# AI Policy for an Accountancy Practice: 7 Clauses to Copy

Canonical URL: https://business.theaieducator.io/posts/ai-policy-for-an-accountancy-practice
Publication: Dan Fitzpatrick Insights
Author: Dan Fitzpatrick
Topic: AI Governance and Risk
Published: 2026-11-18T16:46:11.000Z
Modified: 2026-11-18T16:49:40.634Z

A complete AI policy for a small accountancy practice, free to copy: seven clauses in full wording, each traced to the PCRT guidance of January 2026 or to ACCA, plus a job-by-job table and the test that shows whether your team has taken it in.

## In brief

An accountancy practice's AI policy needs seven clauses: who it covers, approved tools, client information, review, what clients are told, what staff tell each other, and training and mistakes. Write the review clause first: treat AI output as a junior's first draft that a qualified person checks at source and signs. The policy works when three people in the practice give the same three answers.

## Key takeaways

- The PCRT bodies' AI guidance of January 19, 2026 says members and regulated firms remain responsible for their work irrespective of the use of AI tools in its creation.
- A general small-business AI policy leaves out what a practice needs most: client confidentiality, checking at source and a named reviewer.
- The Junior Rule treats AI output as a first draft from a bright junior: a qualified person reads it, checks every source and figure, and signs it.
- In Elden v HMRC a tax tribunal found that unverified AI case summaries amounted to professional incompetence by the regulated firm involved.
- Removing a client's name does not make data safe for a public AI tool if the figures, trade or location still identify them.
- Staff should say when they used AI on a piece of work, because a reviewer cannot check what they do not know was machine-written.
- A finished policy passes the Readiness Test when a partner, a trainee and the payroll clerk give the same three answers.

An AI policy for an accountancy practice needs seven clauses: who it covers, which tools are approved, what client information may go into them, who reviews the output, what clients are told, what staff tell each other, and what happens when something goes wrong. All seven are below, free to copy, each traced to the professional bodies' own guidance.

*Sources checked November 18, 2026. I am not an accountant or a lawyer, and this is not regulatory advice. It is written for practices in the UK. If you practice elsewhere, the seven headings travel but the sources do not.*

## Why does an accountancy practice need its own AI policy?

Because the general version leaves out the three things your professional body cares about most: client confidentiality, checking at source, and who answers for the work. My [AI policy template for a UK small business](https://business.theaieducator.io/posts/uk-small-business-ai-policy-template) will do for a florist or a joinery firm. A practice that files returns in other people's names needs more.

The profession has now said so in writing. On January 19, 2026, the professional bodies behind Professional Conduct in Relation to Taxation, known as PCRT, published [topical guidance on the ethical use of AI tools](https://www.icaew.com/technical/tax/working-in-tax/pcrt/topical-guidance-application-of-pcrt-to-ethical-use-of-artificial-intelligence-tools) in tax work. It adds no new principle. It reads the five you already know (integrity, objectivity, professional competence and due care, confidentiality, professional behaviour) with AI in the room, and it opens with a sentence worth pinning above every desk:

> "Members are ultimately responsible for any work they produce, and for regulated firms any work which the firm prepares, irrespective of the use of AI tools in its creation."

The next line is blunter: "If a member fails to adhere to the principles set out in PCRT they are liable to be subject to the disciplinary process." And where the guidance says "member," it means the firm and its staff too.

ACCA has been just as plain. In a [technical article for practitioners published in March 2026](https://www.accaglobal.com/uk/en/technical-activities/technical-resources-search/2026/March/Key-areas-of-risk-when-adopting-AI.html), three lawyers from Kingsley Napley wrote: "As a starting point, all firms should have firmwide policies in place which cover responsible use of AI tools."

Small practices are heading into this whether they have written anything down or not. ICAEW's [Sole Practitioners and Small Firms in Focus 2026 research](https://www.icaew.com/technical/practice-resources/practice-news/sole-practitioners-and-small-firms), reported on November 16, 2026, invited practising certificate holders at more than 9,000 ICAEW member firms with between one and ten principals. Fieldwork ran in August 2026 and 8% responded. Of those, 41% of sole practitioners and 67% of small firms expect to use AI more. ICAEW's own reading is that confidence takes more than picking a tool: firms also asked for training, better integration with the software they have, and more confidence about data and cyber security.

