The rise of capable large language models has changed the way many small business owners approach tax research. A question that used to require a call to the accountant or an hour with the IRS website can now produce a coherent, confident answer in under a minute. The answer is also frequently wrong in ways that are difficult to detect without domain expertise.
The issue is not that AI cannot be useful for tax work, it can, in specific applications. The issue is the mismatch between how confident the answer sounds and how reliable it actually is. A licensed professional with malpractice liability cannot substitute AI judgment for their own, and neither should a taxpayer preparing a return or making a filing-level decision.
What AI Tax Research Does Well
Several categories of tax research work reliably with current AI tools, particularly when the user provides context and verifies the output.
| Application | Why It Works |
|---|---|
| Explaining a concept in plain language | LLMs are trained on extensive educational content and summarize well |
| Outlining the structure of a problem | Identifying the relevant question types and decision points is pattern recognition |
| Drafting correspondence and disclosures | First drafts compress substantially; editing remains straightforward |
| Organizing and summarizing your own records | Works on data you provide rather than on training-data recall |
| Comparing high-level differences between entity types | Conceptual comparisons are well-represented in training data |
For each of these applications, AI functions as a fast research assistant. The user still reads, edits, and verifies, but the starting point arrives in minutes instead of hours.
Where AI Tax Research Fails
The failure mode that matters most in tax work is hallucinated citations. An LLM asked "what IRC section governs this?" will very often produce a citation that looks correct, reads confidently, and does not exist. This failure is uniquely dangerous in tax because:
First, tax positions must be supportable under audit, a fabricated citation is worse than no citation. Second, the surface appearance of a correct citation (Section 199A, Rev Proc 2024-13, etc.) is indistinguishable from a real one without verification. Third, taxpayers and non-specialist advisors often do not have immediate access to tools that can verify whether a cited authority exists and says what the AI claims.
Recent testing across major general-purpose LLMs continues to find hallucinated citations at rates between 15 and 30 percent on tax-specific research queries. The rate drops dramatically when the tool is a purpose-built professional tax research platform (Checkpoint Edge, Blue J, Thomson Reuters' AI Assistant) because those tools retrieve from curated current-law databases rather than relying on training-data recall.
What You Still Need a CPA or Enrolled Agent For
The work that requires a licensed professional is not arbitrary, it is the work where the stakes, the liability structure, and the complexity combine to make unsupervised AI output unsafe.
Signing a Tax Return
Federal law requires a paid preparer to sign returns they prepare and subjects the preparer to penalty for negligent or willful inaccuracies. AI cannot hold this license, cannot carry this liability, and cannot sign the return. The licensed professional is the only path to a filed return with preparer representation.
Representing You Before the IRS
Only attorneys, CPAs, and Enrolled Agents can represent taxpayers before the IRS in examinations, appeals, and collections. AI tools cannot respond to an IRS notice on the taxpayer's behalf, cannot attend an examination, and cannot file a Tax Court petition. When the IRS engages, a licensed professional is the only option.
Making Filing-Level Positions on Uncertain Law
Any position with a reasonable-basis, substantial-authority, or more-likely-than-not standard requires the judgment of a licensed professional. AI can outline the issue and the arguments; it cannot determine whether a specific position meets the threshold on a specific set of facts.
Entity Structure Decisions With Long-Term Consequences
S-corp elections, partnership conversions, and C-corp formations each have ten-plus-year consequences that an AI tool is not equipped to weigh against the specific facts of the business. These are human-judgment calls informed by facts the AI does not see.
Audit Defense and Penalty Abatement
Every audit is fact-specific, and every abatement request requires positioning tailored to the specific penalty, facts, and history of the taxpayer. AI drafting is useful at the starting-point level; the winning submission is the result of professional judgment applied to the draft.
The Right Way to Use AI in Tax Research
A disciplined workflow for AI-assisted tax research looks less like "ask AI a question and accept the answer" and more like "use AI to accelerate the specific parts of research where the failure mode is acceptable."
The safest and highest-value use of AI for tax research is as a drafting and organization tool, not as an authoritative source. Outline your question with AI, draft your correspondence with AI, summarize your own records with AI. Verify every citation against the actual source, and route any filing-level decision to a licensed professional before it becomes a filed position.
The Verification Protocol
Every AI-generated tax research output must pass through the verification protocol below before it is relied upon for any decision with filing or penalty consequence.
