By Thomson Reuters Institute’s 2026 count, generative-AI use among tax and accounting professionals has roughly doubled in a year, and most firms now use it or plan to. So “can I trust AI for tax research?” needs a follow-up: trust it to do what?
The answer depends on what you ask AI to do and what sources it can use. Drafting an explanation is different from identifying the authority that governs a client’s position. Finding a real citation is also different from applying it correctly.
One answer, two different jobs
When you ask a model a tax question, it may do two jobs in the same paragraph. Language work includes framing the issue, explaining a rule, and organizing the prose. Authority work means identifying the governing Code section, regulation, guidance, or case.
On screen, they look equally polished. They do not carry the same reliability.
The language work can be useful. A capable model can sharpen a vague client question into a more precise research issue and explain a supplied provision clearly, but the output can still omit a fact, exception, authority, or date that changes the result.
Authority work needs a different control. A model may give you a confident reference to §199A(d)(4)(C), even though §199A(d) stops at paragraph (3). For why that happens, see the full explanation.
A clear explanation can rest on an invented citation, a real citation that does not support the sentence, an outdated rule, or an incomplete source set. The tone will not reliably warn you.
Where AI helps—and what still needs review
Instead of labeling tasks “safe” or “unsafe,” match each task to the check it needs:
| Task | How AI can help | What you still need to check |
|---|---|---|
| Explain a concept or supplied rule | Draft a clearer explanation and identify follow-up questions | Check the explanation against the operative text, definitions, exceptions, and facts |
| Draft or restructure a memo or client letter | Organize verified facts and authority into readable prose | Give it verified sources, check every material claim, and keep the conclusion with the practitioner |
| Summarize a document you supplied | Locate themes, provisions, and possible issues quickly | Confirm quotations, page references, omissions, and whether another authority controls |
| Brainstorm issues or research paths | Produce a starting checklist | Treat the list as incomplete until a professional reviews it |
| Identify governing authority | Suggest provisions and search terms | Retrieve and open current primary authority; check hierarchy, year, and jurisdiction |
| Quote law or state a threshold, date, or calculation | Extract or calculate from supplied material | Compare word-for-word, confirm the effective period, and independently recalculate |
Use the model to move faster, but verify anything that must be exact. A plausible-looking subsection, threshold, date, or calculation can be harder to catch than a broad conceptual mistake.
What changes when AI can retrieve sources
Before deciding how much to trust an answer, ask what the system used. It may be relying on model memory, searching the web, reading documents you supplied, or querying a connected research source. Those are different workflows. OpenAI’s current web-search documentation, for example, describes search as a separate tool that gives the model current information and citations.
When a tool retrieves a section number and text from primary authority, it gives the model source material rather than asking it to produce the citation from memory. That makes the result easier to audit. It does not prove that the source is complete, controlling, current, or correctly applied.
Retrieval changes a citation from an unsupported string into evidence you can inspect. Professional review determines whether that evidence actually supports the conclusion.
That distinction is a useful way to compare AI tools for tax professionals.
Accuracy is more than one number
An answer can be right in one way and wrong in another. Check these separately:
- Citation existence. Does the cited section, regulation, notice, ruling, or case exist?
- Citation support. Does the cited text support the exact proposition attached to it?
- Legal correctness. Is the authority controlling or otherwise appropriate for this conclusion?
- Completeness. Did the answer miss an exception, definition, contrary authority, or required source type?
- Effective-date accuracy. Does the rule apply to the relevant tax year, transaction date, and jurisdiction?
- Calculation accuracy. Do the inputs, thresholds, phaseouts, and arithmetic reproduce independently?
A retrieval benchmark can show how one workflow performed on one set of questions. It is not “ChatGPT’s tax accuracy rate,” and it should not be used to compare ChatGPT with Claude unless the tested models, prompts, responses, grading method, and repeated runs are disclosed.
A three-question citation test
Before you rely on any AI tax answer, whether from a chatbot, a platform, or a purpose-built assistant, run it through three questions:
- Where did the citation come from? Was it model memory, web search, a document you supplied, or a connected source? If the system cannot show you, treat the citation as an unverified lead.
- Can you open it? A real section resolves to a real page on uscode.house.gov or the eCFR. A fabricated one resolves to nothing, or to text plainly about something else. No link, no claim.
- Does the source support the sentence? Read the relevant text and its surrounding definitions or exceptions. A real citation attached to the wrong proposition is still a bad citation.
This quick test catches obvious failures, but it is only the start. The full version, applied end-to-end, is the division of labor in an AI-drafted research memo: the model helps organize and draft, primary authority supplies the legal evidence, and the practitioner keeps the judgment.
So, can you trust it?
Yes, conditionally. Use AI to frame the issue, find sources, compare provisions, and draft. But do not treat fluent writing—or even a real citation—as proof that the conclusion is correct. Open the source, confirm that it supports the claim, look for what is missing, and check the math yourself.
Match the tool to the task, show where the answer came from, and keep professional review over every material tax conclusion.
This article is general information, not tax advice. Capabilities of AI tools change quickly; confirm any citation against primary source before relying on it for client work, and test any tool against the questions above.