AI is already in the internal audit plan. What is less clear is whether it is well covered. Risk in Focus 2027, the survey the European institutes of internal auditors published in September with 797 responses from chief audit executives gathered between March and April 2026, puts the two figures side by side.
The category of digital disruption, new technology and AI climbs from third place to second, with 52% of CAEs placing it among their top five risks. It is also the area where audit effort grew most. 57% say they increased it, "the largest growth area in the survey." And yet CAEs themselves give internal audit's coverage of AI a low adequacy rating.
More hours is not more coverage
Given that AI risk management still lags behind, the report recommends advisory, capability-building and governance-focused approaches rather than pure assurance. That makes sense. But advising on a framework does not tell you whether systems in production behave the way the framework promises.
Outside Europe the pattern looks similar. In North America, the IIA's 2025 Pulse of Internal Audit, with 405 responses gathered in October and November 2024, finds that 41% of functions use generative AI in their own work. Only 15% provide assurance over their organisation's use of it, and 25% provide advisory services.
There is a second reading the Risk in Focus report itself points to. Some CAEs do not audit AI in a dedicated engagement but inside others, such as data governance or cybersecurity, and one describes carrying out a "chain of audits" to cover the connected risks. The report concludes that internal audit support may be higher than the survey shows. If that is your case, the coverage exists but cannot be seen. And what cannot be seen in the plan cannot be defended in front of the committee.
What changes with agents
The report captures the shift in the words of a CAE from the technology sector. "These programs started with simple prompting but now we are moving towards the execution of tasks, which opens up a whole different range of risks." Working out what it means for an autonomous agent to act as if it were a team member is what determines the access permissions it needs.
The report also warns about one very concrete point. AI applications integrated into corporate systems can scan, locate and retrieve documents across the whole enterprise, but access permissions were not built to handle that capability.
What to put in the plan
- An agreed definition of adequate coverage. The report asks CAEs to discuss with stakeholders what adequate AI coverage means and how it is reflected in a risk-based plan. Without that definition, a low adequacy rating cannot be fixed because there is nothing to measure it against.
- The AI already audited inside other engagements, made visible. Mark in the plan which AI tests travel inside data governance, cybersecurity or third-party work. Whatever falls outside all of them is the real gap.
- Permissions against search capability. Using an ordinary user's credentials, which sensitive documents can the corporate AI assistant reach. Payroll and human resources data are the examples the report cites.
- Decisions that must be explainable. Which decisions about customers or employees a model supports, and whether anyone could explain each one to a third party. The report asks for assurance that such decisions are not left to opaque models.
Spending more hours on AI does not mean having it covered. That difference is what the plan has to show.
None of this replaces the foundation, which is knowing which agents and AI uses exist and what limits they have, as we covered when looking at governing AI agents. What it adds is the question the committee will end up asking. Of all the hours spent on AI, how many give assurance on how it actually works.
