§ InsightsNo. 217

FRC generative AI guidance and what it expects of your practice

What the FRC's March 2026 guidance on generative and agentic AI expects of firms, and how a practice preparing accounts can apply it before December 2026 year ends.

Two accountants reviewing a printed set of accounts, illustrating FRC generative AI guidance for accountancy practices

This article is stated as at 21 November 2026. The FRC's AI guidance is new and may be revised or added to, and how far ISQM (UK) 1 applies to accounts preparation work depends on the engagement. Take advice on your own circumstances before acting.

The FRC generative AI guidance, published on 30 March 2026, names three risks to audit quality (deficient output, misuse of output and a methodology that does not meet auditing standards) and four types of mitigation. It is written for audit firms, but any chartered practice that uses AI to prepare or review statutory accounts can apply the same framework. And it is clear that the firm and the engagement partner keep full responsibility for the work.

What does the FRC generative AI guidance cover?

The FRC’s Generative and Agentic AI Guidance covers the use of generative and agentic AI tools in audit engagements and how firms can mitigate the risks these tools pose to audit quality. It defines generative AI as technology that “can generate content in response to prompts, using learned patterns from large training datasets”, and agentic AI as systems that “orchestrate and execute multiple components and/or tasks toward a goal, with some degree of autonomy”.

Its main audience is the central technical teams in audit firms who build AI tools and the methodology around them. A smaller practice has no central technical team, so the partner who approves the firm’s tools is doing that job. The guidance deals only with risks to audit quality and leaves organisational, cybersecurity and financial risk to one side, which a practice still has to deal with (particularly data protection) through its own processes.

The FRC says it “is not a response to identified quality deficiencies”, and that it codifies good practice and lays a foundation for later work. Neither the press release nor the guidance states in terms whether it is mandatory, and the FRC says the nature and extent of the mitigations “is a matter of professional judgement”. It does say plainly that regulatory accountability for deploying AI tools and for the quality of audit outputs “remains unchanged”, and that “the human auditor is always accountable”. The factsheet puts it in standards terms: the firm and engagement partner retain full responsibility for audit quality, under ISQM (UK) 1 and ISA (UK) 220 respectively.

Does it apply if we only prepare accounts and do not audit?

The guidance is addressed to audit, so for a practice that mainly prepares unaudited statutory accounts it applies by analogy. ISQM (UK) 1 is titled as covering firms that perform audits or reviews of financial statements “or other assurance or related services engagements”, but whether your accounts preparation work falls within its scope depends on the engagement and on your regulator’s position, so confirm the position for your own engagements.

The three risks translate directly into accounts review. A disclosure checker that misses a related party note is deficient output. A manager who treats a clean report as clearance for the whole file is misusing output that may itself be fine. A firm procedure that lets an automated check replace the manager’s read through of the accounts against the trial balance is a methodology problem, however good the tool is. The partner who approves the accounts carries the same responsibility as before, which is set out in our page on signing off AI reviewed accounts. While you confirm the scope question, apply the four mitigations below to any AI used on accounts work, because the partner’s responsibility for the accounts is the same whether or not ISQM (UK) 1 applies.

The three risks the FRC names, applied to accounts review

The FRC guidance itself splits the risks as follows.

Risk (FRC, March 2026) What the FRC means How it can show up in accounts review
Deficient output The output of the AI tool is deficient A tool summarises a director’s loan note too simply, or misses a comparative that was not updated
Misuse of output The output is misused although it is itself appropriate, split into misinterpretation of output and misunderstanding of methodology A reviewer assumes anything not flagged needs no work, or does not know which checks the tool actually runs
Non compliant methodology The firm’s methodology permits approaches using the tool that fail to meet standards, even where the output is appropriate and used as specified The firm’s procedures let an AI check stand in for review steps the standards still require of a person

The factsheet says the risk of deficient output “can not be eliminated as LLMs have inherent limitations”, including a finite capacity that can lead to over simplification, so the controls have to assume the tool will sometimes be wrong. In its contract review example, the FRC names a misuse risk that every reviewer will recognise: believing there is no responsibility for contracts the tool did not flag as higher risk. The accounts equivalent is treating an unflagged note as reviewed, so the file should show that the reviewer read the whole set of accounts, not only the items the tool flagged.

The four mitigations, scaled to a smaller practice

The FRC groups mitigations into four categories which, taken together, lead to “appropriate confidence in quality of output”. How much a firm does in each is a matter of judgement, and a smaller practice can scale each category to its size.

System design and development. Most practices buy their tools, so the design questions become questions for the provider: what the tool checks, what it does not, how it handles a set of accounts it was not built for, and what it does with client data.

Certification. The FRC describes evaluating inputs, the system’s architecture and component logic, and testing outputs on default cases, edge cases, diverse cases and known issues, followed by recertification and monitoring. The thing certified is “an AI tool in the context of a use case”, so approving a tool in general is not enough: a tool approved for checking disclosures against FRS 102 Section 1A has not been approved for drafting going concern wording. And the FRC accepts that tests “will often be qualitative in nature”. A practical way to do this is to run the tool over files the firm has already reviewed and signed, and compare its findings with the points your managers raised, as our guide to piloting a review tool on past files describes.

