In June 2026 the IRS Office of Professional Responsibility published guidance on responsible AI use in federal tax practice. It is short, and its most useful quality is that it declines to invent anything.
Why "No New Rules" Is the Bad News
The duties it describes are the ones in Treasury Circular 230, the rules governing practice before the IRS. Due diligence. Competence. Confidentiality. Fees that are not unconscionable. All of them were in force before anyone in a tax practice had heard of a large language model.
What the guidance does is state, in writing, that those duties apply to work an AI helped produce.
Practitioners tend to hear regulator silence as permission and new guidance as a deadline. Both readings fail here.
If the IRS had announced a new AI rule effective next January, a firm would have a grace period and a project. Instead the position is that the obligations were always there. There is no start date to plan around, because the standard applied to the return that went out in March.
That is the shape of the exposure. A practitioner who relied on an AI-drafted memo in the spring and did not verify its authorities was not waiting for guidance. They were already short of a standard that had not changed.
The Four Duties, Stated Plainly
The guidance is specific about what compliance looks like, and none of it is exotic.
Review the output. Practitioners must remain vigilant in reviewing what AI produces, validating its factual assertions and its citations. Fabricated authorities are the failure mode the IRS names, and courts have now sanctioned lawyers in several matters for filing them. A citation that looks right is not a citation that exists.
Understand the tool. Competence covers the technology as well as the law, which means understanding how a system produces its content, where it is prone to error or bias, and whether its output is suitable for an IRS matter at all.
Protect the data. Client information goes into secure, firm-approved systems. That sentence quietly rules out a great deal of what actually happens in practices where nobody set a policy.
Bill honestly. Where AI reduces the time a task takes, the bill should reflect it. Charging for hours that the software removed raises an unconscionable fee question, and that is a Circular 230 matter rather than a commercial one.
What Is Actually at Stake
This is the part that gets skimmed, so it is worth naming.
Circular 230 proceedings can end in censure, suspension or disbarment from practice before the IRS, monetary sanctions, and orders to refund fees. Those are not theoretical instruments. They exist, OPR uses them, and the guidance exists precisely to remove the argument that nobody said AI was covered.
A practitioner facing that conversation will be asked what was reviewed and by whom. The answer has to be a record.
The Workable Version
Three things, none of which requires a policy committee.
Every AI-assisted item that reaches a client or the IRS gets a named reviewer and a note that the authorities were checked. Client data goes only into approved systems, and somebody finds out what people are actually using. Where AI compressed the work, the fee conversation happens before the invoice rather than after a complaint.
MetaWurks is built for the middle one and helps with the first. Documents ingested into the platform are not used to train models and are not exposed to other users, which is what "secure and approved" has to mean in practice rather than on a policy page. Role based access controls decide who can open which client's records, and audit logs record who opened what and when, so the question of who looked at this, and when, has an answer.
It does not check a citation. A person still does that, and the IRS has now written down that a person must.
The guidance changed nothing about your obligations. It removed the excuse.
Join the Conversation
For the last AI-assisted document your firm sent to a client or the IRS, can you name who verified its citations?