Automation bias is a tendency to favour output from an automated system even when human reasoning or contradictory information questions it, and an AI tool that reads a set of accounts first creates exactly that risk for the reviewer who reads them second. This article is for partners and managers who sign accounts that an AI tool has been used on. It sets out what the professional standards say about scepticism, how far the FRC’s audit guidance on AI reaches into accounts work that is not an audit, and what a reviewer should record on the file.
What automation bias is and where the Code deals with it
The ICAEW Code of Ethics, which applies from 1 July 2026, lists automation bias among the examples of bias a professional accountant should watch for. It describes it as “a tendency to favour output generated from automated systems”, even when human reasoning or contradictory information raises questions about whether the output is reliable or fit for purpose (paragraph 120.12 A2). The Code suggests consulting others to ensure appropriate challenge, and receiving training in identifying bias, as actions that might mitigate it (paragraph 120.12 A3).
Two other parts of the Code bear on a review. Paragraph 120.5 A1 says an inquiring mind involves considering the source, relevance and sufficiency of the information obtained, and the Code says that requirement applies to all accountants regardless of the professional activity undertaken. Paragraph R320.11 requires an accountant who intends to use the output of technology to determine whether that use is appropriate for the intended purpose. Among the factors in 320.11 A1 are whether the technology has been appropriately tested and evaluated, and the extent of reliance on its output.
Our guide to signing off accounts that an AI tool has reviewed covers the wider question of responsibility. This article stays with the narrower one of what the file should show.
How far the FRC’s audit guidance reaches
On 26 June 2025 the FRC published “AI in audit: illustrative example and documentation guidance”. It says that the material is not prescriptive, and that the requirements against which firms are assessed remain only those in the ISQMs and ISAs (UK). For automated tools and techniques that use AI, it says documentation will be required on the audit file, citing paragraph 8 of ISA (UK) 230. It also covers other tools that use AI, with examples that include tools that can review work to identify omissions or inconsistencies. For those, it says there may be no requirement to document their use on the audit file, although a team may choose to if that would help a reviewer understand the work performed.
That guidance is written for audits. Preparing or reviewing accounts in a non-audit engagement is not governed by ISA (UK) 230. We treat the guidance as a yardstick, applied by analogy. The case for doing so rests on the Code. An inquiring mind applies to every accountant, and the Code notes that auditing, review and other assurance standards require professional scepticism. A practice that signs accounts it has not audited still has to decide how it will show that its reviewer did not simply accept what a tool said.
The FRC’s March 2026 guidance on generative and agentic AI is aimed at the central technical teams within audit firms, so it is a further step removed. It is still useful on one point. It says review of a tool’s output should evaluate whether the output is complete, accurate, internally consistent and consistent with the reviewer’s understanding of the entity, and that the reviewer should apply professional scepticism and be aware of the risk of automation bias. It adds that review may involve forming an independent view to compare the output against, and that a tool may sound fluent and confident even when it is wrong.
What a reviewer records on the file
ISA (UK) 230 paragraph 8 asks for documentation sufficient to enable an experienced auditor, having no previous connection with the audit, to understand the work. Paragraph 9(c) asks for a record of who reviewed the work and the date and extent of the review, and paragraph 11 asks the auditor to document how any information inconsistent with the final conclusion on a significant matter was addressed. Those are audit requirements. They also describe a sensible test for any file: could someone with no prior knowledge of it see what was checked, by whom, and what happened to each point?
The FRC’s list of what it may be appropriate to document on the audit file for an automated tool or technique that uses AI includes the version number of the tool and model used, the data or prompts that were input, and how the team used the outputs to conclude. The list is guidance for audit files and is not a requirement for accounts work. Using those headings as a starting point, we would expect a reviewer’s note to carry the following.
