§ Choosing an accounts review tool: questions to ask and how to test it

Choosing an accounts review tool: questions to ask and how to test it

Guides

Choosing accounts review software is easier once you know what each approach to review can and cannot see. This guide is for partners and managers in UK chartered practices who are weighing a tool to help review statutory accounts prepared under FRS 102 Section 1A or FRS 105. It compares five approaches without naming products, sets out a scorecard, explains how to test a tool on files you have already signed off, and lists the questions on data and responsibility that any vendor should answer in writing.

Five approaches to review, side by side

Each approach below finds some review points and misses others. What matters is which, and who decides what to do about them.

Manager review

An experienced manager reads the accounts against the working papers and what they know of the client. A manager knows the director has a second company, or that the lease was renegotiated, so manager review catches judgement points that no single document reveals, as well as the routine casting and disclosure points.

What it misses depends on the person. A reviewer who knows a client well can read what they expect to see, and disclosure coverage rests on memory or on whichever checklist the manager uses.

Cold file review

A cold file review looks at a completed file after the accounts have gone out. ICAEW’s Practice Assurance guidance for smaller firms suggests reviewing some of your files cold from time to time, and checking a sample of accounts prepared by the firm to confirm they are materially correct and comply with statute and other relevant requirements.

Cold review is good at patterns, such as a template carrying the wrong regime statement or a team that never updates the policies note. It works after the event and on a sample, so it protects future files more than the file in hand, and its findings go into the firm’s quality management for someone to act on. Our guide to file review, compiled accounts and evidence of review covers how those findings feed back.

Disclosure checklists built into accounts production software

Accounts production software commonly carries a disclosure checklist or validation rules linked to its templates. They are good at completeness against the framework the software knows, such as a required note that is missing or a figure that does not agree to the mapped trial balance.

A checklist answers the questions it asks, and it relies on the preparer’s answers. A related parties question ticked “not applicable” is accepted whether or not the director lent money to a connected company. It cannot tell whether a policy note describes what was actually done, or whether narrative typed by hand agrees with the numbers, and it sees only accounts built in that software.

General-purpose AI chat assistants

A general-purpose chat assistant can explain a paragraph of a standard, draft a note, or comment on a set of accounts pasted into it, and may spot an inconsistency between pages.

Its limits matter for review work. ICAEW’s guide to generative AI and ethics notes that many generative AI tools are not able to draw on the latest news and legislative developments, so the tool may not know which edition of a standard applies to the period in front of it. It can produce a paragraph reference that does not exist, and the same guide warns against restating as fact outputs that may be hallucinations and asks for professional scepticism when reviewing the output. Its output is a conversation, which is hard to put on file. On data, ICAEW’s regulatory news of 9 November 2026 says entering client data into public AI tools can risk unauthorised disclosure. An enterprise licence under a reviewed contract may answer the data point, subject to the vendor questions below.

Specialist AI review of statutory accounts

A specialist tool is built to read a set of UK statutory accounts against FRS 102, Section 1A, FRS 105 and the Companies Act, and to return review points for a person to assess.

It is limited to what it is given: it does not know what the director said at the planning meeting unless that is in the file. It will raise some points that are not errors, and quality varies between products, which is why the rest of this guide is about testing. The judgement stays with the reviewer and the partner: the tool proposes points, and people decide which are real and whether the file can be signed off.

ApproachStrongest atJudgement sits with
Manager reviewPoints that need knowledge of the clientThe manager, then the partner
Cold file reviewPatterns across the firm’s filesWhoever acts on the findings
Software disclosure checklistCompleteness against its own templatesThe preparer answering it, then the reviewer
General-purpose AI chatExplaining and draftingThe user, for every answer
Specialist AI reviewReading every page against the frameworkThe reviewer and the partner

An evaluation scorecard

Score each candidate against the same criteria, on the same files. These are the five we would use.

