Where AI tools can help in a small accountancy practice

The starting point in the guidance

The topical guidance on AI published in January 2026 by the bodies behind Professional Conduct in Relation to Taxation (PCRT) lists the areas where AI tools are being used in tax and accountancy work: collating client data for tax returns, drafting reports, client due diligence, transaction work and technical research. It also says outputs from AI tools should not be used as authoritative tax or legal advice, and must be reviewed by a qualified professional in the context of the specific client (PCRT AI topical guidance, January 2026).

The same guidance asks members to treat AI output as if it had been prepared by a less experienced junior colleague, and to review it with appropriate scepticism. That is a useful test for every use described here. Ask what you would check if a trainee had done the work, then check the same things.

Reading documents and receipts

• What it does. A tool reads a scanned receipt, invoice or bank statement and pulls out the supplier, date, amount and VAT.
• Where it goes wrong. Misread digits, a total read as a net figure, VAT picked up from an invoice that does not show a VAT number, or a date read in the wrong order.
• The review it needs. Compare a sample of extracted lines with the documents, and check every item above a value the practice sets. Agree the totals extracted to the bank statement or the supplier statement. Keep the original document with the record, since the extraction is not the evidence.

Proposing codes for bank lines

• What it does. A tool proposes a nominal account and VAT treatment for each bank line, from the payee and description.
• Where it goes wrong. The same payee can need different codes, for example a supermarket used for both stock and personal shopping. A proposal can look confident and still be wrong for this client.
• The review it needs. Look at every line coded to a VAT-bearing account and every line over the practice's review limit. Reconcile the bank account to the statement. Look at the profit and loss for accounts that have moved more than expected against last year.

Drafting letters and emails

• What it does. A tool produces a first draft of a client letter, a reply to HMRC or an internal note.
• Where it goes wrong. Wrong facts about the client, invented rules, out-of-date rates, a tone that does not suit the reader, or confidential detail included where it should not be.
• The review it needs. The PCRT guidance says correspondence prepared with AI should be reviewed so its tone and content are appropriate before it is sent, and gives the example of a reply to HMRC (PCRT AI topical guidance). Check every rate, date and rule in the draft against an official source before the letter goes.

Summarising long material

• What it does. A tool summarises a meeting recording, a long email chain, a lease or a set of minutes.
• Where it goes wrong. It leaves out the one condition that matters, merges two people's views, or states something that was not said.
• The review it needs. The PCRT guidance gives the example of meeting minutes produced by AI and says they should be reviewed to make sure they are complete, reflect the discussion and are professionally worded (PCRT AI topical guidance). The person who was in the meeting checks the summary.

First-pass review of figures and variances

• What it does. A tool compares this year's figures with last year's, or with a budget, and lists movements for someone to look at.
• Where it goes wrong. It explains a movement with a plausible reason that is not the real one. It may miss an error that leaves the figures looking normal, such as a sale posted to the wrong year.
• The review it needs. Treat the list as a starting point. The reviewer still forms their own view of what should have moved, and asks the client about the reasons. The explanation on file is taken from the client or the records.

Technical research

ICAEW's article on validating AI content, published on 7 September 2026, says AI should never be treated as an authoritative source and that every technical claim should be checked against a trusted source (ICAEW, using AI with confidence). The PCRT guidance refers to the tax tribunal case of Harber v HMRC [2023] UKFTT 1007 (TC) as an example of why the existence of case law cited by a tool should be confirmed (PCRT AI topical guidance).

For a practice, that means a tool can help find where to look. The answer that goes to the client is taken from legislation, HMRC's manuals or a professional body, and the reference is kept on file.

Security risks the NCSC describes

The National Cyber Security Centre (NCSC) lists the main weaknesses of large language model tools: they can present incorrect statements as fact, they can be biased, they can be manipulated by prompt injection, and their training data can be tampered with (NCSC: AI and cyber security, 13 February 2024).

Prompt injection matters to a practice that lets a tool read incoming documents or emails. Text hidden in a document can carry instructions to the tool. A tool that reads client documents should not also be able to send messages, make payments or change records without a person approving each action.

Questions