§ Signing off AI-reviewed accounts: professional responsibility

Signing off AI-reviewed accounts: professional responsibility

Guides

Professional responsibility for AI-reviewed accounts sits exactly where it sat before any software was involved: with the directors who approve and sign, and with the partner whose practice prepared and reviewed the file. This guide is for partners deciding whether and how to use AI in accounts preparation and review. It sets out what the Companies Act, the ICAEW Code of Ethics and the FRC’s AI guidance say, what the CCAB has so far published only in draft, and how a practice can show on file that a person reviewed what the tool produced.

Who remains responsible

Section 414(1) of the Companies Act 2006 requires a company’s annual accounts to be approved by the board and signed on its behalf by a director, and s414(2) puts that signature on the balance sheet. Section 393(1) says the directors must not approve accounts unless they are satisfied that the accounts give a true and fair view. Under s414(4), a director who knew the accounts did not comply, or was reckless as to whether they did, and failed to take reasonable steps to secure compliance or to prevent the accounts being approved, commits an offence.

The practice’s responsibility runs alongside the directors’. It owes the client the service in the engagement letter, and its partners owe compliance with the ICAEW Code of Ethics. When an AI tool reviews a draft, three parties are still in the frame:

  • The directors, who approve and sign the accounts and remain responsible for them in law.
  • The practice, which chose the tool, decided how it is used on client work, and answers for the engagement.
  • The partner who signs off the file and decides whether each point the tool raised is right and whether anything was missed.

The CCAB’s draft statement on AI makes the same point: the use of AI tools “does not diminish the personal responsibility or accountability of the professional accountant”, and accountants “cannot abdicate responsibility for outputs produced by their use of AI tools”. For tax work, the PCRT topical guidance on AI published on icaew.com on 19 January 2026 says members “remain ultimately accountable for any work produced” (paragraph 3.5).

The ICAEW Code of Ethics and AI

The ICAEW Code of Ethics 2026 applies from 1 July 2026 and is derived from the IESBA Code issued in October 2025. Its provisions on technology sit within the existing principles and sections.

Professional competence and due care

R113.1 requires an accountant to maintain professional knowledge and skill and to act diligently and in accordance with applicable technical and professional standards. Paragraph 113.1 A3 says maintaining competence needs a continuing awareness of technology-related developments relevant to the accountant’s work. R113.2 requires reasonable steps to ensure that people working under the accountant’s authority have appropriate training and supervision. ICAEW’s own guidance on generative AI and ethics makes the practical point: such tools “do not make up for a lack of professional knowledge and skill”.

Objectivity and reliance on technology

The Code’s definition of objectivity (110.1 A1 and R112.1) is judgement uncompromised by bias, conflict of interest, or “undue influence of, or undue reliance on, individuals, organisations, technology or other factors”. A reviewer who clears a file because the tool found nothing is relying on it, and the Code asks whether that reliance was due.

Using the output of technology

Part 3 of the Code, for accountants in public practice, has a specific requirement. R320.11 says that when an accountant intends to use the output of technology in a professional activity, the accountant shall determine whether the use is appropriate for the intended purpose. Paragraph 320.11 A1 lists the factors, which include:

  • the nature of the activity performed by the technology and the expected extent of reliance on its output;
  • whether the accountant, or an expert they can call on, can understand, use and explain the technology;
  • whether the technology has been appropriately tested and evaluated for the purpose intended;
  • the firm’s oversight of its design, implementation, monitoring and updating, and the controls over who uses it;
  • the appropriateness of the inputs, including data and the decisions people make while using it.

Paragraph 300.6 A2 adds examples of threats, such as data that is not sufficient for the tool or a tool that is not appropriate for the purpose, and a self-review threat where the technology was designed or developed using the firm’s own expertise.

Integrity and confidentiality

R111.2 prohibits being knowingly associated with information that contains a materially false or misleading statement, or statements or information provided recklessly. A review point a tool invents, passed to a client unchecked, is a risk under that paragraph. On confidentiality, 114.1 A1 covers the collection, use, transfer, storage or retention, dissemination and lawful destruction of client information. ICAEW’s generative AI guidance speaks of “not loading confidential information into public Generative AI tools”. Our guide to client data and data protection when a practice uses AI tools covers the UK GDPR side.

The CCAB draft statement

In June 2026 the CCAB published a draft statement to the accountancy profession on the ethical use of artificial intelligence, labelled as draft guidance for consultation, with six draft case studies. The CCAB AI hub describes it as “a draft for engagement and consideration by the profession” and we found no closing date on it. It applies the five fundamental principles to AI, lists algorithmic bias, automation bias, hallucinations and opacity among the risks, and suggests practices tell clients about AI use in the engagement letter, and it is not yet final. Checked against the CCAB AI hub on 21 November 2026.