CriterionWhat to look forHow to test it
CoveragePoints under FRS 102 Section 1A, FRS 105 and the Companies Act, against the right edition for the periodPilot files from both regimes, with periods beginning before and on or after 1 January 2026
TraceabilityEach point cites a paragraph of the standard or a section of the ActLook up a sample of the references yourself
False positivesPoints raised that are not errorsCount them on your pilot files and record why each was wrong
Output for the fileA record a reviewer can respond to, clear and keepAsk to see the file copy after points are cleared
Data handlingWhere files go, who can see them, what is keptWritten answers to the vendor questions below

Coverage

The FRC’s Periodic Review 2024 amendments to FRS 102 are, in most cases, effective for accounting periods beginning on or after 1 January 2026, and the FRC says the most significant changes apply to leases and revenue recognition. A tool that reviews a period beginning on or after that date against the old lease and revenue wording has a coverage problem that a clean report will hide. Test the Act as well: section 414(3) requires the regime statement in a prominent position above the signature, and a tool should notice when it is missing or names the wrong regime. Our checklist of what a reviewer at the standard of an experienced audit manager looks at on small company accounts gives more points to test.

Traceability

A point without a reference has to be researched before anyone can act on it, and a reference that does not exist is worse than none. The FRC’s guidance on documenting AI tools in audit makes a related point: where the outputs of a tool are all checked back to the cited sources, it may be appropriate to document less about the tool. That guidance is written for auditors and describes itself as not prescriptive, but the logic carries across to accounts work.

False positives

Count the points that are not errors on your own pilot files and group them by cause, because a tool that is consistently wrong about one disclosure is easier to live with than one that is wrong at random. Too many false points train reviewers to dismiss output without reading it. The opposite risk is a reviewer who stops checking the points at all, the overreliance that ICAEW’s ethics guide describes for auditors as automation bias.

Output for the file

ICAEW’s Practice Assurance guidance says to keep evidence of any review on file. Ask what the file copy looks like: each point, its reference, the preparer’s response, who cleared it and when. For a firm’s own record of the tool, the FRC’s audit guidance lists items such as the version number of the tool and model used and the team’s consideration of the outputs of the tool. A practice can borrow that list for accounts work even though nothing requires it to.

How to run a pilot on past signed files

The fairest test uses past files that have been through manager review, because you already know the answer.

  1. Settle the data position first. Past files are still client data, so agree terms with the vendor and complete your own data protection assessment before any file leaves the practice.
  2. Choose signed files across both Section 1A and FRS 105, including periods beginning on or after 1 January 2026 and files with known difficulties, such as a directors’ loan or a going concern note with real judgement in it.
  3. Before running anything, write down the review points raised on each file from the review notes, and any found later at cold review or through a client query. This is the answer key.
  4. Run each file through the tool as it was when signed off. Where the vendor allows, also run altered copies with errors you have planted, such as a removed regime statement or a changed comparative.
  5. Mark every point the tool raises: matches a known point, a real point the original review missed, not an error, or unclear. Check every reference cited.
  6. List what the tool missed from the answer key and look for a shared cause, such as one regime or one kind of judgement.
  7. Have the reviewers who would use the tool write up the result, and keep it with the decision to adopt or not.

Treat “real points the original review missed” with care: ask the partner who signed off the file whether each one would have changed the accounts.

Questions to ask any vendor

ICAEW’s ethics guide says accountants must ask the right questions of suppliers and get the evidence to gain comfort, there about the data used to train models. Its Practice Assurance guidance suggests, as a top tip, an agreement with any cloud service provider which includes clauses on who owns the data, confidentiality and data security. Ask in writing and keep the answers with the pilot write-up.

  • Where are our files stored and processed, including backups and any underlying model provider? The ICO treats making personal information accessible to a separate organisation outside the UK as a transfer, and every restricted transfer must be covered by a transfer mechanism. Ask which one applies and, where a safeguard is relied on, whether a transfer risk assessment has been done.
  • Is our data used to train or improve any model, the vendor’s or anyone else’s? Ask for the contract term that answers this. The ICO’s AI guidance says a provider that builds and trains its own model is likely to be a controller for that element of the processing, which is a different relationship from the one most practices assume.
  • Which sub-processors are involved, and how will we hear about changes? Under Article 28(3) of the UK GDPR the contract must say the processor will not engage a sub-processor without the controller’s prior specific or general written authorisation.
  • What is logged, for how long, and who can read the logs?
  • Who can see our files, including support staff and any model provider, and under what controls? Article 28(3) requires the processor to obtain a commitment of confidentiality from anyone it allows to process the personal data.
  • What happens to our data at the end of the contract? The contract should let you choose whether data is deleted or returned, and should allow for audits and inspections by you as controller.
  • Which edition of each standard does the tool apply to a given period, and how are users told when that changes?

