This article is stated as at 22 November 2026. FRC and ICAEW guidance on the use of AI in audit and accounts work is still developing and may change what firms are expected to record. Take advice on your own circumstances before acting.
An AI review tool that agrees with you on every point isn’t doing much reviewing, so sooner or later a partner will overrule one. Documenting professional judgement in an AI review comes down to four things on file: the facts you knew when you decided, your reasoning, any consultation, and who cleared the point and when. No FRC or ICAEW rule deals with the override itself. But both say the human professional stays accountable, and the override note is the firm’s evidence that the judgement was the partner’s own. That applies to audits directly and to compilation work by analogy.
Who is responsible when you overrule an AI review point?
Using a tool doesn’t change who is responsible. The FRC’s factsheet to its Generative and Agentic AI Guidance, published on 30 March 2026, says that “the firm and engagement partner retain full responsibility for audit quality, in accordance with ISQM (UK) 1 and ISA (UK) 220 respectively.” The accompanying news release says the human auditor is always accountable and that the guidance does not alter that position.
The PCRT guidance on the ethical use of AI tools, published on 19 January 2026 and hosted by the ICAEW, says the same for tax work. Members are ultimately responsible for the work they produce, whatever AI tools were used, and should avoid “blindly accepting the AI output”. It suggests treating that output as if a less experienced junior colleague had prepared it. That framing helps with overrides, because no manager would clear a junior’s query without saying why.
The directors carry their own responsibility. Under section 393 of the Companies Act 2006, directors must not approve accounts unless they are satisfied that the accounts give a true and fair view. Section 414 requires the board to approve the accounts and a director to sign the balance sheet. The client is responsible for the accounts, and the firm and partner are responsible for the quality of the firm’s work. A good override record covers both: it sets out the partner’s reasoning, and where the point was the client’s to resolve, it shows the point was put to the directors and what they said.
If the tool was right and the partner overruled it, the partner is accountable either way. The record decides whether a later reviewer sees a considered judgement that turned out to be wrong, or a flag that was waved through. Our guide to professional responsibility at sign off covers the wider position.
What documenting professional judgement on an AI point should show
The FRC’s 2025 AI in Audit documentation guidance gives one direct pointer. In its illustrative example, a journal testing tool flags high risk items, and the engagement team follows up each one and “documents their rationale for this conclusion”. It doesn’t say what that rationale should contain.
For that, the closest guide is the ICAEW’s material on documenting significant professional judgements in an ISA (UK) audit. It names three things it calls essential: the facts and circumstances known when the judgement was made, the reasoning behind the decision, and the results of any consultation on significant or contentious matters. It also points to ISA (UK) 200 (paragraph A30 in the current edition), which sets the test the record has to pass: an experienced auditor with no previous connection with the audit must be able to understand the significant judgements. Add two practical items to those three, who cleared the point and when, and the tool’s original wording, and the record answers most questions a later reviewer could ask.
| Element of the record | What to write | Why it matters later |
|---|---|---|
| The point as raised | The tool’s flag, unedited and kept on file | Shows what the reviewer actually saw |
| Facts known at the time | The documents, explanations and figures looked at | A judgement is assessed on what was known then |
| Reasoning | Why the point doesn’t apply, or doesn’t affect the accounts | The part a later reviewer must be able to follow |
| Consultation | Who was asked and what they concluded, if anyone | Expected on significant or contentious matters |
| Who and when | Name, role and date the point was cleared | Shows who exercised the judgement |
Take a hypothetical example, with illustrative dates. The tool flags that the turnover policy note describes a maintenance service the client appears to have stopped providing. A clearance that reads “N/A, reviewed” gives a monitoring reviewer nothing to follow. A clearance that reads “Client still invoices maintenance contracts, confirmed from the sales ledger, reviewed at B3, and with the finance director on 12 October 2026; policy wording still accurate. Cleared by [manager], 14 October 2026” gives the facts, the evidence, the reasoning and the person in two lines.
Keep the tool’s original output. If reviewers can edit or delete points in the tool, decide how the original wording is kept before the first live file, or you won’t be able to show later what the partner saw and why they disagreed. Our article on what each review point should show goes through the audit trail in more detail.
Do you have to document every AI point you dismiss?
No. The ICAEW page cites ISA 230.A7: it is “neither necessary nor practicable for the auditor to document every matter considered or professional judgement made”. The FRC’s 2025 guidance says that “professional judgement must be applied when determining what documentation is appropriate”. It also warns that over documentation can divert time and resource from areas where they would do more for audit quality.
Scale the record to what the point touches.
| Type of point being dismissed | Suggested record |
|---|---|
| Repeats a point already cleared, or is plainly mistaken on the face of the file | A few words of reason |
| A disclosure or presentation point resolved from evidence already on file | A short note with the evidence reference |
| Touches a significant judgement, a matter for the directors, or something the client must resolve | A full note: facts, reasoning, consultation, who and when, cleared by the partner |
In our view, a few words of reason should be the minimum even for the first group. The tool’s flag stays on the file, and an unexplained dismissal next to it invites the very question the file should answer. A bare “disagree” fails the ICAEW test, because it gives a reader with no connection to the job nothing to follow. A memo for every dismissed flag goes too far the other way, and it buries the points that matter.
Should a manager or the partner clear AI points on significant judgements?