Staff education and governance. The FRC says training “is an ongoing activity, rather than something which is only relevant at the launch of a tool.” Governance means deciding who may use which tool for what, and writing it down in an AI use policy for the practice.

Human in the loop review and oversight. The FRC defines the human in the loop as someone who “directs, authorises or reviews the system’s actions or outputs at runtime”. Final outputs are reviewed by appropriate reviewers, who apply professional scepticism and stay aware of the risk of automation bias. In practice the reviewer needs to be experienced enough to disagree with the tool and should record where they did, which is the subject of our note on evidencing human review of AI output.

What to have in place before December 2026 year end files

The FRC does not say in terms that a firm must have a written AI policy, but its governance category points that way, and ICAEW’s audit regulation guidance stresses policies, training and oversight. ICAEW also warns that staff may bring AI tools they use personally into work without authorisation, which risks data protection breaches and the release of confidential client information. Before the December 2026 year end files arrive, we would have the following settled:

  • a written policy listing the tools the firm has approved, the specific uses each is approved for, and a plain statement that public tools are not to be used on client data
  • a named partner responsible for approving and periodically retesting each tool and use
  • a short file note on each engagement where AI was used, saying which tool, for what, and how it shaped the review work
  • training dates in the calendar beyond the first session, covering automation bias and the tool’s known limits
  • AI use added to the firm’s quality risk assessment

The file note follows the FRC’s board minutes example, in which the auditor records on file that a generative AI tool was used and how it informed further work, while the documentation about the tool itself is held centrally. The same example says the summary “does not replace the auditor’s risk assessment”, and that the misuse risk is treating the summary as a complete account of the period. Mitigations in that example include prewritten prompts, training on automation bias and mandatory human review of the minutes themselves.

On quality management, the FRC encourages firms to consider how their AI use fits within ISQM (UK) 1, which requires a system of quality management providing reasonable assurance that the firm and its people conduct engagements in accordance with professional standards. Where ISQM (UK) 1 applies, a new tool affects the firm’s technological resources and belongs in the risk assessment and in the responsible partner’s annual evaluation. ICAEW’s practice guidance adds a useful working rule for staff: treat AI output as a starting point, and check every technical claim against a trusted source.

Independence: the November 2026 guidance on technology and audit clients

On 16 November 2026 the FRC published Ethics and Technology guidance on sharing technology derived insights with management or those charged with governance, and on giving audited entities access to technology that produces insights directly. It states that it “does not set any additional requirements”.

The FRC says generative AI outputs, “which often sound highly plausible, assertive and confident”, heighten the management threat, and that technology “may enable a significantly higher volume of insights”, which may increase the risk that management act on an insight without evaluating it critically. It also says continuing access that lets the client’s staff monitor controls or generate insights for their own purposes is less likely to be part of the audit. Communications the auditor is required to make are not treated as a non audit service.

For audit clients, giving management ongoing access to the firm’s analytics needs thinking through against the Ethical Standard. For accounts clients the independence rules are different, but the point about confident sounding output still applies to anything you pass on, and it is one reason to think about telling clients you use AI in their accounts.

Key dates

Date What happened
15 December 2022 ISQM (UK) 1 effective
26 June 2025 FRC publishes AI in audit: Illustrative example and documentation guidance
25 September 2025 Updated edition of ISQM (UK) 1 published
30 March 2026 FRC publishes Generative and Agentic AI Guidance and factsheet
16 November 2026 FRC publishes Ethics and Technology guidance

Questions practices ask

Who is responsible if an AI tool gets the accounts wrong?

The firm and the partner who signs remain responsible. The FRC says regulatory accountability for deploying AI tools and for the quality of the output is unchanged, and its factsheet says the firm and engagement partner retain full responsibility for audit quality. A provider's terms may give the firm some contractual recourse, but they do not move professional responsibility away from the people who approve and sign the work.

Do we have to approve an AI tool separately for each task?

The FRC's approach points that way. It describes certification as applying to an AI tool in the context of a use case, so a tool tested for disclosure checks has not been tested for drafting notes or summarising minutes. A practical approach is to list each approved use in the firm's policy and test the tool on past files for that use before staff rely on it.

Can staff use a free public AI tool on client information?

ICAEW warns that staff bringing personally used AI tools into work without authorisation risks data protection breaches and the release of confidential client information. Most practices should say clearly in their AI policy that client data may only be used in tools the firm has approved, and should consider whether a data protection impact assessment is needed before approving one.

How often should staff be trained on the AI tools we use?

Staff should be trained more than once, because the FRC says training is an ongoing activity and not something only relevant when a tool is launched. Sensible refreshers cover the tool's known limits, how to spot automation bias and what to record on the file, and are worth repeating when the tool changes or when file reviews show staff relying on its output too readily.

Sources