Six months ago I almost submitted a research memo with a fake IRC citation in it.
I'd asked ChatGPT to help me research home office deductions for an S-Corp client. It produced a clean analysis citing "IRC Section 280A(g)(2)" with specific language about owner-employee deductions. The format was perfect. The wording sounded authoritative. I had no reason to doubt it.
My manager caught it. Section 280A(g)(2) doesn't exist.
That moment changed how I use AI for tax research. Not whether — how.
If you're an accountant trying to figure out where AI fits into your tax research workflow, this guide is the version of this article I wish I'd read before that meeting with my manager.
Let's start with where AI genuinely helps. There are four specific tasks where I now use AI as the first step in my research workflow, every time.
When you get a tax question you've never researched before, the hardest part is often figuring out where to start. What sections of the code apply? What recent guidance exists? What are the major positions on this issue?
This is where AI shines. A well-structured prompt — "I need to research [question]. Give me the relevant IRC sections, key Treasury Regulations, recent IRS guidance, and major court cases that touch on this area" — gives you a research roadmap in 30 seconds.
You wouldn't trust the actual citations without verification. But you've now narrowed your search from "the entire Tax Code" to "these 4 sections and 2 court cases." That alone saves an hour.
If you've ever stared at IRC Section 199A and wished someone could just explain the QBI deduction without the jargon, AI is excellent for this.
The trick: read the actual section yourself first (or pull it up alongside). Then ask AI to explain it. The combination of source text and AI explanation is faster than reading either alone.
Where this gets dangerous: if you skip reading the source and just trust the explanation. AI can simplify accurately or simplify wrongly, and you can't tell the difference without the source.
Once you've done the research, AI is excellent at organizing it. Standard memo structure (Issue / Facts / Analysis / Conclusion) is exactly the kind of repetitive task AI handles well.
Workflow: write your research notes in bullet points, paste them into ChatGPT or Claude with a prompt like "Format this as a tax research memo with standard sections," and you get a polished draft in under a minute.
You still review and edit. But you skip the blank-page problem entirely.
For multi-state work, AI can give you a quick comparative view. "How do California, New York, and Texas treat S-Corp owner-employee health insurance deductions?" produces a structured comparison faster than reading three state tax guides.
Same caveat: verify everything. But the structure of the comparison saves real time.
Now the part that matters more than the wins. There are three places AI fails in tax research, and accountants who don't know these failure modes are setting themselves up for the same kind of meeting I had with my manager.
This is the big one. AI fabricates IRC sections. It fabricates Treasury Regulations. It fabricates court cases. It fabricates Revenue Rulings. And it does so with complete confidence.
Here's the part that makes it especially dangerous: the fabrications follow real citation patterns. "IRC Section 280A(g)(2)" looks exactly like a real subsection because it follows the format of real subsections. "Smith v. Commissioner, 142 T.C. 234 (2018)" looks exactly like a real case because it follows the format of real cases.
You cannot tell from looking whether a citation is real or fabricated. The only way to know is to verify against the actual source.
This is non-negotiable. Every IRC section, every Treasury Reg, every case, every Revenue Ruling — open the actual source on the IRS website or in your research database and verify the citation says what AI claims it says.
AI models are trained on data with a cutoff date. ChatGPT and Claude don't automatically know about tax law changes that happened after their training data was collected.
This is a real problem because tax law changes constantly. Provisions from the Tax Cuts and Jobs Act sunset. New legislation passes. The IRS issues new guidance. AI may give you advice based on rules that are no longer current.
For any tax research where the rule might have changed in the last 12-18 months, verify against current IRS publications. Don't assume AI knows about the latest guidance.
Tax law is full of provisions that look simple but have significant exceptions, limitations, or interactions with other code sections. AI is excellent at producing the simple version. It's not as reliable on the nuance.
Example: AI will tell you (correctly) that S-Corp owners can take the QBI deduction. What it might miss: the wage limitations, the SSTB restrictions for certain professions, the phase-out thresholds, and the interaction with reasonable compensation rules.
If your research question has multiple moving parts, AI gives you the first layer accurately. The deeper layers are where you need to dig manually.
After my manager-meeting wake-up call, I built a 3-step workflow that I now use for every tax research task. It's slow enough to catch AI mistakes and fast enough that AI still saves me time overall.
Start with a structured prompt that frames the research question, asks for relevant authorities, and explicitly tells AI to flag uncertainty.
The prompt I use:
"I need to research the following tax question: [question]. For my client's situation: [relevant facts].
Provide:
1. Brief answer (yes/no/it depends)
2. Relevant IRC sections
3. Applicable Treasury Regulations
4. Recent IRS guidance (Rev Rulings, Notices, etc.)
5. Key court cases if applicable
6. Common pitfalls or audit triggers
7. Recent changes I should be aware of
IMPORTANT: This is a research starting point only. I will verify all citations independently. If you are not certain about any citation, explicitly say so. Do not guess or fabricate."
The "do not guess or fabricate" line doesn't eliminate hallucinations entirely, but it noticeably reduces them. AI is more likely to hedge when explicitly told to.
This is the step nobody likes, and it's the step that determines whether you stay employed.
For every citation AI produced:
For anything that doesn't verify, you have two options: find the real source manually, or remove the claim from your analysis. Never include an unverified citation in client work.
This step takes 10-20 minutes for a typical research question. It's the trade-off for the time AI saves you on the front end.
For your own protection and for the next person who might pick up this file, document your research process.
Include in your workpapers:
This last point matters. Your firm probably has a position on AI use. Document that you used it appropriately.
Quick comparison based on six months of daily use:
Claude (Pro plan) is my primary research tool. It's noticeably more cautious about citations than ChatGPT. When it doesn't know something, it's more likely to say so. It also handles longer research questions better — you can paste in a full set of client facts and get a more thorough analysis.
ChatGPT (Plus plan) is my secondary tool. I use it for the more creative parts of research — drafting memos, comparing positions, generating questions to consider. For pure citation work, I trust Claude more.
Specialized tax AI tools are improving but still expensive and inconsistent. Most accountants will get more value from learning to use Claude or ChatGPT well than from paying for specialized tools.
Whatever tool you use, the verification step is the same. The tool doesn't change the rule — every citation gets verified, no exceptions.
Some accountants read articles like this and conclude that AI is too risky for tax work. I disagree, but I understand the impulse.
Here's how I think about it: the alternative isn't "use AI carefully" vs "don't use AI." The alternative is "use AI carefully" vs "do the work manually."
Manual research has its own failure modes. Tired accountants miss things. Time pressure creates shortcuts. Junior staff don't know what they don't know. None of this is unique to AI.
The accountants who get in trouble with AI aren't the ones using it carefully. They're the ones using it as a substitute for thinking. The verification step exists specifically to keep AI from becoming that substitute.
Done right, AI is a research multiplier. It gets you to the answer faster, with more thorough background, and with better-organized output. Done wrong, it puts fake citations in your memos and ends meetings the way mine did.
Pick the version of "done right." Your career will thank you.
Want the exact tax research prompts I use, with built-in verification reminders? They're in the free PDF — section 3, prompts 23-30. Download it here.
Get 50 copy-paste prompts delivered to your inbox — free.
Get the Free PDF →