A policy is how a practice of six gives itself that confidence without hiring anyone.

## What happened in Elden v HMRC?

A firm of chartered accountants put AI-written case summaries in front of a tax tribunal without checking them, and the tribunal called it professional incompetence.

I have not read the judgment itself. I am relying on the account given by Susan Cattell, Head of Tax Technical Policy at ICAS, in ["AI at the tax tribunal"](https://www.icas.com/news-insights-events/news/tax/ai-at-the-tax-tribunal/), published January 26, 2026. The case is Gary Elden v HMRC [2026] UKFTT 41 (TC). The taxpayer represented himself at the hearing, but his representative, a firm of chartered accountants, had prepared the skeleton argument. It summarized five cases. HMRC said the summaries were inaccurate and had been produced by AI.

The firm argued that it did not matter whether AI had been used. The tribunal half agreed, and then turned the point against them: "the human who relies on its use bears the responsibility for the accuracy." It read the five cases, concluded the summaries had come from AI, and found they "had not been verified for accuracy with sufficient care as should be used when producing submissions for a Tribunal hearing. This lack of sufficient care amounts to professional incompetence on the part of any regulated individual or firm involved in the production of the skeleton argument."

It also said it would consider referring to their professional body anyone whose work fell below the standard it had a right to expect.

Nobody in that practice set out to mislead a judge. Somebody was busy, the summaries read well, and no rule in the building said who had to open the cases. That missing rule is clause 4.

## What is the Junior Rule?

The Junior Rule is the rule I suggest an accountancy practice writes first, and it is borrowed almost whole from the PCRT guidance.

Paragraph 3.4 of that guidance says: "The output from an AI tool should also be regarded as if it were prepared by a less experienced junior colleague and reviewed with appropriate scepticism." Every partner already knows how to manage a bright trainee in their first week. You are glad of the help. You read everything. You do not let them post anything to HMRC on their own. And when a client rings to complain, you do not say the trainee did it.

So here is the rule, in my wording:

> The Junior Rule is how I suggest an accountancy practice treats anything an AI tool produces: as a first draft from a bright junior in their first week. Someone qualified reads it, checks every source and figure it relies on, and puts their own name to it before it leaves the building.

It is a suggestion from me, not a regulation. The idea underneath it belongs to the professional bodies.

## What should the seven clauses say?

Each clause below is short enough to read aloud at a team meeting, and each has a blank for a name. Copy them, fill in the square brackets, delete what does not fit your practice. PCRT is guidance about tax work; I have written the clauses for the whole practice, because your bookkeeper and your payroll clerk use the same tools.

**Clause 1. Who this covers and who owns it**

> This policy covers everyone who does work for [Practice]: partners, employees, trainees, subcontract bookkeepers and anyone else who can open a client file. [Name] owns it, keeps the list of approved tools, and reviews it every six months or whenever a tool changes.

*Where it comes from:* PCRT's "member" includes the firm and its staff, and paragraph 1.6 says "A member is responsible for the work done by staff and others under their supervision." In a small practice the named owner is what I call an AI steward, and [the one-page steward role description](https://business.theaieducator.io/posts/should-a-small-business-hire-an-ai-lead) sets out the hours and the duties.

**Clause 2. Approved tools**

> Client work may be done only in the AI tools listed here, on accounts the practice controls: [tool, and what it is for]. Personal accounts and free versions are not approved for client work. Before a tool joins the list, [Name] gets written answers from the supplier on what data it collects, where that data is stored, how long it is kept and whether it is used to train models. AI features that appear inside software we already use are treated as new tools, and so are new versions.

*Where it comes from:* the four supplier questions are ACCA's. Its article adds that if a vendor cannot answer them clearly, "firms should think carefully before making a purchase." PCRT says an updated tool "should be approached with caution" until it has shown it works at least as well as the version before.

**Clause 3. Client information**

> Nothing that could identify a client goes into an AI tool that is not on the approved list. Taking the name off is not enough: if the figures, the trade or the location would let someone work out who it is, it is still client information. If a client asks us not to use AI on their affairs, we record the request on their file and [Name] agrees with them what it means in practice.

