Dear Deb: The IRS Has AI Guidelines. What About Bookkeepers?

Editor’s Note: This article is part of a series. View all the articles in this series here: Dear Deb Series 

Dear Deb,

I recently read an article in The Woodard Report about the IRS issuing guidance for tax professionals using artificial intelligence. I own a bookkeeping firm. We don't prepare tax returns. We work with tax preparers who rely on our financial statements.

We're beginning to use AI for reconciliations, transaction categorization and other bookkeeping tasks. I want to take advantage of the technology while protecting my clients and my firm.

Who is providing AI guidance for bookkeepers? Should I have a written policy? What should it include, and how do I keep it current when the technology changes every week?

Signed, Trying to Keep Up Without Messing Up

Dear Trying to Keep Up,

That's a great question, and I think every bookkeeping firm needs to ask it. I read that same Woodard Report article. It got me thinking about what firms that don't prepare taxes can do.

On June 24, 2026, the IRS issued its Introductory Guidelines for Responsible AI Use in Federal Tax Practice. The guidelines explain how tax practitioners must uphold their existing responsibilities when using AI, including reviewing their work, protecting client information and supervising employees.

Unlike tax preparers, we don't have one organization telling us how to use AI responsibly. Circular 230 doesn't apply to us simply because we prepare books that someone else uses for a tax return. That doesn't mean we can do whatever we want. Depending on the services we provide, there may be state rules, professional standards, or privacy laws we need to follow. The FTC Safeguards Rule is one example. It applies to certain financial businesses and whether we're included depends on what we actually do.

We need to develop our own AI guidelines. We need to decide which tools employees can use, how we protect our clients' information and what an acceptable review of AI-assisted work looks like. Employees need instructions they can follow and adapt as technology changes.

Learn what AI can do before deciding how to use it

A restaurant client of mine changed point-of-sale systems. Overnight, the reports looked very different. The old reports summarized sales into categories such as wine, liquor, beer and soft drinks. The new reports listed every item purchased: 4 Rum and Cokes, 10 Coors Lights, 2 Michelobs and so on. My head hurt looking at it. I thought, "Oh my gosh, now what am I supposed to do with all this?"

I experimented with Claude to see whether it could reorganize the information into our accounting categories. It could. Now I can generate the import file in seconds. I still review the results.

Before that experiment, I didn't know AI could reorganize POS data this way. Once I discovered what it could do, I could rethink our process: what AI could handle and what still needed me.

For your first experiments, use fictional information. You can learn plenty without introducing client confidentiality questions on day one.

A checked box doesn't mean the work is correct

Ever had QuickBooks match the wrong transaction in a bank feed because the amounts were identical? I bet you have. Just because an account reconciles doesn't mean everything is correct. The same goes for AI-assisted work. We still need to check the results.

I learned that lesson years before AI. My firm processed restaurant payroll with tips paid out in cash daily. Our checklist required verification that reported tips "IN" matched reported tips paid "OUT." I reviewed the completed checklist and trusted that the employee had performed the verification.

Around Christmas, the restaurant owner called about incorrect paychecks. Tips hadn't been deducted properly for three months. It was a $7,000 error and we reimbursed the client.

I assumed that checking the box meant the employee had verified the numbers. I was wrong. She checked it even though the amounts didn't match. I trusted the checklist without looking at the underlying figures. Today, as AI takes over more of our bookkeeping work, I ask myself the same question: Am I reviewing the results, or am I trusting that someone, or something, already did?

What about client confidentiality?

Confidentiality is an area where I'm still learning. I suspect plenty of other firm owners are too.

I've talked with colleagues who won't use AI because they're worried about protecting client information. I understand their hesitation. There's so much conflicting information about what's safe that it's difficult to know what to believe. I also think they're missing opportunities to make their work easier. I'd rather learn how to use AI responsibly than avoid it altogether.

When I started processing restaurant POS reports with Claude, I used my Pro subscription with model training turned off. The reports contained the restaurant's name, date and sales information. I was focused on solving the bookkeeping problem and hadn't examined all the implications of using a consumer AI subscription for client work.

I've since upgraded my firm to Claude Team. I wanted commercial data-processing terms and administrative controls. I also needed to understand how the provider handles client information.

And this isn't just about ChatGPT or Claude. We're already trusting QuickBooks, Digits, Double and other vendors with our clients' information. As they add AI features, I think we need to start asking more questions about how they use and protect that data.

I'm now limiting uploads to the information the task requires. Claude doesn't need a restaurant's name to summarize rum and Coke sales.

Remember the tax preparer

Tax preparers depend on the information we provide. When AI assists with transaction classification, we still need to recognize questionable entries.

Suppose an AI-powered feature suggests Uniform Expense for purchases from a clothing retailer because similar transactions were previously classified there. The client insists they're business expenses. I'd investigate, correct any personal expenses classified as business expenses and bring unresolved tax questions to the preparer's attention.

I struggle with whether we need to tell clients that AI helps us prepare their books. I don't tell them when I use Excel for calculations or when QuickBooks automatically categorizes transactions. Why should AI be different? Some clients are uncomfortable with AI, and I wonder if telling them about every tool we use creates unnecessary worry.

We've always had a responsibility to protect client information, and AI doesn't change that. The tax preparer needs accurate books and an explanation of any unresolved questions.

Keep your AI standards working

I recommend reviewing your firm's AI guidelines quarterly, or sooner if you introduce a new tool, change how you work, or discover a recurring problem. Don't wait three months to correct something you discovered yesterday.

Your clients didn't hire you because you know how to click ’reconcile’. They hired you because you understand what a reconciled account should mean.

Start with one AI-assisted process. Document what the technology does, how someone reviews its output and who approves the result. Then ask that person to show you how they know the work is correct. That conversation tells you more than another signature on a policy.

You may discover your process works beautifully. You may also discover you've been trusting a checked box.