KPMG has used more than 100 AI agents to examine the financial statements of an unnamed FTSE 250 company in an experimental audit, according to reporting published on 18 November 2026. The agents worked alongside a conventional audit team and, in some cases, investigated issues more deeply or asked questions that human auditors had not raised. The exercise also exposed a clear limitation: a human auditor had to intervene when an agent could not match a regulatory requirement to the company’s documents. KPMG did not rely on the experiment to formally approve the accounts. The lesson for business leaders is not that AI can sign off an audit. It is that high-control workflows can be redesigned around multiple agents only when data, evidence, exception handling and human accountability are designed into the operating model.
Decision-maker takeaway
When automating a regulated or financially material process, treat data quality and control design as prerequisites. Define what the system can inspect, what it can conclude, where it must stop and which person remains accountable. A fast workflow without traceable evidence is not an operational improvement.
What KPMG’s pilot did and did not demonstrate
The reported project, known as Project Exponential, used more than 100 agents across stages of an audit rather than applying one tool to a single task. The agents collected information, analysed documents, recorded findings and prepared draft observations. They could work with one another and suggest follow-up steps. KPMG’s UK head of audit said the agents sometimes went deeper into issues than the human team. The trial was conducted in parallel with a conventional audit and was not used for the formal audit opinion.
The human intervention is as important as the headline result. Business Standard’s account of the Financial Times reporting says an auditor had to step in when an agent failed to match a set of regulations with company documents. Staff set objectives, provided company information and prior audit context, reviewed the agents’ work and challenged their conclusions. KPMG plans to extend the work to as many as 20 larger audits, but the reporting also notes that the technology is not yet ready for thousands of engagements and that some businesses do not have sufficiently clean data.
That combination of promise and limitation is typical of complex automation. A system can be excellent at searching, comparing and generating questions while still struggling with ambiguous documents, unusual circumstances or conflicting rules. The more important the decision, the more dangerous it is to confuse fluent output with verified evidence. A pilot should therefore be evaluated on the quality of the complete workflow, including the cases where the system cannot proceed.
Why the operating model matters more than the number of agents
Deploying many agents does not automatically create a better process. If each agent works from inconsistent data, uses a different interpretation of a policy or passes incomplete information to the next agent, the system can multiply errors as quickly as it multiplies tasks. The business needs a defined process map, an authoritative source for each data element and a clear record of how one step hands work to another.
For an audit workflow, that means connecting source documents, accounting records, control descriptions, transaction data, risk assessments and regulatory criteria. For a finance team, the same principle applies to supplier invoices, purchase orders, approvals, receipts and ledger entries. For an insurer, it may apply to policy documents, claims evidence and underwriting rules. The system must know which version of a document is authoritative and whether a piece of evidence is complete enough to support a conclusion.
Process owners should identify the decisions that can be standardised and the exceptions that require judgement. A routine document classification may be automated. A conclusion about the application of a complex regulation may need a specialist reviewer. A payment release, a statutory audit opinion or a safety-critical operational decision should have explicit approval requirements. The objective is to reduce avoidable work while making high-consequence decisions more transparent.
Data quality is a production dependency
AI agents cannot reliably reconcile records that are incomplete, duplicated or poorly labelled. Before a pilot, check whether documents are searchable, whether entity and supplier identifiers are consistent, whether dates and currencies are structured, and whether the source system preserves the history of changes. Establish ownership for master data and a process for correcting errors. If a team cannot identify which record is authoritative, the agent will inherit the ambiguity.
Data quality should be measured in operational terms. Track missing fields, duplicate records, unmatched transactions, stale supplier details, inconsistent document versions and the number of manual corrections required before a case can proceed. Do not assume that an AI model will clean a data estate as a side effect of being deployed. The model may identify anomalies, but the organisation still needs a responsible owner to correct the underlying system.
Also consider access. Agents should receive only the information required for their task. Sensitive employee, customer and financial records should not be exposed to every component in a multi-agent workflow. Use service identities, scoped permissions, encryption, retention controls and monitoring. The business should be able to identify which agent accessed which document and why. If a supplier changes the model, hosting location or subprocessor, the organisation should know whether a new risk assessment or approval is required.
Design an evidence trail that a person can challenge
Every material conclusion should be traceable to its source. Record the document identifier, page or data field, time of access, model and agent version, prompt or task instruction, intermediate result, rule applied and reviewer decision. If the system proposes a follow-up test, retain the reason for that proposal and the evidence that closed it. This makes it possible to reproduce a result, investigate an error and explain the process to an auditor, regulator or board.
Traceability should not mean storing every piece of information forever. Set retention periods based on legal, regulatory and business requirements. Protect the logs against unauthorised changes and make sure they can be exported in a usable format. The organisation should be able to reconstruct a case even if it changes software providers. A black-box system that cannot provide the evidence behind a material recommendation creates a dependency that may be unacceptable.
Human review must be substantive. A reviewer should be able to inspect the source evidence, understand the system’s reasoning and disagree with the output. A simple approval button is not enough if staff are under pressure to accept whatever the agent produces. Give reviewers time, training and authority to challenge results. Monitor whether override rates are unusually low, because a near-zero override rate may reflect excellent performance or automation bias.
Regulatory guidance is already shaping the design
The Financial Reporting Council published new Ethics and Technology guidance on 16 November 2026 for audit firms using technology, including AI. It focuses on maintaining auditor independence when firms communicate insights generated through technology to audited entities or give those entities access to tools that produce insights. The FRC says technology can improve audit quality and provide useful information, but sharing certain insights can create threats to integrity, objectivity or independence if the work falls outside the financial statement audit.
