AI Policy for Accounting Firms: The Template You Can Fill In Today

An accounting firm’s AI policy is a short internal document that says where the firm uses AI, what data may never go into an AI tool, and who reviews AI output before it reaches a client. It needs ten sections and it fits on three pages. The point is not compliance theatre. Your clients have already assumed you use AI, and the firms that can answer “here is exactly where, and here is who checks it” keep the trust. The firms that improvise when asked do not.

Most firms have not written one. That is the gap, and it closes in an hour.

The three US rules that already govern your AI use

If you operate in the United States, the IRS has now said in writing what it expects, and two older rules apply to AI directly.

IRS OPR Issue 2026-19. On 24 June 2026 the IRS Office of Professional Responsibility published Introductory Guidelines for Responsible AI Use in Federal Tax Practice. It adds no new rule. It explains how Circular 230 already applies:

• Review everything (§10.22, §10.37). “Practitioners cannot rely solely on AI.” Every AI-created document is reviewed before it reaches a client or the IRS, and every citation, fact and calculation is checked.
• Know the tool (§10.35). Competence now includes understanding how the AI you use produces its answers, and where it fails.
• Firm procedures (§10.36). Whoever oversees the firm’s tax practice must have AI policies in place: staff training, protocols for data handling and accuracy checks, and vetting of third-party AI tools, all documented. This is the policy this article is about.
• Fees (§10.27). Billing for time that was not actually spent, or double billing for AI-assisted work, may violate §10.27. The IRS says cost savings “should be passed on openly” and the AI work disclosed.
• Client data (§10.51(a)(15)). “Practitioners must strictly handle all client data using only secure, enterprise-approved AI.”

IRC §7216 makes it a criminal offence to disclose tax return information without the taxpayer’s written consent. The exceptions are narrower than most firms think: one covers another preparer who makes no substantive determinations, the other covers contractors who maintain or repair your tax software, and only if each person receiving the data gets a written notice of the rule. A general-purpose AI tool that analyses a return or drafts advice does not clearly fit either, and the IRS has not said that it does. Three details catch firms out:

• Location means the person, not the server. No rule requires tax data to stay in the US. But if anyone who can see the data is outside the US, including a vendor’s support staff, you need consent before the disclosure, even when the data never leaves a US server.
• Consent for individual (1040) clients is its own document. It can be attached to your engagement letter, but it cannot be a clause inside it, and it must use the wording Rev. Proc. 2013-14 requires. For business clients, the engagement letter itself can carry it.
• Consent is never retroactive, and consent with no stated duration expires one year after signing.

The FTC Safeguards Rule treats any accountant or tax preparer in the business of completing returns as a financial institution, regardless of firm size. You need a written information security program (IRS Publication 5708 is the template), a designated Qualified Individual, and encryption. Two requirements reach your AI tools directly:

• Multi-factor authentication “for any individual accessing any information system”. The rule counts any system “connected to a system containing customer information”. An AI tool, a connector or an agent that can reach client files is inside that line.
• Dispose of customer information no later than two years after it was last used, unless you need it for business or the law requires you to keep it. That applies to what sits in an AI tool’s history too.

Since 13 May 2024 you must also notify the FTC within 30 days when unencrypted information about 500 or more people is taken. Firms with fewer than 5,000 customers are spared only the written risk assessment, penetration testing, the written incident response plan and the board report, not the rest.

None of these rules was written for AI. All of them govern it.

What applies where

The EU’s transparency rules are in force too, though for most firms they bite only on what the firm publishes. They do not require a notice on every piece of client work. The US column has been live for years and is enforced more aggressively than most small firms assume.

The ten sections

An accounting firm’s AI policy needs these, in this order:

1. Who it applies to. Employees, partners, interns, and contractors. The most common gap is not the employee using an approved tool. It is the contract bookkeeper using a personal AI account on your client’s payroll file.
2. Where you use AI, and where you do not. Real tasks, not categories. “Reading data off invoices” and “drafting internal summaries”, not “improving efficiency”.
3. A person signs off. Nothing reaches a client without documented human review. This is the whole policy in one line. Everything else is detail.
4. What may never be entered. ID numbers, bank statements, payroll, client lists, signed agreements. And the part firms miss: removing the client’s name does not anonymise the file. A location, an unusual transaction, and an industry identify a business as reliably as its name does.
5. Approved tools, and how you vet new ones. Firm accounts only, and the vendor checklist below answered in writing before anything is added. AI agents get their own account, never a staff member’s login.
6. Subprocessors and consent. An AI vendor handling client data is a subprocessor. It goes on the list, under an agreement, and clients are told.
7. Your legal baseline. Keep it in an appendix so the policy does not go stale every time a deadline moves.
8. What you tell clients, and how you bill. One paragraph for the engagement letter, and a line on how AI-assisted work is priced. This is the only part a client reads.
9. When AI gets it wrong. Who is responsible (a named person, never “the system”), how errors are reported, what you keep to reconstruct what happened.
10. Who owns the policy, and when it is reviewed. Undated policies are wrong within a year.

