Picture this. It’s a Thursday afternoon in the middle of tax season. One of your staff accountants is staring at a messy K-1 and the deadline’s breathing down her neck. So she copies the whole schedule, client name and all, and puts that client data in ChatGPT on her phone. She gets her answer in half a minute and moves on.
Nobody at the firm ever finds out.
Honestly, that’s the scary part. Not that it happened, but that it’s probably happened more than once and you’d have no way of knowing.
If you’re here because you just found out your staff put client data in ChatGPT, don’t panic. It’s fixable. But how you handle the next few days matters a lot.
Why Client Data in ChatGPT Is a Bigger Deal Than It Feels Like
Most people who do this aren’t being careless on purpose. They’re busy, AI is fast, and nobody ever told them it was off-limits. I get it. But in our world, “nobody told me” doesn’t hold up very well.
The one that catches people off guard is IRC Section 7216. Most accountants think of it as the consent rule. It’s actually a criminal statute first. If tax return information gets disclosed without the client’s written consent, you’re looking at civil penalties of $250 per disclosure, and potentially worse. Now think about how many times someone might’ve pasted client info during a busy season. That adds up fast.
Then there’s your AICPA confidentiality duty, which applies even when 7216 doesn’t. And the FTC Safeguards Rule, which treats tax and accounting firms as financial institutions regardless of size. Yes, even if it’s just you and one assistant. That’s exactly why WISP compliance isn’t optional for CPA firms anymore.
The IRS has gotten involved too. Back in June 2026, the Office of Professional Responsibility put out Alert 2026-19, its first real guidance on AI under Circular 230. The short version? They expect firms to have some control over how AI touches client data.
And here’s something almost nobody talks about. A ChatGPT conversation isn’t privileged. If there’s ever a lawsuit, that chat history full of client numbers could end up in discovery.
What to Do When Client Data Ends Up in ChatGPT
Treat this the way you’d treat a smoke alarm. You don’t start redesigning the kitchen. You find out what’s burning.
Start by getting everyone to stop. Not just the person who did it, the whole team. Keep the tone calm. If people feel like they’re about to get fired, they’ll stop telling you things, and right now you need the full story.
Next, figure out exactly what went in. Was it just a generic tax question with some numbers, or did it include names, Social Security numbers, full returns? Was it a free account or a paid one? Ask for screenshots of the conversations before anything gets deleted.
Once you’ve got a record, have the employee delete those chats and turn off any setting that lets the tool train on their data.
Then pick up the phone. Call your attorney about 7216 and your state’s breach notification rules, and call your professional liability carrier early. Waiting too long to notify your insurer can come back to bite you.
You may also have notification duties. It depends on what was exposed and how many people were affected. Under the Safeguards Rule, if unencrypted info on 500 or more people is involved, the FTC expects to hear about it within 30 days of when you found out.
And write everything down as you go. What happened, when you learned about it, what you did next. If anyone ever asks questions later, that record is going to be worth its weight in gold.
Fixing the Actual Problem
Here’s the uncomfortable truth. If one person did this, chances are others have too. They just haven’t been caught.
Banning AI outright sounds tempting, but it rarely works. People just keep using it quietly on their phones where you can’t see it. What actually works is giving your team a safe option and some clear, simple rules.
Your AI policy doesn’t need to be some 40-page binder. Two pages that people actually read will do far more good. It should tell people which tools they’re allowed to use (and that personal accounts are a no-go for client work), what kind of information can go in and what absolutely can’t, and that a real person reviews anything AI helped create before it goes to a client. That last part matters more than you’d think, because AI has a habit of confidently making up tax citations that don’t exist.
If client data genuinely needs to go through an AI tool, get proper 7216 consent and update your engagement letters to reflect it.
And make it safe for people to come forward when they slip up. A team that’s scared to admit mistakes is a team that hides them.
It Looks a Little Different Depending on Your Size
If you’re a solo practitioner or a small shop, keep it simple. Pick one business-grade AI tool with real data protections, write a one-page policy, and fold it into your WISP. If you don’t have one yet, our free WISP template for CPA firms is a good place to start.
For mid-size firms with a few locations, the challenge is consistency. What one office thinks is fine, another might not. You’ll want MFA on every account, one central list of approved tools, and some controls on browser extensions, because those can quietly read everything on a screen, including your tax software. This is where layered security solutions earn their keep.
Regional and multi-partner firms usually need more than a policy. At that volume, you’ll want data loss prevention tools, proper vendor vetting, and someone whose job it is to keep an eye on all this. A dedicated vCIO through remote managed IT services can be a lifesaver here, especially since these AI tools change every few months.
And if you’re already recovering from a breach, you know people are watching closely, from clients to insurers and maybe regulators. Use that. A written policy, training records and a tested response plan show that you didn’t just clean up the mess, you fixed what caused it. It’s also worth reading up on newer threats like ClickFix attacks targeting CPA firms, because AI misuse is rarely the only gap.
The Bottom Line
Your team is going to use AI. That ship sailed a while ago. What you get to decide is whether they use it safely, with rules you’ve set, or quietly, in ways you’ll only find out about after something goes wrong.
If you’ve already found client data in ChatGPT, start on those first steps today. Then put a policy in place so you’re not reading an article like this again next tax season.
Not sure where your firm stands right now? Our team works on IT for CPA firms every day. Get a free IT audit and we’ll show you where client data might be slipping through and what to fix first.
FAQs
Is it actually illegal for accountants to use ChatGPT?
Not at all. Drafting emails, brainstorming, explaining a tax concept in plain English, that’s all fine. The trouble starts when identifiable client information goes into a tool without the right protections or consent.
If we delete the chats, are we in the clear?
Unfortunately, no. Deleting them helps limit further exposure, but the disclosure already happened. You’ll still want to look at your legal and notification obligations.
Are paid tools like ChatGPT Enterprise or Copilot safe for client data?
They’re a lot safer than free personal accounts, since business versions usually come with contracts that keep your data out of model training. But they still need to be set up properly, and you’ll still need a policy and 7216 consent where it applies.
We’re a tiny firm. Do we really need a written AI policy?
Yes. The FTC Safeguards Rule doesn’t care how small you are, and the IRS now expects firms to have some handle on AI use. The good news is one or two pages is plenty to start.
Should we let go of the employee who did it?
In most cases, no. Especially if there wasn’t a policy in place to begin with. Make it a learning moment, document it, and fix the system so it doesn’t happen again.