ICAEW AI Guidance for Small Practices: A Client Notice and Staff Checklist

ICAEW's generative AI guide, written for accountants, asks a small practice to do five things: keep confidential client information out of public AI tools, write a policy, be transparent about AI use, review outputs with professional scepticism, and keep responsibility with people. In practice that is one client notice and one staff checklist.

The guide is guidance rather than a rule book. It applies the five fundamental principles that the ICAEW Code of Ethics takes from the international ethics code (integrity, objectivity, professional competence and due care, confidentiality and professional behaviour) to a new tool, which is why it is a useful yardstick for a small firm with no compliance team. Each of its points becomes a line in the client notice, the terms-of-business paragraph or the one-page checklist below, and the places where your own adviser must decide are marked.

Follow me on Instagram@sagnikteaches

What the guide actually says, in the order a practice needs it

ICAEW's Generative AI guide is a set of about a dozen pages, running from a brief introduction (dated September 2023) through risks and limitations, getting started, using generative AI, prompt engineering, ethics, legal considerations, cyber security and do's and don'ts to a glossary. I read the using, ethics, getting started, legal, cyber security and do's and don'ts pages on 4 November 2026. Page contents change, so read the current guide before you rely on any wording here. The points that matter most to a small practice:

Connect on LinkedInSagnik Bhattacharya

• Confidentiality. The ethics page says respecting client confidentiality means not loading confidential information into public generative AI tools, even where the information has become publicly available. It notes limited visibility and control over who the information is shared with, how it is secured and how long it is retained.
• Personal data. The legal considerations page is stricter than the ethics page: it says organisations should not put confidential information or personal data into a generative AI tool, because that may not be allowed under data-protection law or may breach a duty of confidence owed to someone else, and the organisation cannot control how the data is then used.
• Terms and data handling. The using-generative-AI page asks you to consider your own terms and the provider's, plus data location, consent and secure data handling such as access controls, anonymisation and encryption, and compliance with data-protection rules.
• Policy. It says easy and free access to tools requires organisations to say what is and is not allowed, and recommends a policy with principles, guardrails or restrictions.
• Transparency. Accountants should be transparent about AI use by labelling or acknowledging where output was produced by generative AI.
• Review. Review, challenge and critique all generative AI outputs, as you would a member of your team or output provided by an audit client; professional scepticism applies equally to AI output.
• Responsibility. For your own use of AI, the do's and don'ts page is the plain line: do not abdicate responsibility, since outputs need human oversight. Separately, where a firm helps a client design or run AI systems, the ethics page warns against assuming management responsibility for the decisions and controls around them; that responsibility should stay with the client.
• Start small. The do's list recommends small projects that are not business critical before scaling, and monitoring outputs against expected results.

I have not reproduced the guide's wording beyond short phrases. The cyber security page adds two points a small practice can act on at once: keep a list of approved tools and discourage staff from using anything else, and plan how you would respond to an AI-related incident before one happens. The legal page is the one to read in full with your adviser, because its line on personal data is stricter than many practices' current habits. For the wider question of what clients should be told, AI for chartered accountants: what to tell clients covers a longer discussion.

Subscribe on YouTube@sagnikteaches

The five decisions every small practice has to make

Before drafting any notice, a partner needs to decide five things, because the notice just reports them.

1. Which tools are allowed. Name them. "Approved tools: [tool] on the practice's business plan." Everything else is not approved for client work.
2. What may go in. Use three bands. Never: client names, identifiers, tax references, bank details, anything under a confidentiality undertaking. Only with checks: figures and documents anonymised so that no person or client can be identified. Free to use: templates, public rules, your own training material.
3. Who reviews. A named reviewer for any AI-assisted work that reaches a client, at the same level of seniority you would apply to a trainee's work.
4. Where it is recorded. A line in the job file saying AI was used, for what, and who reviewed.
5. What the client is told. The notice below.

