For accounting firms using Microsoft 365

An AI-use policy for a small CPA firm: a starter guide.

A practical primer from FirmSteward. Short reads, no corporate feel.

An AI-use policy — sometimes called an AI acceptable-use policy — is a short written set of rules that tells staff which AI tools they may use, what kind of information they may enter into them, and who decides when something is off the table. It is a guardrail, not a law firm opinion. It is a set of boundaries a small team can actually remember, not a compliance filing.

This guide is aimed at a small CPA firm or a sole practitioner. If you have a handful of people using ChatGPT, Copilot, or other generative tools with client tax, payroll, or identity data in reach, a simple policy reduces the risk that sensitive information ends up somewhere you did not intend. Here is how to write one that is good enough to matter and short enough to be read.

Why a small firm needs one at all

Before the details, the reason. Authority is uneven: some staff experiment with public AI tools, others avoid them entirely, and nobody has a shared answer for what is okay. That is exactly the gap a policy fills. A short policy does not need to solve every risk to be useful. It needs to create one predictable answer to the most common question — can I put this in here or not?

When there is no rule, people do their best guess, and guesses vary. A clear boundary turns a gray zone into something staff can follow without asking for permission on every single task. That alone is worth the hour it takes to write.

Say no to over-engineering

Most small firms do not need a twenty-page document, a governance committee, or an approval workflow for every prompt. Over-engineering is its own failure mode: a policy nobody reads, because it is long, formal, and detached from how the firm actually works. A dense document also ages badly, because every outdated clause makes the whole thing easier to ignore.

Start with a page or two. One page is often ideal. The goal is not to cover every possible AI scenario — it is to give staff clear defaults, a short list of absolute prohibitions, and a named person to ask when something is unsure. If a policy does not fit on a screen, a busy team will not keep it in front of them. A short document you revise is far better than a long one you file and forget.

The core rules that matter most

Get the essentials right and the rest falls into place. For a small firm, five rules carry most of the value:

• An approved list of tools. Name the tools staff may use for firm work. Everything else is off-limits unless approved.
• A hard line on client data. State clearly what is never to be entered: full tax ID or Social Security numbers, client identity details, payroll data, and account numbers — unless you have a tool you have explicitly approved and vetted for that purpose.
• The no-trade-secrets rule. Nothing confidential about clients, the firm, or its people goes into a public tool without approval.
• Human review before anything is used. AI output is a draft, not a conclusion. Staff verify the work before it goes to a client or into a return.
• A question to raise a concern. One named owner, and a simple path for staff to ask "can I do this?" without friction.

Keep the wording plain. Instead of "artificial intelligence services," say "AI tools like ChatGPT and Copilot." Staff should recognize the thing being described without a glossary.

Define approved versus prohibited use

Vague policies die in the gap between "use common sense" and "here are the rules." Concrete examples make the difference. Give both a positive list and a negative one.

Approved use is the low-risk work: drafting a client email, summarizing a long document, drafting internal notes, brainstorming a process, or generating a first pass at a memo that staff will check. List the tools permitted for these tasks.

Prohibited use is the high-risk work: entering client identity or tax data into a public tool, pasting a file without confirming the vendor's data practices, using AI to auto-file without review, or relying on an unverified output for a financial conclusion. Name the hard no's plainly.

This "approved / prohibited" split is the heart of the document. It tells staff what they may do freely, what they may never do, and that everything in between goes to the owner for a quick yes or no. Keep the examples few enough to remember — five approved uses and five prohibited ones are plenty. If a list gets long enough to scan, it has stopped being a decision tool and become a wall of text.

Ownership and review

A policy with no owner is a policy that drifts out of date. Name the person responsible for AI decisions — at a small firm this is usually the owner or the managing partner. That person approves new tools, answers staff questions, and updates the document when things change.

AI changes fast, so schedule a lightweight review. A reminder every quarter — the tools you approved six months ago may not be the ones you use now — is enough. You do not need an annual compliance project. You need a check-in where the owner asks: is this list still right, and is anything happening in practice that it does not cover?

A sample short policy outline

Here is a one-page skeleton you can adapt. Keep the language yours, and adjust for your firm's size and client mix.

This outline is a starting point for an educational policy, not legal, compliance, or security advice. Have it reviewed by your own counsel and advisors before you consider it final — but do not let that stop you from writing a draft today.

Getting staff buy-in

A policy is only as good as whether people follow it, and people follow rules they helped shape and understand. Introduce it as a conversation, not a directive. Share the draft before it is final, ask staff what questions they already have about AI, and let them flag the gray areas you missed.

Frame it around the practical benefit: clear rules mean staff do not have to guess, and they are not personally exposed to a mistake they did not know was a mistake. Reinforce it in normal team meetings and when you approve a new tool, rather than once a year. And make it easy to ask — the person who asks before entering something risky is doing the right thing, and that should feel welcome, not like a demerit.

Finally, model the behavior you want. If the owner or manager ignores the policy in daily work — pasting client details into a new AI tool without checking — staff will reasonably read the document as optional. The single most reliable way to get buy-in is to treat the policy as something you visibly follow, not something you impose on everyone else.