The AFM Says Outcomes From (Gen)AI Auditing Tools Must Be Verifiable or Replicable

The Netherlands' Autoriteit Financiële Markten (AFM) report 12 Building blocks for controlled use of auditing tools is dated December 2025. The review behind it concerns auditing tools "in the broadest sense, from audit-documentation software to data-analytics tools," and (Gen)AI, a term a footnote ties to the "general-purpose AI model" defined in Regulation (EU) 2024/1689. The building blocks rest on four preconditions, stacked like layers, and the top layer is for (Gen)AI. Parts of the report follow, not the whole.

The Review and the AFM's Expectations

The report comes from what the AFM calls an exploratory review of the use and scope of auditing tools in statutory audits. It was carried out at eleven audit firms with a regular license to perform statutory audits and at two whose license also extends to statutory audits of public interest entities. At each, the AFM examined the firm's quality control system and the implementation of various auditing tools in two different statutory audit files. It also held discussions with stakeholders, including suppliers of eight different auditing tools, sector organizations and professional bodies.

The building blocks "help assess what is going well and where improvement is needed," the report says. It puts questions to firms, among them which building blocks they should or could use, and says "it is up to audit firms and auditors to answer these questions, based on factors such as the nature of the individual audit or the auditing tools in use at the firm." Choices to include or exclude preconditions from review, it adds, "should be seriously considered to ensure compliance with applicable laws and regulations." The AFM expects audit firms to use the building blocks to evaluate their strengths and weaknesses and improve where it is required, and says "in the coming years the AFM will pay more attention to the controlled use of auditing tools in further supervisory reviews."

The Layers Underneath

The first three preconditions are that the audit firm maintains a robust risk management framework and effective internal controls, that the input data is relevant and reliable, and that the auditing tool is implemented in a controlled manner. The report says "the lower three layers of the structure must be firmly in place before deploying advanced technologies such as (Gen)AI."

On monitoring, the AFM observes that audit firms lack structural monitoring of the use and effect of auditing tools within their audit practice. "As a result, incorrect use, technical errors or potentially significant deficiencies may go unnoticed, negatively affecting audit quality." It observes that auditors do not always perform adequate procedures to verify the relevance and reliability of input data. And it says the external auditor should ensure the entire audit trail of data is detailed enough "for an experienced auditor to be able to fully understand the work performed."

"May go unnoticed" is where the risk sits. Without that monitoring, what would bring a tool's technical error to anyone's attention?

The Top Layer

The report adds three building blocks for (Gen)AI, which it says "introduces new, inherent risk factors due to the nature and complexity of the technology." With these tools, "we see a greater risk of overreliance compared to traditional auditing tools, due to the way outcomes are presented." Some tools mimic human traits, it says, and outputs that seem persuasive because of their structure or presentation can result in a "halo effect."

In response to a data request from the AFM, firms licensed to audit public interest entities reported that 12% of the auditing tools they plan to use in 2025 incorporate AI or generative AI. Such a tool may be fully powered by AI or include AI as one of its components.

The first of the three is that the tool is deployed safely and in a controlled manner. Because of the nature of these tools, the report says, it is more difficult to maintain control over the processing of input data and to ensure that tools "continue to operate as intended by the audit firm." It describes methods to evaluate their precision, one of which is a benchmark outcome tested regularly against specific tasks, including the results of approved prompts from a prompt library checked after model updates.

The second is that the outcomes are verifiable or replicable. "The output of (Gen)AI tools cannot always be precisely reproduced, but outcomes must be verifiable or replicable." Where output is used as audit evidence, verification can come, for example, from having the tool reference external sources or source documents. If exact verification is not feasible, the report says, audit firms may opt to make the results replicable. Audit file documentation requirements when using these tools are "equivalent to those for conventional auditing tools." Its example of bad practice is a tool that is a black box, its outcomes provided "without visible reference to specific source documentation."

Responsibility, and Where (Gen)AI Is Used

The third building block is that the external auditor is ultimately responsible for the outcomes. "(Gen)AI may facilitate the work but cannot replace professional judgment." The auditor is responsible for assessing, weighing and validating the outcomes before using them.

"Our review shows that it is not always clear or visible where and how (Gen)AI has been applied in the audit process." It is important, the report says, that the external auditor knows where it was used, "so that appropriate responsibility can be taken during the review process." Its research found that nearly all audit firms use (Gen)AI to some extent, for example chatbots to summarize meeting minutes, as a sounding board for risk assessments, or to answer simple methodological questions, and that firms are experimenting with tools such as agents that can autonomously prepare and, to some extent, perform audit procedures. As good practice it describes a tool that makes the auditor validate the outputs before proceeding to the next step, and displays reliability scores with an alert about the risks.

A review can only take responsibility for what it can see. Where a file does not show where (Gen)AI was used, the review may pass work it never weighed.

One of the questions the AFM expects audit firms to ask themselves is "Which procedures receive less human attention due to (Gen)AI tools?"