*Where it comes from:* PCRT paragraph 4.2 says "The input of client data into publicly available AI tools is likely to constitute a breach of client confidentiality, unless the client has consented to this." Paragraph 4.5 explains why: "When information is entered into publicly available AI tools, control over that data is relinquished." The guidance also warns that a public tool may identify a client from "an uncommon service they are known to provide." If the worst has already happened, my [Paste Plan for the day an employee pastes customer data into ChatGPT](https://business.theaieducator.io/posts/employee-pasted-customer-data-into-chatgpt) covers the first hour.

**Clause 4. Review: the Junior Rule**

> Anything an AI tool produces is treated as a junior's first draft. A named, qualified person reviews it before it reaches a client, HMRC or a tribunal. Every case, section of legislation and piece of HMRC guidance that an AI tool cites is opened at the source and read. Every figure is traced back to the client's records. "The software produced it" is never an explanation we give.

*Where it comes from:* PCRT paragraphs 3.4 and 3.5, and the Elden decision. The guidance says "Confirming the existence of case law or legislation referenced within the output can help mitigate the risks associated with hallucinations." If your team has never been shown how to check, start with [how to train staff to check AI output](https://business.theaieducator.io/posts/train-staff-to-check-ai-output).

**Clause 5. What we tell clients**

> Our engagement letter says that we may use AI-enabled software in our work, and that a qualified member of our team reviews and remains responsible for everything we send. Where AI is fundamental to a piece of work, we tell the client before we start. If a client asks how a figure or a conclusion was reached, we can explain it and show the source.

*Where it comes from:* PCRT paragraph 1.3 suggests "an appropriate statement in the engagement letter which specifies the potential use of AI tools (eg indicating that AI-enabled software may be used)." Paragraph 1.4 says that where a client queries an element of the work, members "should be able to explain how the conclusion was arrived at, even if the work was generated by AI." For the wider question of when a business owes its customers a word about AI, see [AI disclosure for a small business](https://business.theaieducator.io/posts/ai-disclosure-small-business).

**Clause 6. What we tell each other**

> Anyone who uses AI on a piece of client work says so when they pass it for review, and notes on the file which tool was used and for what. For technical research we keep the prompt and a copy of the sources we relied on.

*Where it comes from:* PCRT's integrity table: "When a staff member uses an AI tool in their work, they should disclose this to senior colleagues." It also suggests "downloading and retaining copies of relevant webpages and screenshots of prompts" as evidence that reasonable care was taken. A reviewer cannot check what they do not know was machine-written.

**Clause 7. Training, mistakes and review**

> Nobody uses an approved tool on client work until they have been shown what it does, where it fails and what this policy asks of them. If AI output that was wrong reaches a client or HMRC, or client information goes into the wrong tool, tell [Name] the same day. You will not be in trouble for reporting a mistake. You will be for hiding one. [Name] decides what we tell the client and our professional indemnity insurer.

*Where it comes from:* PCRT paragraph 3.3 requires members "to ensure that staff receive appropriate training for any AI tool used." ACCA asks for "Procedures for management of adverse incidents relating to the use of AI, including escalation processes." Whether your cover responds is a separate question, and I have gone through it in [does business insurance cover AI mistakes](https://business.theaieducator.io/posts/business-insurance-cover-ai-mistakes).

## Which jobs can AI help with, and what has to happen first?

Most of them, and the check gets heavier as the reader gets more official. The table is my reading of the PCRT guidance, set against the jobs a small practice does every week. It is not the professional bodies' table.

| Job | Can AI help? | What happens before it leaves the practice |
|---|---|---|
| Routine client email | Yes, in an approved tool | The sender reads it for accuracy and tone, and owns it |
| Notes from a recorded client meeting | Yes, if the client knows it is recorded | Someone who was in the room checks the notes are complete and accurate |
| Sorting and coding client records | Yes, in an approved tool | Sample check; anything odd is traced to the source document |
| Technical research on a tax point | Yes, as a starting point | Every case and section is opened and read; the trail goes on the file |
| Tax return or computation | Yes, with human review switched on | A qualified person reviews the data and the draft before anything is filed |
| Letter to HMRC | Draft only | The person signing reviews content and tone |
| Tax planning advice | Draft only | Checked against this client's facts and today's law; uncertainty is spelled out to the client |
| Tribunal submission | Only if every authority has been read in full | The person who read them puts their name to it |

Verdict: nothing in that table is banned, and nothing in it goes out unread.