Choosing the tools is the stage with most small-practice risk. The tools tutorial for chartered accountants lists candidates, and you should check each vendor's current business terms, including whether content is used for training, where data is stored and whether a data-processing agreement is available.

A client notice you can adapt

Keep it to one page, in plain words, and put it on your website and in your engagement pack. This is an illustrative draft, not legal advice; your adviser should check it against your own terms.

How we use artificial intelligence

We use generative AI tools to help with some administrative and drafting tasks, such as preparing templates, summarising public guidance and drafting routine letters.

What we do not do: we do not put your confidential information into public AI tools. Where we use an AI tool on work connected with your affairs, we first remove anything that identifies you or anyone else, we use a business-grade service whose terms we have checked, and we tell you first where required.

Who is responsible: every piece of work we send you is reviewed by a member of our team who takes responsibility for it. AI does not make decisions about your affairs.

If you would rather we did not use AI on your work, tell us and we will note it on your file.

Questions? Contact [name] on [phone]. We review this notice every [six] months.

Each paragraph traces to a point in the guide: transparency, confidentiality, responsibility and review. The opt-out line is a practice decision, not a requirement; include it only if you can honour it. A promise you cannot keep is worse than silence.

Where the notice goes

Three places: a page on your website, a short summary in the engagement letter, and a reply template for client questions. When a client asks "do you use AI?", a partner should be able to answer in two sentences from the notice, not improvise.

Terms-of-business wording

Your terms of business are where the notice becomes binding. Ask your solicitor or adviser to approve final wording, but this gives you a starting point to discuss, again illustrative:

Use of technology. We may use software, including artificial intelligence tools, to assist us in providing our services. We remain responsible for the work we deliver and review it before it is provided to you. We will not input your confidential information into any AI tool unless the tool is covered by terms that protect confidentiality and we are satisfied it is appropriate for that information. Our current approach is set out in our AI notice, available on request.

Two notes on it. First, "AI notice, available on request" means you can update the notice without reissuing every engagement letter. Second, the wording promises review and responsibility, which are the two things a regulator, an insurer or an unhappy client will ask about first. If your terms already have clauses on confidentiality, data protection (including the UK GDPR where it applies to you) and the use of subcontractors or processors, check that any AI tool provider fits under those clauses, because a cloud AI service is a processor in many setups. For how the other side of the table words this, see mentioning AI use in client contracts and proposals.

The staff checklist: one page, tick before sending

Staff need something they can follow at 5pm on a busy day. This checklist is built from the guide's points and is meant for the desk, not the policy folder. Print one copy for each person.

Before you use AI on a task

• Is the tool on the approved list? If not, stop.
• Does the prompt contain a client name, reference, bank detail, address or anything identifying? Remove it, or do not use AI.
• Is this a task on the "free to use" band (templates, public rules, your own notes)? If it is a client-data task, have you checked with a partner?
• Would you be comfortable if the client saw exactly what you pasted?

Before the output goes anywhere

• Have you checked every figure, rule, date and citation against a primary source? AI tools can give answers tailored to a particular jurisdiction without making that clear, as ICAEW's guide notes about accounting, audit, tax and legal content.
• Is any statement of law or standards current? Many tools are not up to date with the latest legislative developments.
• Has a second person reviewed work that goes to a client?
• Have you noted AI use in the file: tool, task, reviewer, date?

If something goes wrong

• If client data went into a tool not on the list, tell a partner the same day. Do not delete quietly and hope.
• Write down what happened, which tool, what data and when. The partner decides whether to notify the client or anyone else, with advice.

The point of the "tell the same day" line is cultural: staff who fear a telling-off hide mistakes, and a hidden data slip is much worse than a reported one. Rolling out an AI policy so staff actually follow it covers how to make that stick.

A six-person practice walked through

An illustrative practice has two partners, three accountants and an administrator, with about 180 clients, mostly sole traders and small companies. The partners meet for 90 minutes and settle the five decisions: one approved business-plan tool, three data bands, a named reviewer on anything client-facing, a one-line file note, and the notice. Two weeks later, the notice is on the site and the checklist is on every desk.