The ICO’s pages on processor contracts and on AI accountability are under review following the Data (Use and Access) Act (the AI page was last updated 15 March 2023), so check the current versions.

Professional responsibility when a tool is part of the review

A tool changes the evidence in front of the reviewer, and leaves unchanged who answers for the accounts. Section 414(1) of the Companies Act 2006 requires annual accounts to be approved by the board and signed on its behalf by a director, and the practice’s own review procedures remain its own responsibility.

ICAEW’s ethics guide says generative AI tools do not make up for a lack of professional knowledge, so the reviewer clearing a point needs to understand it well enough to disagree with it. It also says being transparent and honest in the use of generative AI is essential, which is worth reflecting in how you describe your review to clients. The FRC’s audit guidance, whose scope includes generative AI, lists among its considerations the material on how to interpret a tool’s outputs, including strategies to mitigate automation bias.

Our guide on signing off accounts that an AI tool has helped review goes further.

Client data, confidentiality and the contract

A set of small company accounts usually contains personal data, starting with directors’ names and the details of any loans to them. ICAEW’s regulatory news of 9 November 2026 says accountants’ confidentiality duties still apply when using AI and points firms to subsection 114 of the ICAEW Code of Ethics on the principle of confidentiality. Its 2024 warning on AI tools noted that once client data is in a large language model there is limited visibility and control over who the information is shared with and how long it is retained.

The ICO’s guidance says that in the vast majority of cases the use of AI will involve processing likely to result in a high risk to individuals’ rights and freedoms, which triggers the legal requirement for a data protection impact assessment. Expect to need one before a pilot. ICAEW’s Practice Assurance guidance also suggests, where you use cloud computing, telling your clients where their data is in your engagement letter, which a new tool may change.

Our guide to client data, UK GDPR and the AI tools a practice uses takes the data protection questions further.

We build a tool in the specialist category. Infinite Accountant is an AI review of statutory accounts at the standard of an experienced audit manager, for UK chartered practices, by invitation. We would expect it to be tested against this scorecard like any other, and the judgement and the sign-off stay with the partner. For a longer read, see our guide The Signed Review.

Further reading

Questions practices ask

Can the disclosure checklist in our accounts production software replace manager review?

It covers a different part of the job. A checklist tests completeness against its own questions and relies on the preparer’s answers. It cannot tell whether a policy note describes what was done, which is much of what a manager reads for.

Can we paste client accounts into a general-purpose AI chat assistant?

Not into a public one. ICAEW’s ethics guide points to not loading confidential information into public generative AI tools. An enterprise service under a contract the firm has reviewed is a different position, and should face the same vendor questions as any other tool.

How many files should a pilot use?

Enough to cover both Section 1A and FRS 105, periods beginning before and on or after 1 January 2026, and the judgement areas you care about most, such as directors’ loans and going concern. Files with a good record of review points are worth more than a larger sample without one.

What should we do with false positives found in a pilot?

Count them and record why each was wrong. Ask the vendor how each type would be handled, and keep the answer with the pilot write-up.

Does using an AI review tool change who is responsible for the accounts?

No. Section 414(1) of the Companies Act 2006 requires the board to approve the accounts and a director to sign them. A tool is one input to the practice’s review, and the judgement and sign-off stay with the partner.

Do we need a data protection impact assessment before trialling a tool?

Expect to. The ICO says that in the vast majority of cases the use of AI will involve high-risk processing that triggers the legal requirement for one, and past signed files are still client data.

Sources