How to Use AI for Tax Research and Document the Sources

Can a CPA firm use AI for tax research to support a return position?

Yes, as a starting point. Using AI for tax research can surface the Code sections, regulations, rulings and cases faster than a manual search, but the AI answer cannot be the support itself. Treasury regulations say conclusions in treatises, legal periodicals and opinions of tax professionals are not authority, and an AI answer is not authority either.

That distinction matters at signing time. Under 26 U.S.C. 6694(a), a preparer who knew or reasonably should have known that an undisclosed position lacked substantial authority faces a penalty of the greater of $1,000 or 50 percent of the income derived from that return. The penalty turns on the authorities behind the position. A well-written AI summary is not one of them.

What counts as authority for a tax return position?

The list is in 26 CFR 1.6662-4(d)(3)(iii): the Code and other statutes, final and temporary regulations, revenue rulings and procedures, court cases, congressional intent in committee reports, private letter rulings and technical advice memoranda issued after October 31, 1976, IRS notices and announcements, and similar sources. The same paragraph states that conclusions in treatises, legal periodicals, legal opinions or opinions rendered by tax professionals are not authority.

The regulation also says the weight of an authority depends on its relevance and persuasiveness, and that an authority which merely states a conclusion is less persuasive than one that reasons its way there. An AI tool that states a conclusion with a citation attached is, at best, a pointer to an authority someone still has to read.

Substantial authority itself is an objective test of the law applied to the facts. Under 26 CFR 1.6662-4(d)(2) it is less stringent than more likely than not (a greater than 50 percent likelihood of being upheld) and stricter than reasonable basis. Whether a position clears that bar depends on the weight of the real authorities, not on how confident the AI sounded.

What do Circular 230 and the AICPA standards expect when AI is involved?

They expect the same diligence as any other research method, plus judgment about the tool. 31 CFR 10.22 requires due diligence in preparing returns and in determining the correctness of representations to clients, and presumes diligence when a practitioner relies on others' work only if the practitioner used reasonable care in supervising and evaluating that work.

For written advice, 31 CFR 10.37 requires the practitioner to base advice on reasonable factual and legal assumptions, relate applicable law and authorities to the facts, and not rely on another person's advice when the practitioner should know it should not be relied on. 31 CFR 10.34 bars signing a return with a position that lacks a reasonable basis or is unreasonable under section 6694(a)(2).

For AICPA members, the revised Statements on Standards for Tax Services, effective January 1, 2024, added a standard on reliance on tools and a standard on data protection. Both apply directly to an AI research tool: when it is reasonable to rely on the tool, and what client information can go into it.

What should a tax research memo include when AI helped find the answer?

A tax research memo should show the authorities, read and evaluated by a person, with the AI's role recorded. A reviewer or an IRS agent should be able to follow the reasoning without opening the AI tool.

A practical file for each position includes:

• The question as framed, with the relevant facts and tax year.
• The tool used, the account type (firm business plan or other) and the date of the session.
• Every authority the tool surfaced, with a note that someone pulled and read the primary source.
• Authorities the preparer found independently, including any that cut against the position.
• The conclusion, the standard it meets (substantial authority, reasonable basis with disclosure, or more likely than not for tax shelters and reportable transactions) and who reached it.
• Any citation the tool produced that did not exist or did not say what the tool claimed.

That last item is the one firms skip. Recording it trains the team to verify and gives the reviewer a reason to look harder at that engagement.

How do you keep client data out of an AI tax research prompt?

Write the prompt as a hypothetical. Research rarely needs the client's name, SSN, EIN or return, so leave them out and describe the facts generically. Uploading a return to a tool outside the firm's approved list can be a disclosure under Section 7216, which the regulations define as making tax return information known to any person in any manner whatever.

The firm-level control is technical as well as written. Approve one business-grade research tool whose terms exclude training on firm content, block personal accounts on firm devices, and set data loss prevention rules for SSNs, EINs and bank numbers so a careless paste is stopped before it leaves the browser. Our guide to keeping client tax documents out of public AI covers those rules in more detail.

Who should review AI-assisted research before the return is signed?

The reviewer who signs or approves the return. Section 10.22 ties the presumption of diligence to reasonable care in supervising and evaluating the work relied on, so review is where that care is shown.

A simple rule works for most firms: any position where AI was used in the research gets a reviewer sign-off that names the primary authorities and confirms they were read. For positions on the edge of substantial authority, the reviewer also decides whether disclosure on Form 8275 is warranted. Firms documenting AI under SQMS No. 1 can fold this sign-off into the same records.

What to do this quarter

Before busy season starts, set the research standard once so staff are not improvising in March:

• Name the approved AI research tool or tools and block the rest on firm devices.
• Add a one-page AI research memo template with the fields listed above.
• Turn on DLP rules for SSNs, EINs and account numbers in browsers and Microsoft 365.
• Run a 30-minute session with preparers on the difference between an AI answer and an authority, using a fabricated citation the team catches together.
• Add an AI research check to the review checklist for every return with a judgment position.
• Update the WISP to name the approved tools and the data that may not go into them.

How s90 helps

s90 manages IT and compliance for CPA firms and tax practices in Los Angeles and Orange County. We set up and enforce the controls behind an AI research policy, including approved tools, blocked personal accounts and DLP for tax identifiers, and we keep the evidence ready for a peer reviewer or an IRS inquiry. If you would like a second set of eyes on your AI controls before filing season, we are glad to talk.