| Record | What it shows | Anchor |
|---|---|---|
| Tool and version | Which tool produced the output, at which version | FRC AI in audit, audit file list |
| Inputs | The documents, data or prompts given to the tool | FRC AI in audit, audit file list |
| Points raised | What the tool reported, and what it was silent on | Our expectation |
| Independent check | Which source documents the reviewer read directly | FRC March 2026, independent view |
| Outcome of each point | Agreed, corrected or rejected, with the reason | ISA (UK) 230 paragraph 11, by analogy |
| Reviewer, date, extent | Who reviewed, when, and how far they went | ISA (UK) 230 paragraph 9(c), by analogy |
The FRC also recognises proportionality. In its example of a language model tool used to research an audited entity, it says that where the tool’s outputs are all checked back to the cited sources, it may be appropriate to document less about the tool, because the corroboration limits the relevance of the fact that an AI-enabled tool was used. The FRC says this about audit files. For accounts work we would apply the same idea with care: the more of the tool’s output the reviewer has checked against source documents, the more the note can focus on what was checked.
A worked example
This example is hypothetical. A tool reads a draft set of small company accounts and reports that the directors’ loan note agrees to the loan account. The manager does not treat that as a cleared point. They open the ledger, find that the director owed the company £18,450 at the year end, and see that the note discloses £14,850.
The file note records the tool and version, the documents it was given, and the statement the tool made. It records that the manager agreed the note to the ledger loan account, found a difference of £3,600, and returned the note to the preparer, who corrected it. It records the date the manager cleared the point, and that the related party note was then also read against the ledger, because a statement the tool had made with apparent confidence turned out to be wrong. The note makes no claim about the tool in general. It records what this reviewer checked, what they found, and what they did next, which is what a later reader needs.
Habits that reduce the risk, and who answers for the accounts
- Read the source before the summary. Where practicable, form your own view of a point and then compare it with the tool’s, as the FRC describes.
- Choose reviewers who can spot a deficient output. The FRC says those reviewing outputs need the appropriate competence to identify potential deficiencies.
- Train for the bias itself. The FRC’s central documentation list includes training material with strategies to mitigate automation bias, and the Code lists training on identifying bias as a mitigation.
- Invite challenge. A second pair of eyes on the points a tool cleared is the Code’s consultation safeguard in practice.
- Treat AI as a starting point for research. An ICAEW practice news article of 7 September 2026 says AI should never be treated as an authoritative source.
None of this moves responsibility. Under section 393(1) of the Companies Act 2006 the directors must not approve accounts unless they are satisfied that they give a true and fair view, and under section 414 the accounts must be approved by the board and signed by a director. The judgement calls and the signature stay with people. Our guide to file review and evidence of review shows how this record fits into a practice’s quality management.
Questions practices ask
Does ISA (UK) 230 apply to accounts preparation?
No. It is an auditing standard and applies to audits. The FRC’s AI guidance cites it for tools used on audits. For accounts work that is not an audit, we use its tests as a yardstick: who reviewed, when, how far, and how each inconsistency was dealt with.
Do we have to record every prompt?
The FRC’s list for audit files mentions the data or prompts input, where applicable, and says professional judgement decides what is appropriate. For accounts work no standard sets the rule. We would record enough that a reader can see what the tool was given.
Does the Code require us to guard against automation bias?
The Code names it as an example of bias to be aware of when exercising professional judgement (120.12 A2). The mitigations in 120.12 A3 are actions that might help, so the Code leaves the choice of safeguard to the accountant.
Does using AI change who is responsible for the accounts?
No. The Act requires the board to approve the accounts and a director to sign them (section 414), and section 393(1) puts the true and fair view on the directors. The tool is one input to the reviewer’s work.
Sources
- AI in audit: illustrative example and documentation guidance, Financial Reporting Council, June 2025 (news release dated 26 June 2025)
- FRC publishes landmark guidance providing clarity to audit profession on the uses of AI, Financial Reporting Council, 26 June 2025
- ISA (UK) 230 (Revised June 2016), Audit Documentation, Financial Reporting Council, updated September 2025
- Generative and Agentic AI Guidance: risks, mitigations and illustrative examples, Financial Reporting Council, March 2026
- ICAEW Code of Ethics 2026, ICAEW, applies from 1 July 2026
- Using AI with confidence: how to validate AI content, ICAEW, 7 September 2026
- Companies Act 2006, section 393, legislation.gov.uk, read 21 November 2026
- Companies Act 2006, section 414, legislation.gov.uk, read 21 November 2026