Golden Sky AI

Do I need client consent under Section 7216 before putting tax returns into ChatGPT, Claude or Copilot?

Usually yes: a return in ChatGPT, Claude or Copilot is a 7216 disclosure. Here's what the consent must say, and why consent alone won't make AI safe.

Golden Sky AI12 min read

Yes, in most cases you do. Section 7216 treats making a client's return information known to anyone outside your firm as a disclosure, and OpenAI, Anthropic and Microsoft are outside your firm, so unless an exception covers the work, you need the client's signed consent first. Consent only settles one rule, though, so a free or personal AI account still shouldn't see a client's return, signed form or not.

What counts as a disclosure when your team uses ChatGPT or Claude?

The way the rule works is it starts with a very wide definition and then carves out exceptions. The regulations define disclosure as "the act of making tax return information known to any person in any manner whatever," and return information covers the client's name, address and identifying number and anything derived from the return (26 CFR 301.7216-1). So when someone on your team pastes a W-2 into ChatGPT or uploads last year's return to Claude, that information has been made known to OpenAI or Anthropic, and the summary that comes back is still return information.

The fines in the statute are small. A knowing or reckless disclosure can mean a fine of up to $1,000, or $100,000 where identity theft is involved, and up to a year in prison (26 U.S.C. 7216), and the civil penalty is $250 per disclosure, capped at $10,000 a year (26 U.S.C. 6713). We'd worry more about your license, because the IRS's Office of Professional Responsibility warned on June 24, 2026 that practitioners "must strictly handle all client data using only secure, enterprise-approved AI" and that "willful mishandling of taxpayer information through AI may lead to disciplinary actions under Circular 230" (IRS OPR Issue 2026-19).

If you want to know whether this is already happening inside your firm, the quiz below can tell you, and it only takes ten questions.

Which exception do firms rely on, and where does it run out?

Most firms point to 26 CFR 301.7216-2(d), which lets you disclose return information to another preparer in the United States to help prepare the return or provide auxiliary services for it, and the definitions count "a person who develops software that is used to prepare or file a tax return" as a preparer. That's why you don't collect consent before keying a return into your tax software.

Practitioners still argue over whether ChatGPT or Claude counts as that kind of provider, and nothing in the IRS's June guidance settles it. The exception also stops at "substantive determinations," and "a substantive determination involves an analysis, interpretation, or application of the law," so when you ask Claude whether a client's rental qualifies for a deduction, the regulation says you need the client's consent first. For individual clients you also generally can't get consent to send a Social Security number to a preparer outside the US, so it has to be masked (26 CFR 301.7216-3), which means you need to know, in writing, where an AI provider processes your data. Until the IRS says otherwise, we'd rather a firm get consent than find out later it was on the wrong side of the line.

Under 26 CFR 301.7216-3, the consent has to be in writing, signed and dated by the client before you disclose anything, and it has to name your firm, the client, the purpose, the specific recipient and exactly what information goes to them. If it doesn't say how long it lasts, it runs for one year from the date the client signed. You also can't make your services conditional on the client signing, so you need a way to do the return without AI if they say no.

For individual clients, Rev. Proc. 2013-14 sets the format, which is a separate document on paper at least 8.5 by 11 inches in 12-point type, or on screen at normal text size, opening with "Federal law requires this consent form be provided to you." A line in your engagement letter about technology tools doesn't meet that, although for business returns outside the 1040 series, "consent to disclosure can be in any format and can be incorporated into an engagement letter" (The Tax Adviser).

When Financial Cents asked firms already using AI how they handle client data, only 3% of the 359 answers mentioned client consent, disclosure or engagement-letter wording (Financial Cents).

Because an individual consent names the recipient, the forms you collect for OpenAI this season won't cover Anthropic if you move to Claude next year. That's one reason we make sure your firm owns the intelligence and isn't tied to one AI provider, and we're always going to tell you who the vendor is for that particular solution, so the consent can name it properly.

No, because consent only deals with Section 7216, and you answer to more than that. You won't believe how many people use free accounts, personal accounts, and put client data into it, and it's just a nightmare.

The AICPA's confidentiality rule asks for either a confidentiality agreement with an outside provider or the client's specific consent before you share confidential client information (Journal of Accountancy), and a personal ChatGPT login gives your firm no agreement at all. The IRS says "federal law requires tax and accounting professionals to create and maintain a Written Information Security Plan" (IR-2026-92), and nobody's personal account is in that plan.

If you don't know which accounts your team is using on client returns today, the quiz below finds that out in ten questions.