Then print a one-page version for the team. The full document is for the file. The one-pager is what people actually follow.

Vetting an AI vendor: the checklist

The IRS now expects third-party AI tools to be vetted and the vetting documented. Ask every vendor these nine questions, adapted from the vendor checklist in the AI Security Guide for U.S. Accounting Firms by Jason Staats, and keep the written answers on file:

1. Training. Is our input or output used to train the vendor’s models? Business plans usually say no. Check that yours does.
2. Usage limits. What do the terms allow? Many software terms forbid bots and automated access.
3. Who sees the data, and from where. Which vendor staff can access it, and are any of them outside the US? For tax data that is a §7216 consent question.
4. Subprocessors. The full list, including the AI model provider behind the product. Your own disclosure covers the vendor, so the vendor’s AI provider must be on its list.
5. Audit report. A SOC 2 report or ISO 27001 certificate, and which period and scope it covers.
6. Retention and deletion. How long prompts, files and outputs are kept, and how you delete them.
7. Breach notice. How quickly the vendor tells you about an incident. You may have 30 days to tell the FTC.
8. Admin controls. MFA, single sign-on, role-based access, and a log of who did what.
9. Exit. How you get your data out, and proof that it is deleted when you leave.

A vendor that will not answer in writing has answered.

If you use AI agents

An AI agent does not just answer. It acts: it opens files, sends messages, fills forms. Jason Staats’s guide sorts its reach into three levels, and each needs its own decision:

• A closed room. The agent works only on what you hand it, inside one application.
• An open browser. The agent can visit sites and use web applications on your behalf.
• The whole computer. The agent can read local files, run programs and use whatever the machine is logged into.

Do not run any of them on default settings. Decide which level each task gets, and write it down.

Give every agent its own account, never a staff member’s password. A shared login breaks per-seat licence terms, and it breaks your audit trail: when something goes wrong, you cannot tell what the person did from what the agent did. Under the Safeguards Rule, the agent’s account needs the same MFA as everyone else’s.

The paragraph to send your clients

Most of the value is here, and it takes ten minutes:

We use AI tools to assist with parts of our work, including [short list]. These tools help us prepare drafts and organise information. They do not make decisions about your accounts or your business. Every deliverable is reviewed and approved by a member of our team, who remains responsible for it. We do not enter your confidential information into public AI tools, and any AI provider that processes your data is bound by a written agreement. Where AI saves time on your work, our fees reflect it. You can ask us at any time how AI is used in your work, and you can ask us not to use it.

Send it to your three largest clients before they think to ask. Being early on this is worth more than being thorough about it.

What to do with the time AI gives back

There is a quieter question underneath the policy one. If AI handles more of the routine work, what happens to those hours?

Part of the answer is no longer yours to choose. For federal tax work, the IRS says billing for time that was not spent may violate Circular 230, and that savings should be passed on openly. Billing the hours a task used to take, after AI has cut them, is now a risk. A fixed fee for the outcome, with the AI work disclosed, is the cleaner model.

The rest of the time is where the relationship is won, and the answer that holds a client is attention: the call before the deadline rather than after it, the review nobody asked for, the plan that accounts for what the owner is doing next year. Clients notice speed for about a week. They notice being understood for a decade.

That is also why the human sign-off in section 3 is not a brake on AI. It is the thing you are selling.

Uku is practice management built for that model: the routine work tracked and automated, the client relationship left where it belongs. See what a firm running on Uku looks like, or read how our API was designed for AI agents from the start — the same API that now lets you connect the AI assistant you already use directly to your firm’s data.

The template, ready to copy

Nothing to download and no form to fill in. Copy the document below, paste it into Word or Google Docs, replace everything in brackets, and delete what does not apply to your firm. Budget an hour, and expect section 2 to take most of it, because that is the one where you have to ask your team what they are actually using.

This is a starting point, not legal advice. Your obligations depend on where you operate and what work you do. Have a qualified advisor review the finished document before you rely on it.