Golden Sky AI

What should a small CPA firm's AI policy say?

Which AI tools are approved, what client data goes into each, who checks the work, how clients are told and billed, and who keeps the policy current.

Golden Sky AI12 min read

Your AI policy should name the AI tools your firm has approved, say which client data may go into each one, and name who checks every output before a client sees it. It should also cover how clients are told and asked for consent, how AI-assisted time is billed, what staff do when something goes wrong, and which partner owns the policy and keeps it current as models and rules change.

Does a small firm really need a written AI policy?

Yes, and the IRS has now put that in writing. Its Office of Professional Responsibility said in guidance dated June 24, 2026 that "firms must deploy internal policies and procedures for compliance with Circular 230 in the AI space" (IRS OPR Issue 2026-19), and Circular 230 makes that the job of the person who runs the firm's tax work, who "must take reasonable steps to ensure that the firm has adequate procedures in effect" (31 CFR 10.36). In most small firms, that's the managing partner.

Most firms haven't got one yet. Karbon's survey of nearly 600 accounting professionals, published in January, found just 21% had an AI policy or strategy in place (Karbon), and among the firms already using AI in Financial Cents' summer survey of 486 professionals, 87% had no formal written policy (Financial Cents). Their people are still using AI, though, on whatever they've got, and you'd be surprised how many put client data into free or personal accounts the firm has never seen.

If you want to know whether that's going on inside your firm, the quiz below tells you in ten questions.

What should each section of the policy say?

A good policy is one a seasonal preparer can read over a coffee and still know exactly what they can and can't do.

  • It covers every partner, employee, seasonal preparer and contractor, and every AI they touch, from ChatGPT and Claude to Copilot inside Microsoft 365 and the assistant in your tax software.
  • It names the approved tools and the plan for each, such as the firm's ChatGPT Business or Claude Team workspace, explains how to ask for a new one, and says free and personal accounts never see client data.
  • It sorts client data by how sensitive it is and says where each kind runs. We implement AI in three ways, and the policy uses the same split: the most sensitive client work stays on a computer in the office and never reaches any hosting or cloud, the next tier runs inside the firm's own cloud account, and only non-sensitive work goes out to Claude's or ChatGPT's own servers.
  • It says what clients are told and when they must agree, since a return put into an outside AI usually needs a signed Section 7216 consent first, and a Journal of Accountancy column suggests AI disclosure language in engagement letters (Journal of Accountancy, July 2026).
  • Selling or handing over client data to train anyone's model is ruled out, since the AICPA's ethics division warned on November 2 that "redaction does not, by itself, turn confidential client information into nonconfidential information" (Journal of Accountancy, November 2026).
  • Every AI solution gets an owner who reviews what it produces, because OPR says "due diligence requires verifying the accuracy of facts, citations, and calculations produced by AI" (IRS OPR), and the client file records the prompts, how outputs were checked and who reviewed them, as the same Journal of Accountancy column suggests.
  • Time is billed as it was worked, because OPR warns against "billing clients for manual labor or time that was not actually spent" under section 10.27.
  • It says what staff do when something goes wrong, like a return pasted into the wrong tool, and ties that into the written information security plan the IRS reminds tax pros they must keep (IR-2026-92). The FTC's Safeguards Rule, which covers tax preparation firms, says a breach of unencrypted data on 500 or more consumers goes to the FTC within 30 days (FTC).
  • It names the partner who owns the policy, gets everyone it covers to sign that they've read it, and is reviewed at least once a year and whenever a model is retired, a vendor changes its terms or a rule changes.

Why does a template policy fall short for a CPA firm?

The AICPA released a Small Firm Generative AI Policy Template for members on April 29, 2026 (AICPA & CIMA), and it shows the shape a policy should take. What it can't know is which tools your people use, where your client data sits and which bodies your firm answers to: peer review for audit work, Circular 230 and Section 7216 for tax, your state board and malpractice insurer, and state AI laws, since OPR notes that "several states, including California, Colorado, Illinois, and Utah, have enacted AI governance legislation" (IRS OPR).

So we write a firm's policy from what its people really do. Before a word goes down, the partners and everyone on the staff tell us about their work, the AI they use and the data that goes into it. Every guardrail is then designed into each solution before it goes in, under a policy written for the governing bodies and regulators that particular firm works under, so it covers the tools people open on a Monday morning, including the ones nobody mentioned until we asked.

The quiz below runs through ten questions on which AI your team uses and where client data ends up, and your answers show which parts of this policy you already have.

Where does the AI policy fit in a firm's whole AI process?

A policy on its own won't make a firm good at AI. It comes out of the whole process we take every firm through, and then it governs everything that follows.

That process opens by making the firm compatible with AI. We check its AI layers, meaning the business itself, each department, the systems and data each department works from and its people, and put right whatever's weak in the tech stack, the data or how it's secured, which is when free and personal accounts get shut down. Then we map every task every person does, its hours, that person's hourly rate, where AI already helps and the new services AI lets the firm offer clients, along with each job's budget and how often it overruns. That map shows which tools are in use, on what data and by whom, so most of the policy comes out of it.

Every possible solution is then costed for the hours it frees, what implementing it costs and its first-year return, and checked for data security, compliance and governance against each regulator the firm deals with, which is where each solution's line in the policy gets written.