## What about AI your client used, or AI built into your software?

Treat both as arriving from outside, because they did. The policy covers what your people type. These two walk in through other doors.

Clients now send in spreadsheets, summaries and even draft computations that a chatbot produced for them. PCRT expects this: "Members should exercise due care and professional scepticism when data is provided by a client or a third party which may have been generated by AI." Ask where the figures came from. It is the question you would ask of a carrier bag of receipts.

Then there is the software on your own machines. PCRT gives the example of "tax-return software that can automate the preparation and submission of a tax filing and can be configured to make human review optional." Optional! Find that setting in every product you run, and write down who is allowed to change it. The same goes for meeting recorders, which I have covered in [AI note takers in a small business](https://business.theaieducator.io/posts/ai-note-taker-small-business).

Law firms face a version of all this with privilege and court citations added; [the AI policy for a small law firm](https://business.theaieducator.io/posts/ai-policy-for-a-small-law-firm) is the sister piece to this one.

## What Good Looks Like

Good is three people in the practice giving the same three answers. That is the measure I would hold a finished policy to, and it has a name.

> The Readiness Test is three questions an organization must be able to answer consistently, from the top to the front line, before it can call itself AI ready: What is AI for here? What may I do with it? Who decides when it goes wrong?

Ask a partner, a trainee and the person who does payroll, separately. If the trainee says "we're not really allowed" while the partner has been drafting client letters in a chatbot since March, you have a document and no policy. Clauses 2 and 3 answer the second question. Clause 7 answers the third. The first is yours alone: decide which jobs in the table AI is for in your practice, and say so out loud. I have written a script for that conversation in [how to tell your team about AI](https://business.theaieducator.io/posts/how-to-tell-your-team-about-ai).

Clause 4 needs a harder look, because "a partner reviews it" is the easiest sentence in the policy to write and the hardest to mean.

> The Loop Test is three questions I ask before accepting "a human checks it" as an answer: Would that person notice a wrong result without being told to look for it? Would they notice in time to stop it mattering? And if they missed it, could the work be put back? Oversight that survives all three is a safeguard. Oversight that fails any of them is a signature.

Take a hypothetical four-partner practice on January 30. A trainee has used an approved tool to draft forty covering notes for self-assessment returns, and one partner will sign them off before midnight tomorrow. Would she notice a wrong relief in note twenty-eight without being told to look? Probably not, unless clause 6 has flagged which notes the tool wrote. In time? Only if review starts before the last afternoon. Could it be put back? A filed return can be amended, but a client who has acted on bad advice cannot be un-advised. The honest fix is dull: fewer notes per reviewer, flagged drafts, an earlier start.

In the advisory work I have done with government bodies (the UK Department for Education, KHDA in Dubai and the Ministry of Education in Kazakhstan), I have held to the same line I would give a four-partner practice: a policy nobody can repeat from memory is not a policy yet.

## How do you put the policy in place this week?

In a week, an hour or so a day; that is the time I would set aside for a practice of six.

1. **Monday: list what is in use.** Ask everyone, without blame, which AI tools they have used on client work in the last month. Expect surprises. If the list is longer than you thought, read [what to do about shadow AI](https://business.theaieducator.io/posts/shadow-ai-small-business) before you say anything you regret.
2. **Tuesday: put the four supplier questions** to your two most-used tools, in writing. What data, stored where, kept how long, used for training?
3. **Wednesday: fill in the seven clauses.** One name in every square bracket. A sole practitioner writes their own name seven times, and that is fine.
4. **Thursday: add the AI sentence to your engagement letter** for new clients, and decide when existing clients will get it.
5. **Friday: read the policy aloud** at the team meeting. Ten minutes. Then find the "human review optional" setting in your tax software together.
6. **A week later: ask three people the three Readiness Test questions.** Fix whichever answer differs.

