Two earlier pieces in this series established what good AI agent governance looks like in practice: AI Governance in Finance: Framework vs Audit Trail argued the case for evidence — being able to reconstruct what an agent actually did — and AI Agent Internal Controls: Access Is the Risk argued the case for access — constraining what an agent was ever permitted to do in the first place. Both assumed the problem was knowledge. This piece argues something less comfortable: the problem is incentives, and it belongs to the CFO and the audit committee.
What does the survey data actually show?
The data shows a governance gap that is structurally rational, not accidental. A June 2026 survey of 1,505 CFOs and senior finance leaders across the UK, US, Australia and India — commissioned by a tax compliance software vendor, conducted by Censuswide (an MRS and British Polling Council member), with fieldwork running 15–22 June 2026 — screened every respondent as someone who had already deployed, piloted or evaluated AI agents in financial processes. That screening matters: nothing in this survey can say how widely AI agents are used across finance functions in general. It speaks only to those already in the game. Bear that in mind before treating any figure below as a sector-wide benchmark; the sample skewed to the UK and US and covered companies with $10m or more in revenue.
With that caveat stated, the numbers are directionally striking. Among finance leaders who have already deployed or evaluated agents: 92% report moderate or significant career pressure to demonstrate that AI agent investments are delivering value. Of those, 71% say that pressure is focused specifically on deployment speed. Only 7% say they prioritise governance over speed.
Now read those figures alongside the governance results. Thirty per cent of finance leaders who have already deployed or evaluated agents have not updated their internal controls within the last year. Forty-four per cent say they are only somewhat confident they could explain agent actions to auditors or regulators. Twenty-three per cent say accountability for a significant AI error is unclear or sits with nobody. Seventy-six per cent lack dedicated in-house AI expertise.
Is this a competence problem or a knowledge problem?
It is neither. The 44% of finance leaders who have already deployed or evaluated agents and cannot confidently explain those agents to an auditor are not ignorant of the risk — they are responding rationally to what they are measured on. Speed of deployment is rewarded. Explainability is not. Controls lag deployment because the appraisal structure does not reward closing the gap. That is not a technology failure and it is not a capability failure. It is an incentive design failure.
The standard framing treats the governance numbers and the incentive numbers as two separate findings: here is how well-governed our deployments are; separately, here is the pressure people felt. Treat them as one finding and the picture sharpens considerably. The 71% feeling pressure focused on deployment speed are the same population producing the 30% who have not updated controls. The causal arrow runs from measurement to behaviour, not from ignorance to risk.
Who owns this problem — and why does that matter?
This is a management-control problem owned by the CFO and the audit committee, not a technology problem owned by IT. That distinction matters for what you do next. If the root cause were technology — agents that are inherently difficult to audit — the solution is better tooling. If the root cause were knowledge — finance teams that do not understand what agents do — the solution is training. But if the root cause is that the incentive structure actively rewards speed and says nothing about explainability or control integrity, then buying better tooling and running more training will not close the gap. The incentive structure will reproduce the same outcome.
The audit committee's role in setting the tone for internal control has been well-established through decades of post-Enron governance reform. AI agent deployment is not categorically different: it is a new class of process change with control implications, and the committee's oversight remit covers it. The CFO's role is to surface the incentive conflict explicitly — not to apologise for the governance gap, but to name its structural cause and request that the appraisal framework be redesigned accordingly.
What does a redesigned appraisal framework look like?
It looks like one that measures what it currently ignores. At minimum, it rewards three things alongside deployment speed: first, the ability to produce an audit trail for agent actions in deployed processes; second, the currency of internal controls relative to what has been deployed; and third, clear documented accountability for agent errors before deployment, not after. None of these require new technology. They require that someone in the appraisal chain asks for them.
The 23% of finance leaders who have already deployed or evaluated agents and report that accountability for a significant AI error is unclear or sits with nobody is not a finding about AI. It is a finding about governance design. Accountability for errors has to be assigned before something goes wrong, or the assignment happens in a crisis — which is when it is most costly and least objective.
What is the right question for the board agenda?
The wrong question — the one that currently dominates board papers — is: are our agents governed? It is the wrong question because everyone answers yes. Every organisation has a governance framework. The right question is: what is the finance leadership team actually measured on this year, and does any of it reward being able to explain what we deployed? If the honest answer is no, the governance framework is decorative. Fix what you measure before you buy anything else.
The three pieces in this series have argued evidence, access, and now incentive. All three are necessary. None is sufficient alone. But of the three, incentive is the one that the CFO can change without a technology budget and without a vendor. It requires a conversation with the audit committee, a rewrite of one section of the performance framework, and the willingness to name the problem accurately. That is a management decision, not a procurement decision.
Common questions
Why do finance teams struggle with AI agent governance despite awareness of the risks?
A June 2026 Censuswide survey of 1,505 finance leaders who had already deployed or evaluated AI agents found that 71% face career pressure focused specifically on deployment speed, while only 7% prioritise governance over speed. The governance gap is not caused by ignorance of risk — it is caused by an appraisal structure that rewards speed and does not reward explainability or control integrity. This makes it a management-control problem, not a technology problem.
What proportion of finance leaders who have deployed AI agents cannot explain them to auditors?
According to the same June 2026 Censuswide survey — commissioned by a tax compliance software vendor, covering 1,505 respondents across the UK, US, Australia and India, all of whom had deployed, piloted or evaluated agents — 44% say they are only somewhat confident they could explain agent actions to auditors or regulators. This figure applies only to that adopter population and cannot be generalised to finance functions in general.
Who is responsible for fixing AI agent governance in a finance function?
The CFO and the audit committee own this as a management-control problem, not IT as a technology problem. Because the root cause is incentive design — appraisal frameworks that reward deployment speed but not explainability or control currency — better tooling and training will not close the gap unless the measurement framework changes first. The audit committee's oversight remit covers this class of process change.
What is the single most important governance question to put on the board agenda?
The most productive question is not 'are our agents governed?' — every organisation will answer yes. The productive question is: what is the finance leadership team actually measured on this year, and does any of it reward being able to explain what was deployed? If the answer is no, the governance framework is unlikely to constrain real behaviour, regardless of how well-written it is.
How widespread is unclear accountability for AI agent errors in finance?
The June 2026 Censuswide survey found that 23% of finance leaders who have already deployed or evaluated AI agents say accountability for a significant AI error is unclear or sits with nobody. This is a governance design finding, not a technology finding: accountability for errors must be assigned before deployment, not resolved during a crisis.