At a lot of firms this coming busy season, someone on staff will paste part of a client’s return into a chat tool to get a quick summary. They won’t mean any harm, since they’re usually just trying to clear a backlog. The question for firm owners is whether the firm has decided, in writing, where that data is allowed to go.
The IRS has now said a fair amount about this. On June 24, the IRS Office of Professional Responsibility issued OPR Alert 2026-19, its introductory guidelines for responsible AI use in federal tax practice. Most of it applies Circular 230 duties practitioners already know, but a few lines deserve a place in every firm’s AI policy.
What the IRS guidance asks for
The alert says practitioners must handle client data “using only secure, enterprise-approved AI,” and it warns against uploading sensitive information to unsecured or public systems. It ties that to the civil and criminal penalties for unauthorized disclosure under IRC Sections 6713 and 7216. It also says firm leaders with oversight duties under Circular 230 Section 10.36 need procedures covering staff training, secure data handling, AI accuracy monitoring and vetting of outside AI tools, with every step documented.
Two other points are easy to miss. The guidance names cross-client leakage as a privacy risk, where data from one client’s file is reused to answer a question about another. And under Section 10.27 on fees, it says billing for manual time that wasn’t spent, or double billing for AI-assisted work, may be a violation depending on the facts.
The consent question has no clean answer yet
Section 7216 requires a taxpayer’s written consent before a preparer uses or discloses return information outside the regulatory exceptions. For Form 1040-series returns, Rev. Proc. 2013-14 sets out a separate, signed consent format. The IRS hasn’t said whether sending return information to an AI tool counts as a disclosure to a third party.
Until it does, the cautious approach is a standalone, signed AI consent, worded with your counsel. Plan for the clients who say no. The regulations say a consent must be knowing and voluntary, and conditioning services on it generally makes it involuntary. That means the firm needs a way to prepare a return without AI for any client who declines, and staff need to know which clients those are before they open the file.
Read the terms for the exact plan you use
“We use an AI tool” tells you very little. The questions that matter are specific to the product and the plan. Does the provider keep prompts and files, for how long, and does it train models on customer data? Does a zero-retention agreement cover the chat app your staff actually open, or only the developer API? Consumer and business tiers of the same product can carry different terms.
Put the answers in your vendor file and the requirements in your contracts. If a vendor won’t say in writing that your data isn’t used for training, treat that as the answer.
Keep each client’s data in its own lane
The cross-client risk the IRS described is a design problem as much as a policy one. A tool that searches every client file to answer a question can pull a figure from the wrong return. Confirm that any AI system you use separates client data, so work on one client can’t draw on another client’s documents. Ask vendors how that separation is enforced, and what happens when one employee can open many client files.
Keep a record of what touched what
The alert asks firms to document AI use and verification. In practice, that means being able to answer a simple question for any return: which tools processed this client’s information, and who reviewed the output before it went out?
That record helps with the fee guidance too. If AI cut the time on a return, the log shows it, and the firm can price its work fairly and explain the change to clients.
When my co-founder and I built agent workflows for a high-volume tax practice, human time came down to less than 5 minutes per return. That only worked because the preparer could see every action the agents took, and anything that needed the preparer’s approval waited for it. The same running record that let the preparer check the agents is the kind of documentation the alert now asks firms to keep.
A short starting list
For most firms, a workable first AI policy fits on one page. It should list the approved tools and the plans you’re on, and say that client return information goes only into those tools. Adopt a standalone AI consent with your counsel and set a no-AI path for clients who decline. Every AI output gets human review before it reaches a client or the IRS, and the firm logs which tools touched which client’s data. Because the guidance will keep changing, review the policy every year.
Name one person to own the policy and the vendor file, and put the first review date on the calendar.
ABOUT THE AUTHOR:
Steven de la Fe, CPA, is co-founder of Precision AI OS, an AI implementation firm in Miami that builds AI agents for service firms, including accounting and tax firms, law firms, engineering firms, insurance agencies and consulting firms. The agents take repetitive work off a firm’s team, so the same people can serve more clients without adding headcount. Before Precision, he and co-founder Phillip Ingelmo spent ten years building Taxfyle from the ground up, taking it from $0 to $25M in annual recurring revenue.
Sign in to get access to this free resource, and all of our whitepapers and reports.
Download this content today!
Register Now Already registered? Click here to Log In
Tags: Advisory, Artificial Intelligence, IRS, tax data, Taxes, Technology