## Where to go from here

If your team is already using these tools and nobody has shown them how to check what comes back, that is what my [AI training for business teams](https://theaieducator.io/ai-training-for-business?utm_source=business.theaieducator.io&utm_medium=referral&utm_campaign=ai-policy-for-an-accountancy-practice) is for. For a quick read on where the practice stands before you write anything, the free [Business AI Readiness Scorecard](https://theaieducator.io/business-ai-readiness?utm_source=business.theaieducator.io&utm_medium=referral&utm_campaign=ai-policy-for-an-accountancy-practice) takes about five minutes. And if you organize a district society or a practice network and want this put to a room of partners, it is one of the subjects I cover in [keynotes for business audiences](https://theaieducator.io/business-keynotes?utm_source=business.theaieducator.io&utm_medium=referral&utm_campaign=ai-policy-for-an-accountancy-practice). For governance work with the partners themselves, [start here](https://theaieducator.io/business?utm_source=business.theaieducator.io&utm_medium=referral&utm_campaign=ai-policy-for-an-accountancy-practice) or [talk to me](https://theaieducator.io/booking?sector=business&utm_source=business.theaieducator.io&utm_medium=referral&utm_campaign=ai-policy-for-an-accountancy-practice).

The software will keep getting better at sounding like a senior. Your policy is what keeps it a junior.

## Sources and further reading

- [Topical guidance covering the application of PCRT to the ethical use of artificial intelligence tools](https://www.icaew.com/technical/tax/working-in-tax/pcrt/topical-guidance-application-of-pcrt-to-ethical-use-of-artificial-intelligence-tools), the PCRT bodies, as published by ICAEW, January 19, 2026.
- [AI at the tax tribunal](https://www.icas.com/news-insights-events/news/tax/ai-at-the-tax-tribunal/), Susan Cattell, ICAS, January 26, 2026.
- [Key areas of risk when adopting AI](https://www.accaglobal.com/uk/en/technical-activities/technical-resources-search/2026/March/Key-areas-of-risk-when-adopting-AI.html), Julie Matheson, Ian Ko and Sam Binymin of Kingsley Napley LLP, for ACCA, March 2026.
- [Sole practitioners and small firms, different futures: the choices shaping accountancy practice](https://www.icaew.com/technical/practice-resources/practice-news/sole-practitioners-and-small-firms), Kat Hearn, ICAEW, November 16, 2026.

*Dan Fitzpatrick is The AI Educator: a keynote speaker, Forbes contributor and five-times bestselling author who has trained more than 150,000 educators in more than 30 countries. [More about Dan](https://theaieducator.io/about?utm_source=business.theaieducator.io&utm_medium=referral&utm_campaign=ai-policy-for-an-accountancy-practice).*

## Frequently asked questions

### Does an accountancy practice need a written AI policy?

Yes, in practice. No UK rule names a standalone AI policy, but the PCRT guidance of January 2026 holds firms responsible for all work whatever tool produced it, and ACCA says all firms should have firmwide policies on responsible AI use. A short written policy is the simplest way to show that care was taken.

### Can accountants put client data into ChatGPT?

Not into a public or personal account without the client's consent. The PCRT guidance says entering client data into publicly available AI tools is likely to breach client confidentiality unless the client has agreed. Use tools the practice has approved, on accounts it controls, and keep identifying details out of everything else.

### Do accountants have to tell clients they use AI?

The PCRT guidance says members should consider it, and suggests a statement in the engagement letter that AI-enabled software may be used. Where AI is fundamental to a piece of work, it suggests telling the client before starting. A client who asks how a figure was reached should always get an answer.

### Who is responsible if AI makes a mistake in a tax return?

The accountant and the firm. The PCRT guidance says members are ultimately responsible for any work they produce, irrespective of the use of AI tools in its creation. In Elden v HMRC the tribunal said the human who relies on AI bears the responsibility for its accuracy.

### What is the Junior Rule?

The Junior Rule is Dan Fitzpatrick's suggested rule for accountancy practices: treat anything an AI tool produces as a first draft from a bright junior in their first week. Someone qualified reads it, checks every source and figure, and puts their own name to it. It follows paragraph 3.4 of the PCRT AI guidance.

### How long should an accountancy practice's AI policy be?

Short enough to read aloud at a team meeting. Seven clauses of two to four sentences each will cover a small practice: scope, approved tools, client information, review, clients, internal disclosure, and training and mistakes. A policy the team can repeat from memory does more than a long one in a folder.

---
Source: [AI Policy for an Accountancy Practice: 7 Clauses to Copy](https://business.theaieducator.io/posts/ai-policy-for-an-accountancy-practice)
Publisher: [The AI Educator](https://theaieducator.io)