The FRC’s AI guidance and what it means outside audit

On 26 June 2025 the FRC published “AI in audit: Illustrative example and documentation guidance”. The FRC says it “is not prescriptive and does not introduce new regulatory requirements”, and the document says the requirements firms are assessed against remain only those in the ISQMs and ISAs (UK). It is written for audit. Applying it to accounts preparation is our analogy, and the FRC does not make it.

The guidance rests on two standards. Citing ISQM (UK) 1 paragraphs 32(f), 32(g) and 57(c), it says ISQM (UK) 1 requires the firm to obtain or develop, implement, maintain and use appropriate technological and intellectual resources, and to document how it has addressed the risks that it might not meet these objectives. ISA (UK) 230, paragraph 8, is the basis for its statement that automated tools and techniques using AI will require documentation on the audit file. Paragraph 8 asks for documentation sufficient to enable an experienced auditor, having no previous connection with the audit, to understand the work done, the results and the significant judgements.

The guidance splits documentation between the firm centrally and the engagement file. Centrally, the firm records what the tool does, when to use it, why the firm is confident it works as intended, the training given (including strategies to mitigate automation bias) and how the tool is appropriately explainable. On the file, it may be appropriate to record the version of the tool and model used, any configuration, the data or prompts put in, and how the team used the outputs to reach its conclusion.

The guidance also covers tools that “review work to identify omissions or inconsistencies”, which is a fair description of an AI accounts review. For tools of that kind it says there may be no requirement to document their use on the audit file, if an experienced auditor would not need it to understand the basis for the report. It also notes that where outputs are all checked back to the cited sources, less documentation about the tool may be needed.

Two later FRC publications are also written for audit firms. The Generative and Agentic AI Guidance of 30 March 2026 is aimed at the central technical teams within audit firms; the FRC’s announcement says “the human auditor is always accountable”. On 16 November 2026 the FRC published guidance for audit firms on maintaining independence when applying technology, including AI.

What we would take from all of this for non-audit accounts work, by analogy only:

  • Keep a short central record for each tool: what it does, what it is approved for, how the practice satisfied itself that it works, and how staff are trained to use it.
  • On each file, record which tool and version ran, on what inputs, and what the reviewer did with each point it raised.
  • Apply the experienced-reviewer test: could a reviewer with no connection to the job see what was checked, by whom, and why each point was cleared?

Inquiring mind, automation bias and evidence of review

The right word for accounts work

The Code ties professional scepticism to auditing, review and other assurance standards (120.5 A3 and 120.16 A1). For accounts preparation, the duty that applies to every accountant is the inquiring mind in R120.5: considering the source, relevance and sufficiency of information, and being open to the need for further investigation. That is the standard the partner applies to AI output on a compiled set of accounts. On an audit or review engagement, scepticism applies as well.

Automation bias

Paragraph 120.12 A2 of the Code defines automation bias 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. The FRC’s March 2026 guidance tells reviewers of AI output to apply professional scepticism, to know the main ways the tool’s output can be wrong, and to be aware of the risk of automation bias. Its worked example reminds staff that a tool can sound fluent and confident even when it is wrong.

In accounts review, automation bias shows itself in two ways. A reviewer accepts a point because the tool raised it, without opening the note. Or a reviewer treats a clean output as a clean file. The second is the more dangerous, because nothing on the page prompts a second look. The Code’s suggested mitigations (120.12 A3) include consulting with others to ensure appropriate challenge and training on identifying bias.

How to evidence human review on file

ISA (UK) 230, paragraph 9(c), requires an auditor to record who reviewed the audit work and the date and extent of the review, and paragraph 11 requires the auditor to document how they addressed information inconsistent with their final conclusion on a significant matter. Those are audit requirements; as a model for an accounts file, they work well. We would expect the file to show:

  1. The tool and version used, and the draft it ran on.
  2. Each point the tool raised, with the reviewer’s response: accepted and corrected, or rejected with the reason.
  3. The areas the reviewer covered personally, whatever the tool said, and in particular the judgement areas listed below.
  4. Any point where the reviewer disagreed with the tool, and how that was resolved.
  5. Who cleared the file, and when, against the version the board approved.

Our guide to building review evidence into a practice’s quality management goes further on file review, and the reviewer’s checklist for FRS 102 and FRS 105 accounts sets out the checks themselves.