Shutting off those accounts is one clause in a wider document, and we've set out what a small CPA firm's AI policy should say, part by part, with the rule each part answers to.

Consent is one check inside the whole process that a firm serious about AI has to go through from start to finish. We've written about how a CPA firm gets going with AI properly, and this is where your question sits in it.

So the first thing is getting the firm ready for AI. We check the business as a whole, each department, the data and tools each department runs on and the people who work there, and fix the tech stack, the quality of the data and data security wherever they aren't ready, including shutting down free and personal accounts. Then we map every task each person does, the hours and hourly rate behind it, where AI already helps with it today and the new services AI makes possible for the firm's clients.

Every solution then gets costed, with the hours of capacity it unlocks, the cost of implementing it and the return it makes in its first year, and put through a compliance, governance and data security check against Section 7216 and every other rule your firm works under, which is where your consent question gets answered, solution by solution, along with where each one runs. So we implement AI in three ways. One is locally on a computer so it doesn't go to any hosting, it doesn't go to any cloud, the next is your own cloud account, and if the work isn't sensitive at all, it runs on Claude or ChatGPT's own server.

To give you an example, which isn't a client result, say a senior preparer billing $175 an hour spends 4 hours a week comparing last year's return with this year's documents to list what each client still owes. Over 48 weeks that's 192 hours, times 0.7 to 0.85 to leave time for review, so 134 to 163 hours of capacity unlocked, worth $23,520 to $28,560 of billable time, and because the task reads the return itself, it would run on a machine in your office. We work that out for every task of every person, then ask what the hours will turn into, because if we can unlock capacity but it doesn't mean anything for the business, then what is the point in unlocking that capacity? We free up your team's time, so the same people can take on more billable work and you don't have to hire to grow.

Those hours are worth most against your job budgets. A return's fee is set from the hours it ought to take, and when it runs over, the extra time gets written off, which pulls down the firm's realization rate (in Britain, its recovery rate). Bring the work in on budget and those write-offs end: revenue rises, gross profit rises, and the same preparers have room for more clients. On tax work that gain comes from the write-offs that stop, never from billing hours nobody worked.

After that, the solutions go on a roadmap of which ones, in what order and how, with the ones worth the most first, under the firm's own AI policy. An operating model shows where and for whom AI is working and the capacity each team has unlocked. Only then do we implement, one solution at a time as the roadmap sets out, each with an enablement pack for the people who use it, and keep every solution up to date as the models and the rules change.

How much work does doing this properly take?

It takes a lot of work, and you should know that before anyone begins. Every person in the firm has to be asked about every task and where AI touches it, including the personal logins nobody mentions until asked, and go through every department's data and tools to find where client returns sit and travel. Every solution has to be read against Section 7216, Rev. Proc. 2013-14, the AICPA code, Circular 230, the FTC Safeguards Rule, your state board and whatever AI laws your state has passed, then costed and ranked, with its provider vetted in writing.

Each consent has to be drafted, signed by every individual client it covers and tracked until it runs out, and after that come the policy, the operating model, each implementation with its enablement pack and named owner, and the upkeep for as long as each solution stays in use. Most firms stall because the partners best placed to lead it are also the busiest billers, which is why so many ask whether they need an AI consultant at all. When we do it, each of your people spends a few minutes describing their work to us, and the rest 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 should your firm do on Monday?

These steps stop the risk while the full process gets going.

  • Ask everyone on your team which AI tools touch client returns, including free and personal logins, and stop return data going into any account the firm doesn't control.
  • Have the provider of the business plan you pay for tell you in writing whether it trains on your data, how long it keeps it and where it processes it.
  • Make sure your written information security plan names the AI tools your people use.
  • Keep individual returns out of any outside AI until your attorney has signed off a consent in the Rev. Proc. 2013-14 format.

If you'd like to see how many of those your firm already has covered, the quiz below checks that across ten questions.

Common questions

Does Microsoft Copilot count the same as ChatGPT under Section 7216?

Yes, because Microsoft is outside your firm, so making return information known to it is a disclosure whatever the tool is called. Find out which Copilot your people use and under which agreement, then ask Microsoft the same questions you'd ask OpenAI.

No, you don't. If someone uses ChatGPT to draft a general newsletter and nothing from a client's return goes in, no return information is disclosed or used. The risk is people drifting from general questions into pasting in a client's numbers, which is why your AI policy needs to say where that line sits.

Does taking the client's name off get around Section 7216?

No, because a return still has an address, an employer and dependents that point back to one household. Anything derived from the return also still counts as return information.

This page is general information, not legal or tax advice, so check your position with your attorney.

Work with us to get AI that's safe to use on your clients' returns

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