Why Are So Many CPAs Already Using AI for Tax Research Without a Process?

Most tax preparers already ask an AI tool a research question. Few have a process for what happens after they get an answer.

That gap is the real risk, not the AI answer itself. A client calls about a home office deduction after a mid-year move. You get a fast answer from an AI tool and relay it. Three weeks later you can't remember exactly what you asked, what the tool said or how you checked it.

The IRS Office of Professional Responsibility addressed this directly in June 2026. Its first formal guidance on AI use in tax practice said long-standing Circular 230 duties, due diligence, competence and confidentiality, apply to AI the way they apply to any research tool: "Practitioners must remain vigilant in reviewing what is produced by AI, validating its factual assertions and citations." AI-generated research is a starting point, not a finished product.

The five prompts below turn that requirement into a habit. Each one has a job. Together they take you from a client's question to a documented, defensible answer.

The Five-Prompt Chain That Turns a Fast AI Answer Into Something Defensible

What does a complete AI-assisted research workflow actually look like? It's five prompts run in order, not five separate tools you pick from.

PromptPurposeWhat You Get
1. ResearchGet a first-pass answer with sourcesA summary with named citations to check
2. TranslateTurn the technical answer into plain languageA client-ready explanation, no jargon
3. VerifySanity-check the answer against real authorityA list of what to confirm before you rely on it
4. SummarizeDraft the follow-up in writingAn email documenting the advice given
5. EscalateCatch questions too risky to answer aloneA flag telling you to stop and get a second opinion

Prompt 1: Research the Question

Use this to get a fast, sourced starting point, not a final answer.

"I'm a CPA researching [specific tax question, with relevant facts: entity type, tax year, state]. Summarize the current federal tax treatment, cite the specific IRC section, Treasury regulation or IRS guidance that applies, and note if there's a state-level variation I should check separately. Flag anything that depends on facts I haven't given you."

The last sentence matters most. It forces the tool to tell you what it's guessing at instead of quietly filling gaps.

Prompt 2: Translate It for the Client

Use this before a call, not during one.

"Rewrite this tax research summary for a client with no accounting background. Explain what it means for them in plain English, in under 150 words. Do not use technical terms without explaining them in the same sentence. End with one clear next step they need to take."

This saves you from talking in code section numbers on a client call. It also gives you a script, so you say the same thing every time you get this question.

Prompt 3: Verify It Against Authority

This is the step most preparers skip. It's also the one the IRS just said is non-negotiable.

"List every factual claim, citation and calculation in this research summary as a separate checklist item. For each one, tell me exactly what I need to look up to confirm it, IRC section, Treasury regulation, Revenue Ruling or current IRS publication. Flag if any citation looks like it could be outdated or fabricated."

Asking the AI to flag its own possible fabrications doesn't replace your judgment. It gives you a checklist instead of a blank page, the difference between verifying in five minutes and forty.

Prompt 4: Draft the Follow-Up Email

This prompt turns a phone conversation into a record.

"Draft a follow-up email to a client summarizing the tax advice we discussed on our call. Include: what they asked, what the guidance says, what they need to do next and the deadline if there is one. Keep it under 200 words. Do not include anything I haven't already verified."

That last line is there on purpose. Don't run this prompt until you've finished Prompt 3. This is the same audit trail KPMG's 2026 survey pointed at when it found firms able to show what their AI did and how a human reviewed it saw error reduction improve three to six times faster than firms that can't. That data came from large finance teams, but the principle scales down to a single client file just as well.

Prompt 5: Flag When It's Too Risky to Answer Alone

Use this before you send anything, as a final check.

"Based on this research and the facts given, tell me honestly: is this a question with a clear, well-established answer, or does it involve ambiguous facts, conflicting authority or a position aggressive enough that I should get a second opinion or consult a specialist before advising the client? Explain your reasoning."

This prompt won't catch everything. It's a second set of eyes, not a substitute for your own judgment. But it costs ten seconds, and it catches the question that looked routine until the AI flagged the conflicting authority underneath it.

Is It Safe to Put a Client's Tax Question Into ChatGPT?

Not with identifying details attached, and not on a free consumer account. Strip the client's name and account details before you ask.

That's the clearest instruction in the IRS's June 2026 guidance, which tells firms to use "only secure, enterprise-approved AI" and avoid public platforms where sensitive data could be exposed. Run the research on the fact pattern, not the file, unless your firm has an enterprise agreement covering data handling.

The guidance also expects you to understand how your AI tool generates answers and where it's likely to get things wrong, the same way you're expected to know the limits of any research database.