Julien Gruhier, head of AI at Croner Intelligence, explains the FRC guidance on potential risks for accountants and tax professionals when using AI tools
The Financial Reporting Council (FRC) guidance on generative and agentic AI is written for auditors, focusing on risks and best practice, but the advice applies to anyone who relies on AI output: from accountants to tax professionals. Most useful is the FRC’s list of the five ways AI responses can be deficient.
The five risk areas
The FRC guidance clearly sets out the five AI risk areas to watch out for:
- Hallucinations are information the model has fabricated.
- Omissions are information that should be there and is not.
- Distortions misrepresent the meaning, emphasis or implications of something real.
- Faulty reasoning produces unsupported or illogical conclusions.
- Inconsistencies are outputs that contradict themselves, or an earlier output, without justification.
Only the first is easily picked up by checking that a source reference actually exists. The other four can only be identified by comparing the AI output against the primary source material.
An AI summary’s biggest risk is what it leaves out
The FRC’s first illustrative example is an AI summary of board minutes.
The main risk it identifies is not invention by the AI, but an auditor treating the output as a complete account of the period.
The mitigations are instructive: split the minutes into indexed sections so every point can be cited, build the summary in separate steps (each meeting, then the period, then a summary), and advises telling the AI tool to include matters when unsure, in other words do not let the AI control the process.
Automation and importance of reviews
The second example, a contract review for revenue testing, separates extraction from evaluation, supplies curated technical material through a retrieval component, and has the tool escalate to a human when it is not confident.
The FRC recommends firms always train staff about automation bias and warns users to watch out for the tendency of AI tools to sound fluent and confident even when they are wrong.
What this means beyond audit
Those design choices are not peculiar to audit. Croner Intelligence follows the same principles in tax and accounting research. Answers are only drawn from retrieved Croner and HMRC material, references are cited to their sources, and where facts are missing the AI tool asks rather than assumes. That reduces the room for hallucination and makes omissions and distortions easier to spot. It does not remove the human reviewer.
The FRC is clear that accountability for AI output stays with humans, in other words the auditors and accountants actually using the tools. That is the important lesson to take away from any use of AI, particularly in a professional workplace where accuracy and accountability are crucial.
*** Find out more about Croner Intelligence, the game changing AI tool for accountants and tax professionals. Book a demo here ***
