accounts, they sometimes dug deeper and asked questions its human auditors hadn't raised. That's a bigger claim than "AI can summarize documents" or "AI can speed up routine work." It points to something more uncomfortable - and more interesting: that agents may affect the judgment at the heart of an audit.
The trial, called Project Exponential, used more than 100 agents working on the financial statements of an unnamed FTSE 250 company. The agents handled tasks across the audit workflow, from collecting and analyzing information to drafting observations and proposing next steps. But this wasn't a statutory audit used to approve the company's accounts. Human oversight was still required, including when an agent failed to match regulatory requirements to the company's documentation.
That distinction matters. A promising trial doesn't prove AI can replace an audit team. It also doesn't automatically mean automation will cut fees or improve audit quality at scale. What it does show is why firms are testing agents beyond basic administrative support. This article breaks down what the trial actually tested, how it could affect audit economics, where the risks are, and what a responsible adoption plan looks like.
What did KPMG's AI audit agent trial actually test?
Project Exponential tested whether AI agents could take on connected audit tasks - not just help with one isolated step. The Financial Times reported that KPMG deployed more than 100 agents on the accounts of an unnamed FTSE 250 business. Cath Burnet, KPMG UK's head of audit, said the agents sometimes went further into issues and raised questions the human team hadn't necessarily thought to ask.
That's an important difference. A conventional automation tool might pull numbers from a ledger or sort documents into folders. An agent-based system is designed to carry work through several stages: find relevant information, analyze it, record what it did, draft an observation, and suggest what should happen next. Agents can also split up parts of a bigger investigation and coordinate across them.
What the agents did - and did not do
The trial covered activities across the audit process, including information gathering, analysis, and documentation. Its value wasn't just the amount of work completed. The point was whether agents could surface inconsistencies or new lines of enquiry that a human auditor should take seriously.
At the same time, this was not a formal statutory audit. An audit partner signed off on the experimental exercise, and the results weren't used to approve the company's accounts. So the right way to read the outcome is as evidence of capability under trial conditions - not proof that an AI system can issue a dependable audit opinion.
KPMG reportedly built and refined the technology over about four months. The firm says it plans to roll it out to up to 20 larger audits, while also noting the system isn't ready for broad use across thousands of engagements. A limited rollout can reveal failure modes and weak spots, but it can't settle questions about performance across different companies, different data setups, and different audit teams.
Why "better questions" is a meaningful claim
An audit isn't only about checking whether the numbers add up. Auditors also have to pressure management's explanations, investigate unusual patterns, and decide whether the evidence is enough. If an agent spots a discrepancy that the team missed, it could improve the audit by expanding what the auditors consider.
But finding a question is not the same as answering it correctly. The agent's output has to be traced back to evidence, checked in context, and challenged by someone who understands the business and the relevant accounting rules. The trial's most useful takeaway, then, isn't that machines have become auditors. It's that agents may change how audit teams find and prioritize the questions that matter.
Why could AI agents change audit economics?
A traditional audit model depends on teams of junior accountants doing detailed work under the supervision of managers and partners. A chunk of that work is repetitive: collecting records, reconciling information, documenting procedures, and following up on exceptions. AI agents could reduce the time spent on some of that workload - especially where information is consistent and easy to access.
The business case follows a familiar logic. If a system processes more information in less time, firms may be able to complete parts of the work faster, or test a broader set of transactions. Senior professionals could then focus more on complex judgments, exceptions, and discussions with management. For big firms like KPMG, Deloitte, PwC, and EY, even small changes to a labour-heavy workflow can affect staffing, delivery capacity, and the cost structure of engagements.
The economics depend on the whole workflow
It's tempting to equate "faster processing" with "lower audit costs." In reality, the economics depend on what comes before and after an agent runs. Someone still has to prepare data, configure and monitor the system, review what it produces, investigate mistakes, and document why the output is reliable. If all that review work stretches to fill the gap, the savings might be limited.
A better way to judge the business case is to measure the full workflow, not just the agent's speed. Ask:
- How much staff time does the process take today?
- How much of the time goes into preparing and validating inputs?
- How much review and rework does the agent's output create?
- Does the system help identify real exceptions, or does it just generate more alerts?
- Does it improve coverage or evidence quality enough to justify the system's operating cost?
The answers will vary by audit task. Automating a stable, well-documented procedure may be genuinely efficient. Applying the same setup to judgment-heavy areas with incomplete evidence could add risk - and increase review effort.
Lower costs do not automatically mean lower fees
Even if agents reduce delivery costs, that doesn't automatically determine what clients pay. Audit fees are shaped by scope and complexity, the effort required, the firm's pricing model, and the value of the service - not simply hours recorded. KPMG has acknowledged the potential for time savings, but it hasn't said those efficiencies will translate into lower fees.
There's also the question of what firms do with the capacity that automation frees up. They could cut hours, examine more data, add review where it's most needed, or redeploy staff to other workstreams. Each choice affects client value and firm economics differently. The durable argument isn't "fewer people at lower cost" by default. It's whether the audit becomes more reliable, more timely, or more thorough for the resources used.
What are the limits and risks of AI-led auditing?
