AI for Indian Tax Research: Common Failure Modes

General-purpose AI assistants fail on Indian tax questions in recognisable ways: citing repealed provisions, presenting outdated rates confidently, blending Indian law with foreign concepts, and generating citations that do not exist. Each failure mode has a specific check.

Stale law presented as current

The most consequential failure right now. An assistant drawing on content written before April 2026 will cite Income-tax Act, 1961 provisions as though current, and content written before September 2025 will state superseded GST rates.

The answer will not signal the problem, because the underlying source did not signal it either — it was correct when written.

The check: ask what date the position reflects and which Act it refers to. For anything rate-related, treat the figure as unverified until checked.

Confident wrong numbers

Rates, thresholds, limits and deadlines are exactly the content most likely to be wrong and least likely to be flagged as uncertain, because numbers are stated with the same fluency as everything else.

The check: treat every number as unverified. This is the same standard you would apply to a number from any secondary source.

Invented or misattributed citations

Section numbers, circular numbers and case citations that do not exist, or that exist but say something different. Where the underlying corpus disagrees with itself — as it currently does on the 2025 Act section mapping — the assistant may confidently pick the wrong one.

The check: open the citation. If it cannot be opened, it does not exist for the purposes of a deliverable.

Jurisdiction drift

Answers blending Indian law with concepts from other jurisdictions, particularly on transfer pricing, permanent establishment, and GST versus VAT, where terminology overlaps but rules diverge. The result reads plausibly to a non-specialist.

The check: ask which country’s rules the answer applies, and be alert to terminology that is not Indian usage.

Missing the exception

Assistants tend to state general rules well and miss the proviso, exception or threshold that reverses the answer in a specific case. Indian tax law is dense with these.

The check: ask explicitly for exceptions and provisos, then verify them in the provision.

Where these tools are genuinely useful

None of this is an argument against using AI in tax practice. Used for orientation on an unfamiliar area, for generating the list of issues a transaction raises, for drafting a first pass at a response structure, or for narrowing a research question before going to primary sources, the time saving is real.

The distinction is between using it to find where to look and using it to decide what the law is. The first is a legitimate productivity gain. The second imports a verification obligation the tool does not discharge.

The practical test for any tool used in tax research: does it take you to the primary source, or does it stand between you and the primary source?

Related reading

• a working verification method — How to Verify a Tax Answer Against Primary Sources
• the 2025 Act transition — Income-tax Act, 2025 vs the 1961 Act: What Actually Changed