Managers can sensibly clear routine points, provided the firm’s procedures say so. Points that touch significant judgements should go to the partner, and the partner’s name should be on the clearance. When junior staff clear those points without escalation, the evidence that the engagement partner took responsibility is weaker, and on an audit ISA (UK) 220 places that responsibility on the partner. Examples include going concern, the treatment of a director’s loan, or whether reserves supported a dividend (our note on how unlawful dividends arise shows how easily that goes wrong).
The FRC’s 2026 guidance notes that AI outputs at decision points may be deficient where the auditor “lacks some relevant technical knowledge or exhibits automation bias”. It defines a human in the loop as “a human who directs, authorises or reviews the system’s actions or outputs at runtime”. The PCRT guidance describes automation bias as a tendency to favour output from automated systems, which is the risk of accepting points too readily. Dismissing them too readily is the mirror image, and it is common once a team decides a tool “always flags that”. In both cases the reviewer has stopped reading the point. Daniel Clark, who chairs ICAEW’s Technology Faculty Board, has said the FRC’s material on automation bias is “buried in just a single paragraph” and that firms must train staff to apply judgement and scepticism to AI outputs. Our piece on evidencing human review of AI output covers the training side.
Does the FRC AI guidance apply to compilations and unaudited accounts?
Not directly. Both FRC documents are written for audit. The 2025 guidance says it is not prescriptive and that the requirements firms are assessed against remain only those in the ISQMs and ISAs (UK). The 2026 guidance is aimed mainly at the central technical teams in audit firms that build AI tools and methodology. The ICAEW’s guidance on compilation reports for incorporated entities is TECH 07/16 AAF, issued in March 2016 and available only to subscribers. We have found no official source that deals with AI review of compilation work.
So for unaudited statutory accounts, the approach above is our reading of the audit guidance applied by analogy, not a requirement. It is still worth following, because a compilation file is reviewed by monitoring reviewers and, when a practice changes hands, by the acquirer’s due diligence team, and both want to see who decided what. Many AI points on unaudited accounts are matters for the directors, who approve the accounts under section 393. Where a point is the client’s call, record that it was raised with the directors, what they decided, and whether that changed your view. Our article on accounts preparation file review sets out the wider evidence a compilation file should hold.
What overridden points tell the firm about the tool
In the FRC’s illustrative example, the engagement team is expected to stay alert for information indicating that the tool’s assessment “may be systemically flawed”. The 2025 guidance refers to ISQM (UK) 1 requiring firms to obtain or develop, implement, maintain and use appropriate technological resources, and its documentation table lists training “including strategies to mitigate automation bias”.
In practice, that means looking at overrides across files as well as within them. If the same flag is overruled on client after client, either the tool is wrong for your client base and needs reconfiguring or reporting to the supplier, or reviewers are dismissing it out of habit. Both are worth knowing, and you can only tell them apart if each override records its reasoning. Firms within the scope of ISQM (UK) 1 can feed this into their ISQM 1 annual evaluation as evidence that the technological resource is being monitored.
The time to settle this is before the tool goes live on signed files or before the next round of file reviews, whichever comes first. Write down who may clear which points, what a clearance must contain, and how the original output is kept, then check a sample of overrides at each file review.
Questions practices ask
Do we need to tell clients we use AI to review their accounts?
For tax work, the PCRT guidance on AI suggests a statement in the engagement letter about the potential use of AI tools. We have not found an equivalent requirement for accounts preparation, but the same reasoning applies, and a short clause is simpler than explaining it later. Check at the same time whether the way the tool handles client data calls for a data protection impact assessment.
Will a monitoring reviewer look at the AI tool's output as well as our file?
None of the FRC or ICAEW material we have seen says what monitoring reviewers will ask for. It is safest to assume they may want to see what the tool raised alongside how each point was cleared, so keep the original output instead of overwriting it. If the file shows only the cleared position, you cannot later show what the partner saw when deciding.
Do we have to document how we use the AI tool itself on an audit?
The FRC's 2025 guidance says automated tools and techniques that use AI will require documentation on the audit file in line with ISA (UK) 230. For other AI tools, it says there may be no requirement to document their use on the audit file. In both cases it expects professional judgement in deciding what documentation is appropriate and warns against over documentation.
Does the FRC's 2026 AI guidance create new rules for engagement teams?
The FRC says its Generative and Agentic AI Guidance, published on 30 March 2026, codifies good practice and is not a response to identified quality deficiencies. It is written mainly for the central technical teams in audit firms that develop AI tools and methodology, with engagement teams as a secondary audience. Regulatory accountability for deploying AI tools and for audit quality is unchanged.
Sources
- Companies Act 2006, section 393
- Companies Act 2006, section 414
- AI in Audit: Illustrative example and documentation guidance (HTML version)
- FRC publishes landmark guidance providing clarity to audit profession on the uses of AI
- Innovative new guidance supports audit firm adoption of emerging AI technologies
- Generative and Agentic AI Guidance (HTML version)
- Generative and Agentic AI Guidance factsheet (HTML version)
- ISA (UK) 230 Audit Documentation
- ISA (UK) 220 Quality Management for an Audit of Financial Statements
- Documenting significant professional judgements in an ISA (UK) audit
- Topical guidance: application of PCRT to ethical use of artificial intelligence tools
- How helpful is FRC guidance on generative and agentic AI?
- TECH 07/16 AAF Chartered accountants' reports on the compilation of financial information of incorporated entities