- Every cited IRC section, opened at IRS.gov or Cornell LII and confirmed to exist and say what the AI claims
- Every cited Revenue Procedure or Revenue Ruling — verified at IRS.gov with matching number and publication date
- Every cited court case — confirmed via the actual court opinion, not AI paraphrase
- Every numerical figure (bracket thresholds, deduction limits, penalty amounts) — verified against IRS Revenue Procedure 2025-32 or the relevant current guidance
- Every conclusion reviewed by a licensed CPA or Enrolled Agent before it informs a filing position
- Any IRS correspondence response drafted with AI reviewed line-by-line before submission
AI With a CPA vs. AI Instead of a CPA
The contrast below is not rhetorical. Each row is a field-observed pattern from the last 18 months of client work.
AI Instead of a CPA
- Tax position based on LLM answer that sounds authoritative
- Citation that does not exist relied upon in correspondence
- Penalty abatement request drafted by AI and submitted unreviewed
- Entity structure decision made from an AI comparison table
- IRS notice responded to with AI-drafted letter, no professional review
AI With a CPA or EA
- Research outlined by AI, verified against primary source, reviewed by licensed professional
- Every citation in client correspondence confirmed against actual IRS or court record
- Abatement request drafted with AI, edited by EA, submitted with professional representation
- Entity structure AI overview used as starting point; CPA applies the fact-specific decision
- IRS notice acknowledged by AI draft, rewritten by professional, signed under professional authority
If you use a general-purpose LLM for tax research, install the habit of ending every query with the instruction: "Cite only primary sources that exist. If you are uncertain whether a citation is real, say so." The instruction does not eliminate hallucinations, but it substantially reduces them, and it conditions you to treat uncertainty as the default rather than the exception.
A Practical Two-Tool Approach for Small Business Owners
Most Florida small business owners who use AI for tax questions will benefit from a simple two-tool workflow:
General-Purpose LLM for Framing and Drafting
Claude, ChatGPT, or Copilot for outlining a question, drafting correspondence, or summarizing your own records. Treat output as a first draft, never as an authoritative source.
Primary Source Verification for Any Cited Authority
IRS.gov for revenue procedures and rulings, Cornell Legal Information Institute for IRC sections, PACER or court websites for case citations. Every cited authority gets verified before it's relied upon.
Licensed Professional for Filing-Level Decisions
Every tax position that will appear on a filed return, every IRS correspondence response, and every abatement or audit response routes through a CPA or Enrolled Agent for review and signature.
Frequently Asked Questions
The tools can help you understand the forms, explain what a line means, and outline how a specific situation might be reported. They cannot reliably produce a filing-grade return on their own, hallucinated figures and outdated law are too frequent. For a simple W-2 only return, commercial tax software (TurboTax, H&R Block) is purpose-built for this and more reliable than an LLM. For anything with self-employment income, rental properties, partnerships, or unusual items, a licensed preparer is the appropriate channel.
Substantially. Purpose-built tools like Thomson Reuters' Checkpoint Edge AI, Blue J's tax analysis platform, and CCH AnswerConnect's AI features retrieve from curated current-law databases rather than from general training data. Hallucination rates drop sharply, citations are verified against actual IRS and court sources, and the output is designed for professional use. These tools cost $100–$400 per month and are generally available only through professional subscriptions.
Ask them why, and then decide. Some firms deliberately avoid general-purpose LLMs for client work because of data-privacy and accuracy concerns, which is a defensible position. Others have integrated professional AI tools behind the scenes and use them to research faster without advertising the workflow. The question to ask is not "do you use AI" but "how do you verify the current-year accuracy of the positions on my return?" A thoughtful answer either way reflects professional judgment.
For general-purpose consumer LLMs, yes, unless you are using an enterprise tier with explicit data-retention controls. For professional tax research tools with SOC 2 certification and tax-specific data-handling commitments, no. Before pasting any client-identifying tax information into any AI tool, verify the data-handling policy. If you cannot verify it, do not paste.
Sources
- Thomson Reuters Institute. 2026 State of Tax Professionals Report — AI Tools and Accuracy Benchmarks.
- American Institute of Certified Public Accountants. AI in Tax Practice — Accuracy and Reliability Study, 2026.
- Internal Revenue Service. Circular 230 — Regulations Governing Practice Before the Internal Revenue Service. IRS.gov
- Cornell Legal Information Institute. Internal Revenue Code — Primary Source. law.cornell.edu/uscode/text/26
- Stanford HAI. AI Index Report 2026 — Hallucination Benchmarks Across Large Language Models. aiindex.stanford.edu
Discussion
Comments
Leave a comment