Practical AI prompting for tax issues and regs

Key takeaways

• AI is not a tax authority. Treat it as a research assistant that organizes facts, surfaces candidate code sections, and drafts a starting analysis.
• Work in two prompting stages: A discovery prompt that turns a client’s quick question into the right tax issues and follow-up questions, then a full research prompt that locks in tax year, jurisdiction, and material facts.
• Never accept a citation at face value. Every code section, regulation, and case citation an AI model gives you needs the same verification you would apply to any other research source.
• Know the authority hierarchy. The Internal Revenue Code, US Treasury regulations, and judicial decisions are the primary authorities. IRS publications and form instructions are secondary. 
• AI-assisted research regulations: The IRS Office of Professional Responsibility has already addressed AI-assisted research directly. Circular 230 due diligence, professional competence, and data security obligations all apply, in full, to work you do with AI.

AI is an assistant, not an authority

I’ve been a CPA for more than 15 years in public accounting and writing code for tax software, and now running my own advisory-focused firm in Dallas. I’ve watched AI move from a novelty to a tool I lean on for genuinely sophisticated work. I’ll even admit I once used ChatGPT to figure out what to cook based on a photo of what was inside my fridge.

These days, Claude is my preference for the heavier lifting, although from time to time, I still reach for ChatGPT.

However you use it, the rule I start every training session with is this: AI is not tax authority. It never will be. When a client comes to you with a quick question, the honest answer is almost always “it depends,” because it depends on the year, the facts, the jurisdiction, and the relationships involved.

AI can help us organize those facts, suggest provisions worth investigating, and draft a preliminary analysis. However, it can’t establish the authoritative basis for a conclusion we’re willing to put our name on. A supportable tax position traces back to a primary authority, not to a model’s output.

The consequences of forgetting this point are real, and they’ve already played out publicly. Deloitte was paid by the Australian government to produce a report reviewing government entities, regulations, and processes. Because the engagement involved a government agency, the report became public record. A university researcher who dug into it found more than 20 instances where citations to laws and judicial decisions were wrong.

Deloitte ended up refunding part of its fee and reissuing the report with a disclaimer that AI had assisted the findings, alongside confirmation that a second set of human eyes had now verified the sources. That’s the scenario every one of us wants to avoid, and the fix is the same one we’re going to walk through here: Take the extra step to verify.

Start with a discovery prompt

Let’s say a client tells you they rented their cabin to their sister, and wants to know whether that counts as personal use or rental use, and if they can deduct expenses against the rental income. That’s not a quick question; it’s a research question wearing a quick question’s clothes.

If you’re not already fluent in the personal use versus rental use rules, don’t ask AI to “do the research.” Start with what I call a discovery prompt. Give the model the client’s exact question, then ask what areas of tax law it could touch and what follow-up questions you should ask the client before you can even frame the final issue. The output should be a list of candidate primary-authority citations and a list of facts you still need to gather.

For the cabin example, a discovery prompt surfaces some key questions. Is the rent at fair value? Does the owner ever use the cabin personally? Does the sister use it as her own personal home rather than a temporary stay? It also flags the framework you need to fill in before you can build a real research prompt:

• The identity and jurisdiction: federal and any applicable state law.
• Activities and amounts: Is this a meaningful dollar threshold, and what is the fair value?
• Use and related parties: Does the family relationship matter under the code?
• Exceptions and limitations: Which tax year are you working in? Something such as 2025 tax legislation can change the answer.

Build the research prompt

Once you have those factors, build the research prompt. Specify the tax year and jurisdiction, and state the material facts precisely—a sister renting a dwelling unit at roughly $1,000 a month, which the client confirms is fair value—and ask directly for the tax reporting consequences. Then add directives that keep the model honest. Require it to label its assumptions and uncertainties, and instruct it not to treat its own output as verified.

Compare two prompts side by side. For example, “Explain the Augusta rule when I rent to family” is the kind of shortcut prompt that circulates on social media. It’s also the weaker choice every time. A stronger prompt names the tax year, asks the model to identify the federal tax issues, potentially applicable code provisions, definitions, exceptions, and missing facts for a dwelling unit rented at fair value to a sibling for use as a vacation home. It also explicitly instructs the model not to assume its sources are verified.

The result is the difference between an answer you can defend and one you can’t.

Don’t stop at a confident-sounding answer

AI models are consistently confident, and confidence is not correctness. A real citation can still be incomplete or misapplied to your facts. This isn’t hypothetical: an attorney representing himself in a tax court case (Clinco v. Commissioner, T.C. Memo. 2026-16) cited AI-generated case law with incorrect citations and interpretations. It was obvious enough to the judge that the attorney was called out for it on the record.

Missing the tax year is its own trap. For 2026, there’s a new 0.5% floor on individual charitable contribution deductions that didn’t exist in 2025 or earlier years, a distinction that changes the answer entirely if the model, or you, gloss over it. Verify every material proposition. Confirm the tax year, review the definitions and exceptions against your client’s specific facts, and check for newer or conflicting authority.

I ran into this myself in the months after the 2025 tax legislation passed, when AI models were still confusing the House version, the Senate version, and the final law.

Know your authority hierarchy

This is really important. Primary authorities such as the Internal Revenue Code, US Treasury regulations, and judicial decisions establish and interpret the law. Secondary authorities, including IRS publications, form instructions, and tax research articles, explain and summarize the primary authority.

AI-generated output sits below both. It’s a useful pointer toward a primary authority, but it is never the authority itself. IRS Publication 527’s guidance on vacation homes is a good example. It is helpful explanatory guidance, but not a controlling authority, so it’s no different from a well-written AI summary.

For the cabin example, that hierarchy leads you to Code Section 280A(g), which addresses the fewer-than-15 days rule and the used-as-a-residence test. Section 267’s family member definitions matter enormously here. A sister counts as family under the code, but a brother-in-law does not. That distinction alone can change whether a family member’s days count as personal use.

Build the verification habit

Every AI-generated proposition needs a checklist, and once you’ve been through the checklist, only then validate the answer for delivery:

• Locate and open the original authority.
• Confirm the tax year.
• Read the surrounding regulatory context.
• Trace definitions and cross-references.
• Review exceptions and limitations.
• Check precedential value.
• Recheck any calculations.

Then, document what you verified—the research question, the facts you confirmed, the authority you relied on, and how you applied it—in a tax memo. I don’t trust my memory a week later, and you shouldn’t either. If a client, auditor, or tax court ever asks how you reached a conclusion, the memo is your answer.

What the IRS says about AI in practice

In June 2026, the IRS Office of Professional Responsibility issued guidance addressing AI use directly, reinforcing Circular 230 due diligence in an AI context. It doesn’t prohibit AI; the IRS recognizes it’s now part of how we work. However, Circular 230 spells out three obligations:

1. Technical diligence: a human review of citations, facts, and calculations.
2. Professional competence: staying grounded in the tax law itself, not just in what a model tells you.
3. Data security: never entering client personally identifiable information or confidential information into an AI model, paid or not.

This same framework works just as well for state income tax or sales tax research as it does for federal questions. Remember that AI can accelerate the process, but it cannot own the answer. That’s still on us.