An AI policy for accounting firms: a one-page template you can adapt

Why your firm needs one now

Your staff are probably already using AI. In accounting forums, people describe colleagues pasting client information into public chatbots, and others spending their time policing it. A policy turns that into a clear set of rules, and it tells people what they can do, which is what gets a policy followed.

Accountants also have duties that follow client information into any tool: the AICPA confidentiality rule, IRC §7216 for tax return information, and the requirement for a written information security plan that covers your service providers (IR-2026-92). The guide to putting client data into ChatGPT explains each one. The template below is built around them.

The template

Replace the parts in [square brackets]. Each clause is followed by a note on why it's there, which you can delete from your own copy.

1. Scope

This policy covers any use of AI tools for firm or client work by anyone working for [Firm], on firm or personal devices. AI tools include chat assistants, AI features built into software we already use, and browser extensions that read or write content.

Why: AI is turning up inside software people already use, and personal devices and accounts are where unapproved use happens.

2. Approved tools

Only the tools below may be used for firm work, on the plans and for the uses listed:
- [Tool], [plan], approved for [uses], [client data allowed: yes / no]
- [Tool], [plan], approved for [uses], [client data allowed: yes / no]

Ask [owner] before using any other tool, including free trials and browser extensions. Sign in with your firm account, never a personal one. [Owner] reviews this list every [six months].

Why: the same tool can have very different terms on a personal plan and a business plan. Approve a plan and an account type, not just a product name. The Copilot guide shows why the account you sign in with matters.

3. Client and firm data

Information falls into one of three groups:
- Fine: anything that doesn't identify a client or reveal their private information, such as made-up sample data, public guidance, or wording with names removed.
- Anonymize first: real work where identifying details can be taken out. Remove names, account and ID numbers, and identifying file names before upload, and keep the key outside the tool.
- Never, in any tool not approved for client data: Social Security numbers, EINs, bank and card numbers, full tax returns, payroll with names, passwords, and anything under a confidentiality agreement.

Tax return information may need the taxpayer's written consent before it goes to an AI provider. Check with [owner] before using AI on tax work.

Why: the AICPA rule bars disclosing confidential client information without consent, and under the §7216 regulations tax return information includes more than names and numbers (26 CFR §301.7216-1). Three groups are easier to apply on a busy day than a long list of rules.

4. Checking AI output

AI output is a first draft. Before it goes into the books, a workpaper, a return or anything a client sees, the person using it checks it against the source: figures tied to the source data, every citation looked up in the original, and every summary read against what it summarizes. The preparer is responsible for the work, whether or not AI helped. [Describe any extra review, such as a second reviewer for client-facing figures.]

Why: AI makes convincing mistakes, and responsibility doesn't move to the tool. For tax practitioners, Circular 230 requires due diligence in determining the correctness of what you tell clients and the IRS (31 CFR §10.22). The guides on flux commentary and tax research show what the checks look like.

5. Telling clients

[Choose your approach, for example:] We describe our use of AI tools in our engagement letter. If a client asks us not to use AI on their information, we record that on their file and follow it.

Why: the AICPA's interpretation on outside service providers expects either a contract that keeps the information confidential or the client's consent (Journal of Accountancy). How a firm handles disclosure is a decision for its leadership, ideally with advice. Whatever you decide, write it down so everyone does the same thing.

6. Records

AI-assisted work is documented like any other work: what was asked, what the AI produced, what was checked, and by whom. Work products go in the client file, not only in an AI tool's chat history. Chats containing client information are handled under our record retention policy.

Why: a reviewer or an auditor will ask how a figure or a conclusion was reached, and a chat history in one person's account isn't a workpaper.

7. When something goes wrong

If client data goes into a tool it shouldn't, or AI output with an error reaches a client or the books, tell [owner] the same day. Don't try to fix it quietly. [Owner] handles it under our written information security plan.

Why: the sooner the firm knows, the more options it has. Your security plan should already say how incidents are handled; IRS Publication 5708 is a template for the plan if you need one.

8. Training and review

Everyone who uses AI for firm work completes [training] before they start. [Owner] reviews this policy every [six months] and whenever a tool is added or its terms change.

Why: a policy nobody has been shown how to follow doesn't change what people do, and AI tools and their terms change quickly.

Rolling it out

1. Name an owner. One person who keeps the tools list, answers questions and handles incidents.
2. Start with the approved tools list. It does more than anything else in the policy, because it gives people a safe option.
3. Walk the team through the three data groups using real examples from your own work, not abstract rules.
4. Show people the checks, not just the rules. A short session on testing a formula, tying a total and looking up a citation does more than a reminder to be careful.
5. Revisit it after a month. Ask what people are actually using AI for, and update the list.

Common gaps in firm AI policies

• No owner, so nobody updates the tools list and questions go unanswered.
• Approving a product but not a plan, so personal accounts on the same tool slip through.
• Banning everything. A blanket ban can push use onto personal accounts, where the firm has no agreement and no visibility.
• Saying nothing about checking output. A data rule alone doesn't stop a wrong figure reaching a client.
• Forgetting AI built into existing software, from email to accounting packages.