Then three real tasks run through it.

• A client letter explaining a late-filing penalty. The accountant asks AI to draft the explanation using no client name and no figures, only "a company filed late". The draft says a fixed penalty and a cut-off date. The accountant checks both against the current official guidance, finds the date wrong, corrects it and records the check. This is the verification step doing its job.
• A staff training note on a rule change. Green band, public material. The administrator drafts the note with AI and a partner reads it before it circulates. No file note needed because no client data is involved.
• A client's bank statement. Raw statement with account numbers. Red band, never in any tool. The accountant uses the practice's accounting software's own features instead and does not paste anything into a general assistant.

Notice that two of the three tasks are low-risk and one is refused. That mix is the point: a policy that says yes to everything is not a policy, and one that says no to everything gets ignored.

Questions to put to an AI vendor before you approve it

The getting-started page of the guide includes a step on vendor or supplier evaluation, and encourages probing questions about training data, how the model operates, accountability and bias. For a small practice, that becomes a short written questionnaire you send before any tool joins the approved list, and file the replies. Questions that earn their place:

1. Is customer content used to train or improve models, by default or by setting? Where is that setting, and can an administrator lock it?
2. Where is data stored and processed, and can you choose a region?
3. How long are prompts and outputs retained, and can we delete them on request?
4. Is a data-processing agreement available, and who are your sub-processors?
5. What access controls does the plan offer: single sign-on, admin controls, audit logs?
6. What happens to our data if we cancel?
7. Does the plan document any limits, such as how many seats or a minimum number of users?

Record the date and the document the answer came from, since vendor terms change. In an illustrative review, a practice found that the same assistant had two different data-use positions: one on its personal plans and one on its business plan. That finding settled the question of which plan staff may use, and it went into the policy as a single line: "business plan only, with company login".

Task bands in a table

A policy is easiest to follow when staff can look up their task. Here is an illustrative table for a small practice; adapt the bands to your own risk appetite and your insurer's view.

The last row follows the guide's warning not to abdicate responsibility. AI can help a person find the relevant rule, but the advice has the person's name on it.

What to say when a client asks about AI

Clients ask in three ways: a casual "do you use AI?", a formal questionnaire from a larger client's procurement team, and an anxious "is my data in ChatGPT?" Prepare an answer for each and keep them next to the notice.

• Casual: "We use AI tools for some drafting and admin, always reviewed by a person, and we don't put your confidential information into public tools."
• Formal questionnaire: a copy of the notice, the approved-tool list with vendor terms, and the review procedure. Keep this pack ready as a PDF.
• Anxious: a direct answer: "No, your information has not been entered into public AI tools. If that ever changes for a specific task, we will ask you first." Only say this if it is true. If you cannot be sure, say what you have checked and what you will check next.

A file note that survives a review

When AI helps with client work, the file note should be short enough that people actually write it. An illustrative example:

AI use note - Client ref [xxxx] - 13 Nov
Tool: [approved tool], practice account.
Task: drafted explanation of late-filing penalty (no client identifiers entered).
Checked against: official guidance page, accessed 13 Nov; corrected the penalty date.
Reviewer: [name]. Sent to client: yes.

Four lines, written in under a minute. If a client, an insurer or a reviewer later asks how the work was produced, the answer is in the file. It also trains staff to remember that a check happened, which is the whole point.

Four slips that quietly break a small practice's AI policy

From the pattern of risks the guide describes, these are the four slips to watch for. First, a free personal account used "just for a quick query" with a client's name in it. Second, over-reliance: a polished paragraph that sounds right and is not checked, which the guide describes as automation bias. Third, out-of-date rules, because the tool does not know the latest legislative changes. Fourth, a notice that promises more than staff do: if the notice says every document is reviewed and a busy week skips it, the gap between the words and the practice is the problem. Test the notice every quarter by pulling three AI-assisted files and checking each has a review recorded.