Writing an AI use policy for the practice

An ICAEW Audit and Beyond article on AI in smaller firms (14 July 2025) advises: “Create a simple AI usage policy”. A practice policy would usually cover:

  • Which tools are approved, for which tasks, and who approves a new one, with the evidence that it was tested for that purpose (R320.11 and 320.11 A1).
  • What client information may go into which tool, and the rule that confidential information never goes into a public tool.
  • Who may run each tool, what training they need first, and how their work is supervised (R113.2).
  • The rule that every output is checked by a person competent to judge it before it is used, and that a clean output is never treated as a completed review.
  • What is recorded on file and centrally, following the points in the previous section.
  • What clients are told, and how a client’s objection to AI use is handled.
  • Who owns the policy, and how often it is reviewed as tools and guidance change.

Our guide on testing an accounts review tool before relying on it covers choosing the tools themselves.

Telling clients and talking to your broker

Telling clients

R113.3 of the Code requires an accountant, where appropriate, to make clients aware of the limitations inherent in the activities and explain their implications. On 13 March 2026 ICAEW reported that its engagement letter Terms of Business had been updated to refer to the use of AI, with new guidance on whether AI or other software tools will be used, their limitations and potential impact on outputs, and responsibilities for data protection, confidentiality and due diligence on technology providers. For tax work, paragraph 1.3 of the PCRT AI guidance suggests an engagement letter statement indicating that AI-enabled software may be used.

Questions to raise with your broker

ICAEW’s Professional Indemnity Insurance Regulations, effective from 1 October 2026, say the required cover is only a minimum and “you should always consider if this is adequate for your firm”, and recommend the help of an insurance broker. They also point firms to other cover, for example cyber cover. We found nothing in them about AI, so we make no statement about what insurers require. The questions we would put to the broker:

  • Does the policy wording treat work reviewed with an AI tool any differently from other work, and does any exclusion or condition touch technology or software?
  • Does the insurer want to be told that the practice uses AI in accounts preparation or review, at renewal or before?
  • Would a claim arising from an error the tool missed, or a point it invented, be handled as an ordinary professional negligence claim?
  • Is a data incident involving an AI supplier covered by the PII policy, a cyber policy, or neither?

Where the tool stops: the judgement areas

An AI review can read every page against every other page, which is a real strength on a set of small company accounts. Whether the going concern basis is appropriate, whether a related party relationship has been missed because nobody mentioned it, whether a director’s loan is recoverable, whether a dividend was lawful on the relevant accounts: each of these depends on facts outside the draft and on a conversation with the client. Our guide to going concern, related parties, directors’ loans and dividends takes each in turn.

The FRC’s March 2026 guidance describes good review as checking that output is complete, accurate, internally consistent and consistent with the reviewer’s understanding of the entity, and notes that the review may involve forming an independent view to compare the output against. In accounts work, the partner’s knowledge of the client is that independent view. ICAEW’s September 2026 article on validating AI content puts it more bluntly: “AI should never be treated as an authoritative source.”

We built Infinite Accountant to carry out the reading part: an AI review of statutory accounts at the standard of an experienced audit manager, for UK chartered practices. The judgement calls and the signature stay with the partner. The Signed Review sets out the principle we work to, that no partner is ever asked to sign a figure that cannot be traced back to its source.

Further reading

Questions practices ask

Can a partner sign off accounts that an AI tool has reviewed?

Yes, provided the partner has done the review. The directors approve and sign the accounts under s414, and the partner’s sign-off rests on their own judgement. The ICAEW Code requires an accountant who uses the output of technology to decide whether that use is appropriate for the purpose (R320.11).

Does the FRC’s AI guidance apply to accounts preparation?

No, because the FRC’s June 2025 and March 2026 AI guidance is written for audit firms, and the June 2025 guidance says it is not prescriptive and introduces no new requirements. A practice can still borrow its approach to documenting tools centrally and on file, as long as it is clear that it is doing so by choice.

Is there final CCAB guidance on AI?

Not yet: the CCAB published a draft statement on the ethical use of AI in June 2026 as draft guidance for consultation. On 21 November 2026 the CCAB AI hub still presented it as a draft, and we found no closing date on it.

Do we have to tell clients that we use AI?

The Code requires accountants, where appropriate, to make clients aware of the limitations of the work (R113.3). ICAEW’s engagement letter Terms of Business now refer to AI, and the CCAB draft suggests stating AI use in the engagement letter.

What should the file show about AI use?

The tool and version, the inputs, each point raised and how the reviewer dealt with it, the areas the reviewer covered personally, and who cleared the file and when.

Sources