AI agents can take financial testing further and raise questions that professionals have missed. But the lesson from KPMG’s trial is not to remove auditors from the process. It is to redesign their work. For enterprises, the opportunity is to expand control coverage without a proportional increase in working hours, while preserving professional accountability and evidence that can withstand scrutiny.
That distinction matters. A system that formulates an impressive question is not yet a system on which an audit opinion can rest. Between the two sit data, business context, quality checks and a mechanism for recognizing when to stop and involve a person.
What KPMG tested, and what remains unproven
According to the Financial Times, KPMG deployed more than 100 AI agents in an experimental audit of a British FTSE 250 company. Working alongside the regular audit team under Project Exponential, the agents gathered information, analyzed documents, documented work and drafted conclusions. They also collaborated with one another and proposed follow-up actions.
- 100+ AI agents participated in the trial.
- 4 months were spent developing the agents.
- Up to 20 additional audits were included in the planned expansion.
These numbers describe an experiment across a substantial workflow, not simply an assistant that summarizes a document. Still, the expansion is a future plan. The trial does not establish that the system is ready to perform thousands of audits.
Cath Burnet, KPMG’s head of audit in the UK, said that in some cases the agents went deeper and asked better questions. That is a qualitative assessment from the firm, not a published comparative measure demonstrating overall superiority to human auditors.
More importantly, although an audit partner signed off on the experimental exercise, KPMG did not rely on its results when formally signing off on the company’s accounts. In one instance, an accountant had to intervene because an agent could not reconcile a set of regulations with the company’s documents. That failure is particularly instructive: it shows what a dependable professional system must be designed to handle.
The value is not just the answer, but the question nobody asked
In financial processes, errors do not always begin with faulty calculations. Sometimes the problem is an untested assumption, a document nobody requested or a contradiction nobody connected across sources.
An agent could help examine relationships between contracts, accounting records and correspondence, then suggest further lines of inquiry. These are potential enterprise applications, not specific findings reported from the trial. Their value depends on distinguishing a useful suspicion from a persuasive explanation unsupported by evidence.
The management question is not whether an agent sounds like an auditor. It is whether the agent produces work that can be reproduced, challenged and used to support a decision.
This is also why domain expertise matters. Defining a test requires an understanding of accounting, the business process and the limitations of the models. Relevant education and practical experience are not decorative credentials on a development team. They are part of its quality-control mechanism. Familiarity with automation tools alone is not enough to design an audit process.
Human oversight does not mean approving every line
It is easy to promise that a person will review every output. Building an economically viable operating model around that promise is harder.
If an employee must reread every document and repeat every test, the organization may gain another layer of work rather than save time. Blanket approval of agent outputs, on the other hand, creates dangerous dependence on a system whose outputs are not deterministic.
The answer is risk-based oversight, governed by rules set by professionals. Someone who previously performed an individual test needs tools to supervise many workpapers, not an inbox overflowing with approval requests.
- Findings: A conclusion without supporting evidence is weak control; the recommended design links findings to evidence and document versions.
- Human intervention: Requiring approval for every action creates friction; focus professional attention on exceptions and high-risk cases.
- Permissions: Broad information access is weak control; assign access according to the task.
- Quality: Polished wording is not a quality measure; use testing, sampling and reproducibility checks.
- Accountability: Ambiguity between vendors leaves gaps; designate a professional owner.
These are design recommendations, not a description of KPMG’s controls. In practice, escalation needs clear triggers: missing evidence, conflicting sources, regulatory ambiguity or an action with a material impact. Routine work also needs sampling. An agent’s statement that it is confident is not a substitute for control.
Watch out: Traceability is not proof of correctness. A complete record makes it possible to inspect how a conclusion was reached. It does not prove the conclusion is right. A wrong conclusion can be thoroughly documented.
KPMG describes detailed records of the agents’ interactions with documents. That is important infrastructure for investigation and review. It does not eliminate the need to verify that the document is appropriate, the relevant rule was selected and the reasoning is valid.
Who gets the savings, and who pays for oversight?
KPMG identified opportunities to save time but did not commit to reducing fees. CFOs and audit committees should not assume that automation at a service provider will automatically translate into lower prices for clients.
Discussions with professional services providers should separate three types of value: shorter turnaround times, broader testing coverage and better-quality findings. The price may stay the same while more transactions are examined. Preparation hours may fall while specialists spend more time resolving exceptions. Measure the net outcome.
For internal deployments, model costs are only one component. Data preparation, system integration, permissions, maintenance, quality checks and professional review time all belong in the calculation. KPMG itself noted that data quality at some companies is insufficient for agents to work efficiently.
There is also the question of junior staff development. If agents take over some foundational tasks, organizations will need another route for employees to gain experience. They cannot expect expert judgment without giving people a way to develop it.
How to start without gambling on accountability
A full financial statement audit does not have to be the starting point. Choose a bounded internal process, such as preparing a reconciliation workpaper or gathering supporting evidence, and compare the system’s performance with the existing process before relying on it.
- Choose a defined process: Establish an accountable owner, operating boundaries and the intended business outcome.
- Prepare evidence and permissions: Check data quality, document versions and need-based access.
- Run in parallel: Compare performance with the existing process without relying on the agent for the decision that carries formal responsibility.
- Evaluate quality and cost: Measure total time, missed findings, false positives and the human review burden.
- Expand based on results: Increase autonomy only when quality and control measures justify it.
This approach requires progress on two fronts: AI literacy among employees and organizational capability to build and manage agents. Employees need to know how to request evidence and challenge conclusions. IT systems must support permissions, monitoring, version control and the ability to stop activity.
An agent embedded in an existing system may require less change to working habits than a tool that asks every employee to start a new conversation each time. It still requires operational discipline behind the scenes.
Regulatory caution is not a side issue. The UK’s Financial Reporting Council emphasized that responsibility remains with the auditor and warned of potential threats to independence when insights extend beyond the boundaries of the audit. That is not a universal regulatory instruction, but it is a principle leaders everywhere should take seriously: using technology does not transfer accountability to it.
The lesson from KPMG’s trial is not to buy 100 agents. It is to build the professional capability to operate them, test them and recognize when not to rely on them. That is the difference between an impressive demonstration and an operational improvement an organization can stand behind.