Supervision: making review real

The guide's phrase about reviewing AI output as you would work from a team member suggests a practical test: what would you check on a trainee's draft? You would check the numbers, whether the conclusion follows, whether the law cited is current and whether the client's circumstances fit. Write those four questions at the top of the review step:

1. Are the facts and figures from a source I can name?
2. Is the law, rate or standard current?
3. Does this fit this client's situation, not a generic one?
4. Is there anything the tool has added that nobody asked for?

Track corrections for the first quarter. A simple tally of errors found in review tells you which tasks suit AI and which do not. If reviewers find a mistake in half of the drafts, the task is not saving time; stop doing it with AI or tighten the prompt.

Data protection and the processor question

When you use a business-grade AI tool on client data, the provider is typically a processor acting on your instructions. That has practical consequences: you should have a data-processing agreement in place, know where the data is stored, know whether it is used for training, and be able to delete it. The using-generative-AI page lists those same points under terms and conditions and data privacy, and the legal page's default is stricter still: no personal data into generative AI tools. If, with your adviser, you decide that a business-grade tool under a processing agreement is acceptable for anonymised work, write down why. Do these checks with the vendor's current documentation and an adviser, not from memory. Also decide whether your privacy notice to clients, which describes how you handle their data, needs a line about AI tools; if it names the processors you use, adding the AI provider is straightforward.

Keeping the policy alive

Put a date on the notice and checklist, and review them every six months or when you change a tool. The tools move fast; ICAEW's guide itself notes that many tools cannot draw on the latest legislative developments, and your list of approved tools will change too. A small review agenda: which tools are in use, any incidents, any client questions, any new task someone wants to try. If you want help testing a draft policy against the real jobs in your practice, that is what my AI implementation consultation covers. The notice and checklist above are enough to start, and they put you where the guide says to begin: small, careful and reviewed.

Questions small practices ask about ICAEW and AI

Is the ICAEW generative AI guide a rule I must follow?

It is guidance, not a rule book. It applies the Code of Ethics principles, which members must follow, to generative AI. That makes it a sensible benchmark for what a reviewer or client would expect. Read it alongside your own regulatory obligations and ask ICAEW or an adviser if you are unsure.

Can we use ChatGPT or Claude on client work at all?

The ethics page says not to load confidential client information into public tools, and the legal page goes further: no confidential information or personal data into generative AI tools. Using a business-grade tool on anonymised work, with checked terms and a processing agreement, is a decision to take with your adviser. Many practices start with templates and training notes.

Do clients have to be told we use AI?

The guide advises being transparent by labelling or acknowledging where output was produced by generative AI. Whether you need client consent depends on what data goes in and what your engagement terms say. Disclose in plain words in your terms and ask before using client data in any new tool.

Who is responsible if the AI gets it wrong?

The practice and the named reviewer, not the tool. ICAEW's guide says to review, challenge and critique AI outputs as you would work from a team member, and that responsibility should stay with people. Build a supervision step into every AI-assisted task.

Further reads

• Automate BACS Payments in an Accountancy Practice: Where AI Fits — Where AI fits in payment runs, and where it must not.
• AI Chatbot Consent Management: What a Professional Firm's Site Needs — Consent records for a professional firm's website chatbot.
• How Should Consultants Use AI? A Client-Data Checklist — A client-data checklist written for advisory firms.
• HR Consulting AI: Sell an AI Policy and Handbook Update Package — What a staff AI policy and handbook update contain.
• Can Lawyers Use AI Note Takers in Client Meetings? — Recording and summarising client meetings: the confidentiality line.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ICAEW Generative AI guide (index page plus the Using generative AI, Generative AI and ethics, Getting started, Legal considerations, Cyber security and Do's and don'ts pages, all read on 4 November 2026); data-protection points are general guidance, not legal advice.