The clearest warning from the trial is that agent capability doesn't remove the need for professional oversight. During the experiment, an auditor had to intervene when an agent didn't correctly match regulatory requirements to the company's documentation. That's not a trivial problem. Audits depend on applying requirements to specific evidence, and a persuasive but wrong match can steer a team in the wrong direction.
AI systems can also generate inaccurate information, miss relevant context, or produce weak conclusions with an air of certainty. When a workflow involves multiple agents, it can be harder to spot where an error entered the chain. If a later agent treats an earlier mistake as a verified fact, a small problem can turn into a documented conclusion that still lacks support.
Data quality is a practical constraint
KPMG has pointed to data quality as a major obstacle. Some companies don't keep information in a clean, structured form that agents can use effectively. Data can be scattered across systems, labeled inconsistently, or missing enough context to interpret transactions reliably.
That creates an unglamorous but real implementation requirement: the quality of an agent's work depends partly on the quality of the information it can access. Firms shouldn't treat data preparation as a one-off technical step. They need to define what records are in scope, how data is checked, how changes are tracked, and what the system should do when evidence is missing or contradictory.
Independence, security and accountability remain human responsibilities
The UK's Financial Reporting Council has warned that AI use must not undermine auditor independence, integrity, or professional skepticism. Those principles affect how teams use outputs, how they challenge management's claims, and whether they allow technology to replace independent evaluation.
Security is another operational concern. Audit work can involve sensitive financial and commercial information. Firms need clear controls over access, retention, and the systems that process information. A tool that improves analysis but creates an uncontrolled path for confidential data is not a good trade-off.
Most importantly, the responsibility for an audit opinion stays with the human auditor. Agents can support analysis, but they don't get professional accountability for the conclusion. Reviewers need enough evidence to understand what the agent did, what information it used, where uncertainty remains, and why its output was accepted or rejected. Human sign-off only matters if the human can inspect and challenge the work - not if approval becomes a checkbox over a black box.
How should audit firms adopt AI agents safely?
The practical question isn't whether to automate an entire audit overnight. It's which tasks fit agents, what evidence shows the system is working, and how you stop or contain it when it isn't. A staged rollout forces those decisions early and limits the cost of learning.
1. Select a bounded task with a clear outcome
Start with a process that has defined inputs, a measurable output, and an existing human review method. Don't begin with a vague mandate like "audit this company." Instead, pick a specific task - collect evidence for a defined procedure, compare records against a documented rule, or draft a list of exceptions for review.
Make the task valuable enough to justify a pilot, but narrow enough to evaluate. If the agent works, the team should be able to point to what improved: turnaround time, coverage, consistency, or the quality of issues identified.
2. Establish a human baseline
Before using an agent, document how the task is done today: how long it takes, what kinds of errors or rework show up, what evidence gets gathered, and what decisions require judgment. Without a baseline, a faster workflow can look successful even if it creates more review work - or misses important exceptions.
Compare agent-supported work with a human-controlled process on the same task and standards. Don't judge success only by how many documents were processed or alerts were generated. Judge whether the evidence is relevant, the reasoning can be traced, and reviewers can resolve findings efficiently.
3. Make evidence and escalation part of the design
For each material conclusion, the system should preserve the underlying source information and the steps used to produce its output. Reviewers need to tell the difference between facts, summaries, recommendations, and assumptions. The workflow also needs a clear plan for what happens when evidence is missing, contradictory, or outside the agent's scope.
A sensible escalation rule is straightforward: uncertainty, an unsupported conclusion, or a possible regulatory mismatch should go to a qualified human - not to another agent that might repeat the same error. Human review should focus on meaningful challenge, not on rubber-stamping a queue of machine-generated answers.
4. Test failure modes before expanding
A pilot should test more than success stories. Include messy inputs: inconsistent data, incomplete documentation, unusual transactions, and cases where the relevant rule is hard to apply. Track false positives, missed issues, rework, and the time reviewers spend correcting outputs.
Set thresholds for stopping or narrowing the use case. If performance drops when the data changes, or if the work can't be reproduced, don't expand. This matters even more when you move from a controlled trial to multiple audits with different clients, different systems, and different business models.
5. Scale only when the operating model holds
Before scaling up, decide who owns system configuration, data access, review standards, incident handling, and ongoing monitoring. Training matters too: auditors need to understand both how to use the tool and how it can fail. Firms should also check whether the process changes staffing, supervision, and engagement economics in ways that preserve audit quality.
The key question isn't "Can the agent do the task?" It's "Can the firm rely on the output, explain how it was produced, and correct it when it fails?" If the answer isn't clear, the use case isn't ready for broader deployment.
Conclusion
KPMG's trial matters because it pushes the conversation beyond "AI as a document assistant." Agents were tested across connected audit tasks, and the firm says they sometimes identified deeper questions than the human team. That's real capability - but it's not proof that autonomous audits are ready for everyday use.
The next phase shouldn't be judged on speed alone. Firms need to show that agents improve quality or coverage, that their outputs can be traced back to evidence, and that human auditors can spot and correct failures. Data readiness, security, and professional accountability aren't add-ons to solve after deployment - they're part of the system from the start.
For audit firms, the sensible path is selective adoption: automate bounded tasks, measure the full workflow, and expand only when the control model holds up. The question isn't whether agents can contribute to an audit. It's which parts they can do reliably, under what safeguards, and with measurable value for clients and the public.