The money comes through the realization rate, which UK practices know as the recovery rate. Each job's fee is set from the hours budgeted for it, so when a job overruns, the chargeable time past the budget gets written off and the firm collects less than the time it put in. When AI keeps a job on budget or brings it in early, the firm gets paid for every hour it bills with nothing written off, so the same work earns more, gross profit improves and the team has room for more clients. That's why we cost every solution against those job budgets, as we explain in how AI moves a CPA firm's realization rate. On tax work the gain is the write-offs that no longer happen and the extra clients, never billing hours nobody worked, which is the line OPR draws under section 10.27.

Here's how one line of the map gets costed, using a made-up tax manager, so it isn't a client result. They bill $210 an hour and spend 3 hours a week drafting replies to IRS notices, which over 48 weeks is 144 hours. We multiply that by 0.7 to 0.85 to keep room for review and edits, giving roughly 101 to 122 hours of capacity unlocked, worth about $21,200 to $25,700 at that billing rate. The notices carry return information, so the policy keeps that solution in the firm's own cloud account or on an office machine, and the manager signs off each reply. We cost every task of every person this way and then ask what each freed hour becomes, because capacity that does nothing for the business isn't worth unlocking.

The solutions then go on a roadmap that sets out which ones go in, the order and the method, most valuable first, with the firm's own AI policy over every one of them, and alongside it an operating model shows where AI is working, who it's working for and the capacity it has unlocked for each team. Implementation comes last, solution by solution in roadmap order, with an enablement pack for everyone who uses each one, and every solution and the policy are kept current each time a model or a rule changes.

How much work goes into a policy that holds up?

It takes more than most partners expect. Everyone in the firm, seasonal staff included, has to be asked what they do, which AI they use and what client data goes in, and each department's systems gone through to find where returns, payroll files and bank statements sit and travel. Every rule the firm answers to has to be read for what it says about AI, including Circular 230, Section 7216, the AICPA code, the Safeguards Rule, the state board, the insurer's terms and any state AI law, and every vendor's terms checked for training, retention and where the data is processed.

Then each possible solution has to be costed and checked against all of that, the policy written to match, everyone trained and signed off, and the whole thing kept current. The partners who know enough to write it are the ones whose billable hours keep the firm running, which is why many firms ask whether an AI consultant is worth it. When we take it on, your people each give us a few minutes on what their work involves, and everything after that is on us.

If you want us to do it for you, book a free call and we'll find you at least $50,000 a year in new business that AI can unlock. It's compliant, data-secure and has governance built in.

What can your firm put in place this week?

These steps lower the risk before the full process is under way.

  • Ask every person, partners included, which AI they use for client work, including free accounts and phone apps.
  • Tell everyone in writing, today, that client data stays out of any AI account the firm doesn't control.
  • Get each paid AI provider to confirm in writing whether your data trains its models, how long it's retained and which country it's processed in.
  • Name the partner who owns AI at the firm, because Circular 230 lets the IRS pick someone if the firm hasn't (31 CFR 10.36).
  • Add the AI tools your people use to your written information security plan.

Before you draft a word of the policy, the quiz below shows you in ten questions which of these gaps your firm has today.

Common questions

Can we just ban AI at our firm instead?

You can, but in our experience a ban with no approved tool pushes the use onto personal phones and free accounts, where nobody at the firm can see it. A short list of approved business tools with clear rules on client data is the safer position, and it fits OPR's instruction to use "only secure, enterprise-approved AI" (IRS OPR).

Does the policy have to name ChatGPT, Claude or Copilot?

Yes, it should name each tool and the plan the firm pays for, because the risk sits in the account, and a Section 7216 consent has to name its recipient anyway. When a solution moves to another model, the policy and any consents move with it, which is far easier when the firm owns its own system and only rents the model.

How often should a CPA firm update its AI policy?

Review it at least once a year, and again whenever a model you rely on is retired, a vendor changes its terms or a rule changes. The Safeguards Rule already has the person running your information security program report in writing at least annually (FTC), so both can be reviewed at the same sitting.

Should our clients see the AI policy?

They don't need the whole document, but they should know AI may be used on their work, which is what the engagement letter wording is for. If your insurer, a client or the state board asks how the firm uses AI, the answer is already written down, and a one-page summary is usually enough to send.

This is general information and not legal or tax advice, and your attorney is the right person to say how it applies to your firm.

Work with us to get an AI policy that fits how your firm really works

At Golden Sky AI, we help companies in regulated industries implement AI that frees up their team's capacity, so they can take on more business without having to hire. Every solution is checked for compliance, governance and data security before we implement it, and you own the intelligence, meaning any AI agent could plug into the solutions we implement, as they're robust and still deliver the same quality of work. That means you're never held hostage by one AI provider.

We'll find you $50,000 a year in new business AI can unlock in 72 hours, or you don't pay.

Sources

Field notes

Get the field notes.

What is working with AI inside professional service firms, written from the buyer’s side.

Sent when there is something worth reading, not on a schedule.

One email when there’s something worth reading. Unsubscribe any time. See our privacy notice.

Find out what AI would earn your firm, risk-free.

Book a call and we'll show you where AI can unlock $50,000 a year in new business for your firm.

Book a call

See where you rank