AI Hallucinations in Tax Work: How CAs Should Check the Answer

A tax answer can be neatly formatted, confidently worded and still be wrong. AI may generate an incorrect section reference, a non-existent case citation or a rule that applied to a different period.

This is commonly called a hallucination. The practical response is not to stop using AI for everything. It is to separate drafting and explanation from facts that must be verified before they affect professional work.

Recognise the errors that matter

In a CA office, risky errors can include an old threshold presented as current, a proposed amendment described as enacted, an exception omitted from a summary or an inferred deadline shown as a fact.

Another common problem is scope: a response may combine company law, accounting guidance and income-tax treatment as though they were interchangeable. Fluent wording can hide that confusion.

Treat a request for more confidence differently from a request for evidence. “Are you sure?” may produce another confident paragraph without resolving the underlying gap.

Step 1: define the question precisely

State the relevant jurisdiction, entity type, transaction, period and purpose. A learning example about an earlier year is not the same task as a current filing decision.

Before answering, list the facts needed to determine which provision applies. Use only the facts I supply. Mark missing information. Separate general background from a conclusion on this case. Do not invent rates, section numbers, filing dates or case citations. Facts and question: [insert].

A useful assistant may first ask for a missing period or entity category. That is preferable to a quick answer based on the wrong assumptions.

Step 2: work from the original material

For a notice, agreement or circular, supply the relevant approved text and ask for a source-bound extraction. For a current legal question, locate and open the applicable official provision and amendments.

Where live search is available, require the response to distinguish material actually retrieved from information recalled by the model. If the source cannot be accessed, that limitation must remain visible.

A real link is not enough: check that it supports the precise claim being made for the period in question.

Step 3: make a claim-check table

Review the draft answer below. List every claim about a section, threshold, rate, deadline, exception or case decision. For each, identify the supporting passage in the supplied material or mark “Not verified”. Separate calculations and assumptions. Do not fill unsupported claims with general knowledge. Draft and material: [insert].

This is a review aid, not independent assurance. The professional must check the passages because the assistant can also misstate what a source contains.

A fictional example

Suppose a circular permits relief only for a specified category. An AI summary says the relief applies to all taxpayers. The date and document number may be correct, but the conclusion is still wrong.

The reviewer should check the qualifying paragraph, restore the limitation and assess which clients are actually covered. Merely confirming that the circular exists would miss the error.

Step 4: verify calculations independently

For arithmetic, use a calculator or tested spreadsheet rather than accepting the explanation as evidence. Check inputs, signs, units, rounding and period boundaries.

A calculation can be mathematically correct but based on the wrong legal assumption. A separately verified formula therefore does not establish that the deduction, rate or period is applicable.

For a proposed Excel formula, test normal cases and edge cases such as blanks, duplicate identifiers and dates stored as text before putting it into a live working.

Step 5: record the conclusion and uncertainty

Keep a short working note containing the actual question, verified material, important assumptions and reviewer conclusion. Preserve unresolved points and obtain further advice where the issue requires it.

Do not add a reassuring disclaimer to an unverified client answer and consider the problem solved. The unsupported content must be corrected or removed.

What a second AI tool can and cannot do

Claude, Gemini or another model can help challenge an explanation or identify missing questions. However, two models agreeing does not prove the law, and disagreement does not show which answer is correct.

Resolve the issue from the source and facts, not by counting how many tools repeat the same sentence.

Frequently asked questions

Can AI be used for routine tax explanations?

Yes, as a draft or learning aid with appropriate review. Keep technical conclusions grounded in the applicable material.

Is a response with citations always reliable?

No. Confirm the source exists, read the relevant passage and check whether the citation actually supports the statement.

Should every short email go through a legal research exercise?

No. Match the review to the task. A wording-only email needs factual checks; a tax conclusion needs legal and factual verification.

Use confidence carefully

assureOffice's approach to practical technology keeps human review in the workflow. Let AI organise questions and evidence, but do not treat fluent language as proof.