The FRC’s earlier Generative and Agentic AI guidance, published on 30 March 2026, explains risks and mitigations for audit tools and makes clear that regulatory accountability for the use of AI and the quality of audit outputs remains with the firm and human auditor. These documents are directly relevant to audit firms, but the underlying principles matter to any business using agents in a controlled process: technology does not transfer accountability away from the organisation.
The Centre for Public Interest Audit’s 2026 Audit Trust Index also found that 23% of audit committee chairs surveyed had limited confidence in how AI was used in audits of their companies’ financial statements. That does not mean most stakeholders reject AI. The same index reported high overall satisfaction with audit quality. It does show that confidence depends on transparency about which tools are used, how outputs are checked and who remains responsible.
Translate the lessons into a business process
Start by selecting a process with a clear objective and a measurable baseline. Examples include supplier onboarding, invoice matching, contract review, compliance evidence collection or customer-claim triage. Document the current steps, systems, hand-offs, approval points and exceptions. Identify where employees spend time searching for information, copying data or checking repetitive fields. Then separate the work into retrieval, analysis, recommendation and execution. Each stage can have a different risk level and permission set.
Run the first version in shadow mode. The agent prepares a result while the existing process continues unchanged. Compare the output with the decisions made by experienced staff. Review both false positives and missed issues. Test difficult cases deliberately: conflicting documents, incomplete records, unusual dates, policy changes, duplicate identifiers and instructions embedded in untrusted files. Ask whether the system can recognise when it lacks enough evidence and stop rather than invent a plausible answer.
Only then move to controlled execution. Give the agent narrow permissions and set thresholds for human approval. For example, it may create a draft record or propose a supplier classification but not approve a bank-account change. It may identify a potential control weakness but not close the finding. The organisation should have a kill switch, a manual fallback and a named incident owner. If the system behaves unexpectedly, the business must be able to stop the workflow without losing the underlying records.
Measure quality, speed and resilience together
Efficiency measures should include end-to-end cycle time, handling time, cases completed, backlog, rework and cost per completed case. Quality measures should include error rate, missed exceptions, evidence completeness, reviewer override, repeat findings and customer or supplier impact. Control measures should include access violations, unapproved actions, audit-log completeness, incidents, time to revoke permissions and recovery performance. No single metric is enough. A faster process that produces more errors can increase total cost and regulatory exposure.
Measure the workload that moves to human reviewers. If agents handle routine work but create a larger queue of complex exceptions, the business may need different skills or staffing levels. Review whether employees have been trained to interrogate outputs rather than simply process them. Update job descriptions and performance measures so that professional scepticism, escalation quality and evidence review are recognised as core work.
For finance and operations leaders, the cost model should include licences, integration, data preparation, model monitoring, security, training, audit support and the cost of maintaining a manual route. The value may come from better coverage, faster identification of unusual transactions or more consistent documentation rather than a reduction in headcount. Those benefits should be measured honestly and compared with the investment required to sustain the system.
A practical control checklist before scaling
Before expanding an agentic workflow, confirm that the business has a process owner, a technical owner and a risk owner. Verify the data sources, permissions and retention settings. Test the system against a documented set of ordinary and difficult cases. Confirm that every material action can be traced to evidence and a responsible person. Check that a human can challenge the result, pause the workflow and restore the previous state. Review the vendor contract for model changes, subprocessors, incident notification, data use, service availability and exit support.
Ask internal audit or an independent reviewer to sample the system’s work. Look for patterns that the development team may overlook, including over-reliance on one document type, systematic errors in a particular business unit or a tendency to treat missing data as a negative finding. Set a regular review cycle for model performance and control effectiveness. The decision to scale should be documented and based on evidence, not on the fact that the pilot has already consumed time and budget.
KPMG’s reported pilot shows that agents can contribute meaningful analytical work in a complex process, but it also shows why human judgement and data readiness remain essential. Businesses should take the same approach to their own workflows. Automate the repeatable parts, expose the uncertain parts, preserve the evidence and keep accountability with named people. That is how a promising demonstration becomes a dependable operating system.
Frequently asked questions
Did KPMG use AI agents to sign off a FTSE 250 company’s accounts?
No. The reported exercise ran alongside a conventional audit and was not relied on for formal approval of the accounts. KPMG’s human auditors reviewed and challenged the agents’ work, and a human had to intervene when an agent could not match regulatory requirements to company documents.
What should a company prepare before deploying AI agents in a controlled workflow?
Prepare reliable source data, process maps, access controls, evidence logging, exception rules, named owners and a tested manual fallback. Run the system in shadow mode first and compare its output with experienced staff decisions before allowing it to execute actions.
Who remains accountable when an AI agent performs audit or compliance work?
The organisation and the relevant professional remain accountable under the applicable standards and rules. The FRC’s guidance makes clear that using AI does not transfer responsibility for audit quality or professional judgement to the tool provider.
Sources and further reading
- Financial Times: AI asked better questions than KPMG accountants in pilot audit, 18 November 2026
- Business Standard: KPMG’s AI audit pilot and human intervention
- Financial Reporting Council: Ethics and Technology guidance, 16 November 2026
- Financial Reporting Council: Generative and Agentic AI guidance, 30 March 2026
- Centre for Public Interest Audit: 2026 Audit Trust Index



