On June 24, 2026, the IRS Office of Professional Responsibility told tax practitioners to bill clients for what AI actually saves them — efficiency from a language model credited back through lower fees, not charged as if a person had still spent the hours. In September, the AICPA told the IRS that instruction doesn't add up.
The guidance treats an hour a chatbot saves as an hour a client is simply owed. The AICPA's objection is that the hour was never free to the firm in the first place — someone paid for the license, the implementation, and the training the tool still needs before anyone can trust its output, and none of that shows up in OPR's arithmetic.
What the alert actually asks for
OPR Alert 2026-19, "Introductory Guidelines for Responsible AI Use in Federal Tax Practice," reaches billing through Circular 230 § 10.27(a) — the rule against unconscionable fees. Per its own language, cited in trade coverage of the alert, a practitioner using generative AI should pass along efficiencies gained from the technology through billing practices that reflect reduced research and drafting time, and fairly credit clients for the resulting cost reductions. Pernee covered this same alert's disclosure gap in September; the billing clause is the half that just drew a formal response.
What the AICPA pushed back on
Eva Simpson, the AICPA's vice president for member value in tax and advisory services, called the instruction an overly simplistic view that ignores the full economics of AI adoption — software licensing, implementation, the training a firm has to run before anyone can use the tool responsibly, and the liability that doesn't disappear just because a machine wrote the first draft.
Value, however it is performed by AI or by the profession, has a cost and it has a benefit.
Jan Lewis, AICPA Chair
A second objection sits closer to how firms actually charge. OPR's framing assumes billing by the hour — an hour the tool now fills faster, credited straight back. AICPA President and CEO Mark Koziel pointed out that a growing share of firms don't bill that way at all: value pricing and bundled engagements price the outcome, not the clock, and a rule written for hourly billing has nothing to say to either. He also noted the alert could be "overstepping a little bit," since it carries no authoritative status of its own. The AICPA says it's now working with the IRS on clarifying language and FAQs — no date attached.
The review discipline the billing question sits on top of
Five checks review already runs — internal consistency, agreement to source, and more — regardless of who, or what, drafted the return first.
Where this actually lands, for the desk doing the work
The argument is about billing, but underneath it is a measurement problem review already has to answer. "Credit the client for what AI actually saved" presumes a firm can say, with a straight face, what AI actually saved — on this return, on this review pass, net of the time a reviewer still spent checking the tool's output. Cycle time, review passes per return, and the recurring-comment rate a team keeps writing back are the same metrics that answer that question, and just as easy to game as OPR's billing rule assumes they aren't.
Neither side of the billing argument changes what still has to happen before AI touches a return: due diligence under § 10.22, independent verification of the output, a record of who checked it. The billing question is downstream of that discipline, not a substitute for it.
Whoever the IRS sides with when the FAQs land, the number a firm needs first is its own — not OPR's assumption of what AI saved, and not the AICPA's objection to that assumption, but a measured account of what a given AI-assisted return actually took to get right.
Measure review time you can actually bill on
Pernee attaches verification time to the return line it covers, so "what AI saved" is a number, not a guess.
