On June 24, 2026, the IRS Office of Professional Responsibility issued OPR Alert 2026-19, "Introductory Guidelines for Responsible AI Use in Federal Tax Practice" — the agency's first guidance mapping a practitioner's existing Circular 230 duties onto AI-assisted work. It doesn't create new law. It tells a preparer which of the old rules apply when a language model, not a staff associate, produces the first draft.
What it tells them to do is specific. What it never tells them to do is mention it to the client.
What the alert actually requires
OPR Alert 2026-19 maps AI use onto six areas of Circular 230 — due diligence, competence, firm oversight, written advice, fees, and data confidentiality (§§ 10.22, 10.35–10.37, 10.27(a); IRC §§ 6713, 7216(a)). None of it is aspirational: a practitioner verifies the output, understands the tool that produced it, and bills for what the tool actually saved — not for hours a person didn't spend.
AI produces a draft
A tool generates a position, a calculation, or a first-pass narrative — nobody has reviewed it yet.
A practitioner verifies it
Every fact, citation, and calculation checked independently. Circular 230 § 10.22.
The tool itself gets scrutinized
Understanding how it generates output, and where it tends to fail. § 10.35 competence.
Billing reflects the efficiency
Time the tool actually saved is credited to the client, not charged as if a person did the work. § 10.27(a).
What's not on this list
Telling the client AI touched the return. OPR Alert 2026-19 never asks for it.
OPR's own list of AI failure modes, cited in trade coverage of the alert, names three: fabricated output ("hallucinations"), embedded bias, and a lack of transparency in how the tool reached its answer — reasoning a reviewer can normally trace back through a human associate's workpapers and generally cannot through a model's.
The rule that isn't there
Search the alert for a disclosure requirement and there isn't one — no instruction to notify a client, get consent, or note AI's involvement anywhere the client would see.
AI should assist, not replace, professional judgment. Final decisions must always rest with qualified professionals.
OPR Alert 2026-19
Accountability doesn't move to the tool no matter how much of the draft it wrote — but a firm staying accountable and a client knowing are two different guarantees, and the alert only requires the first. The AICPA's own commentary, published a week later, doesn't add the second either.
The five checks a review already runs
Pernee's guide covers the review model due diligence under § 10.22 assumes — internal consistency, agreement to source, prior-year continuity, position authority, and planning quality.
Where the gap actually gets closed
For a firm doing tax return review, the gap doesn't stay theoretical. Due diligence under § 10.22 is only as real as the record behind it — a firm that can't show what an AI tool touched on a given return, and who verified it, hasn't satisfied the duty just because someone glanced at the output once. That's the same discipline review already runs on for a different reason: a second, independent check, attributed to a reviewer, attached to the return. Recording which lines AI touched isn't a new process. It's the existing one, with a column added.
No regulator is asking this season whether a client was told. An examiner asking who verified a specific line is a nearer question, and OPR Alert 2026-19 already assumes a firm can answer it.
Give a reviewer a record AI use didn't erase
See how Pernee keeps every AI-assisted position attached to a named reviewer and the return line it explains.
