IRS OPR 2026-19: What It Means for AI in Tax Practice - Part 1

On June 24, 2026, the IRS Office of Professional Responsibility issued Issue Number 2026-19, titled Introductory Guidelines for Responsible AI Use in Federal Tax Practice. Read the headline and you might brace for a new AI rulebook. There isn't one. OPR did not write a separate AI law or a new penalty regime. It took the standards that already govern our work and spelled out how they apply the moment you drop a prompt into a generative AI tool.

Due diligence, competence, written advice, client confidentiality, reasonable fees, firm-level supervision: those duties predate generative AI by decades. OPR's point is that none of them change when the first draft comes from a model.

The central principle is short. The practitioner owns the final work. Not the model, not the vendor. That means the facts, the authorities, the math, the handling of client data, the written advice, and the fee all trace back to you.

For CPA firms, enrolled agents, tax attorneys, and in-house teams, that makes OPR Issue 2026-19 less a warning than a checklist. It gives you a way to look at how AI is already running through your practice and where the controls are thin.

Why Generative AI Creates Different Risks

We have leaned on software for a long time. Preparation packages, research databases, depreciation engines, document management, all standard equipment. Generative AI is a different animal.

It writes. Ask it a question and it produces original content: a summary of an IRS notice, a read on a K-1, a research memo, a penalty-abatement request, a client email, a federal-to-state comparison. Ordinary tasks in any tax practice. The catch is presentation. The output arrives complete, polished, and confident even when the analysis under it is thin or flat wrong.

OPR names the recurring risks:

• Fabricated authorities, quotations, or facts
• Bias in the output
• Little transparency into how the answer was reached
• Weak protection of confidential information
• Retention or reuse of client data
• Over-reliance on AI-generated conclusions
• Errors that are hard to catch because the response sounds credible

The first and last items are the ones that bite in tax research. A model will cite a real Code section and then misapply it. It will name an actual case and mangle the holding. It will state a correct general rule and quietly drop the limitation, the exception, the effective date, the election, the state conformity wrinkle. We have watched this happen, and the consequences were not academic.

A clean answer breeds false comfort. Your job is to judge the substance, not the fluency. Three takeaways from the alert are worth your time.

Takeaway 1: Treat AI Output as a Draft

OPR keeps returning to one word: draft. Whatever the model gives you is a starting point, not a deliverable. Before that work reaches a client, lands on a return, goes into a filing, or gets submitted to the IRS, you review it yourself. That review means you:

• Confirm each cited authority actually exists
• Read the authority, not just the citation
• Check quotations and case references against the source
• Confirm the authority is current and supports the conclusion drawn
• Rework the calculations and formulas
• Flag the missing facts, assumptions, and exceptions
• Confirm client information was handled properly

This holds whether the output is a fifteen-page memo or one sentence in a client email. It holds even when the tool hands you citations. Citations are a real time-saver, since they point you straight to the authority instead of making you hunt for it, but they do not close the loop. You still confirm the source is accurate, current, and on point for the facts.

Takeaway 2: The Practitioner Stays Responsible

Here is the message I would tape to the monitor. Under OPR Issue 2026-19, responsibility does not transfer. Not to the platform, not to the vendor, not to the staff member who typed the prompt. If an AI-drafted protest letter cites a case that does not exist, the practitioner who signed and sent it answers for it. Upload a client's documents into an unapproved system and the firm owns the confidentiality fallout. Cut an engagement's hours in half with AI and the fee still has to be reasonable. More on billing in a later issue.

The standard that governs hand-prepared work governs AI-assisted work. Same bar.

How the Alert Maps to Circular 230 and the Code

Here is how the alert lines up against the rules you already answer to, and what each one asks of an AI-assisted workflow.

None of these provisions are new. The alert just draws the line from old rules to new workflows.

Takeaway 3: Firm Responsibility Goes Beyond Individual Review

The alert does not stop at the person holding the keyboard. Circular 230 §10.36 puts a duty on whoever runs the firm's federal tax practice to build compliance procedures and then confirm people follow them. For AI, that is a concrete list: decide which tools are approved, give staff the right ones so they stop reaching for random consumer apps, limit public platforms, control what client data can be uploaded, set review and sign-off rules, vet vendors, and keep a plan for when something breaks or leaks. A policy in a shared drive does nothing on its own. Staff have to understand it, and managers have to watch how AI is actually being used, not how the memo says it should be.

What Tax Professionals Should Do Now

AI is already in your office whether or not you have blessed it, so start there. Do not wait for a finished governance framework from the IRS before you handle the obvious risks:

• Find out which tools your staff already use, and for what
• Keep confidential information out of public or unapproved platforms
• Check every AI-generated citation, quotation, calculation, and conclusion before it reaches a client or the IRS
• Name a qualified reviewer who owns the final output on each AI-assisted work product

Then write the policy down. A basic one covers approved tools, permitted uses, restricted data, review requirements, training, vendor checks, and what to do when an error or a data incident happens. The right controls depend on your practice, since a solo EA and a two-hundred-person firm will not land in the same place, but the last line reads the same for both. The final work is yours.

The Bottom Line

OPR is not anti-AI. The alert accepts that these tools can make a practice better. It simply refuses to let efficiency buy down responsibility. Accuracy, advice, client data, fees, firm-wide procedures, all still yours. The firms that get the most out of AI will be the ones that pair it with real controls and real human review.

This is the first article in Feather's series on OPR Issue 2026-19. Next up: what AI hallucinations actually cost in practice, and the warning signs that should stop you before you rely on a polished answer.

Have Additional Questions?

If you are building an AI review process, tightening a firm AI policy, or checking a position against the Code, Treasury Regulations, IRS guidance, and state authorities, Feather can help with citation-backed, practitioner-grade answers you can verify. That is the whole point: research you can stand behind.