The AI For Series

Where AI gets Indian tax and audit wrong: seven limits every CA should know

By Chandan Gupta, CA, CIA · 2026-11-01

AI tools are fast, fluent and usually right. In a CA's work, "usually" is the problem. A working paper carries your name; a tax position carries your client's money. So before asking what AI can do for a practice, it's worth knowing exactly where it goes wrong, and how to spot it when it does.

One distinction explains most of it.

Recalled answers vs computed answers

Some answers are recalled: a section number, a rate, a due date or a case name, produced from what the model learnt in training. Others are computed: a script reads your file and counts, matches or totals it, and reports figures you can trace to rows.

Recalled answers can be wrong, and nothing in the tone will tell you. Computed answers can be wrong too, but they come with evidence: row counts, totals, row numbers. The skill is to turn as much work as possible into the second kind, and to check the first kind against the source every time.

Here are the seven limits, with what to do about each.

1. Made-up sections, rates and case law

Ask a question of law and you get a fluent answer with a section, a rate and perhaps a judgment. Much of it will be right. But a section number may belong to a different provision, a rate may be one that applied two years ago, or a case may be invented outright: a plausible party name, a year, a citation in the right format, and no such judgment. Or the case is real, but it held the opposite.

What to do: treat every section, rate, threshold, due date and citation as a lead, not an answer. One question exposes most of these: "Did you read this in this session, or is it from memory?" Never put a case into a submission you haven't read yourself.

2. Stale knowledge of recent changes

A model learns from material up to a cut-off date, and Indian tax law moves faster than any training run. A Budget change, a CBIC circular, a revised ICAI pronouncement or a replaced MCA form may be missing or half-known.

The structural changes are the dangerous ones. With the Income-tax Act, 2025 applying from 1 April 2026 while FY 2025-26 still governed by the 1961 Act, an AI may quote old section numbers, new ones, or a mix, without warning you. Smaller changes catch people too: a TDS threshold quoted from before the Finance Act, 2025, stated with full confidence.

What to do: put the current text in the working folder (the notification, the circular, the relevant chapter of the Act) and say: "For the law, use only the documents in this folder. Quote the paragraph and page for every rule you rely on. If anything you'd say from memory differs, tell me, and follow the documents."

3. Misreading scans and handwriting

AI can read scanned invoices and photos, often well. But a faded 8 becomes a 3, a rubber stamp hides a digit, and a signature across an amount comes back as a confident wrong figure. It won't always say which readings it was unsure of.

What to do: prefer native exports (CSV from net banking, a Tally export, not a printout). When you must use scans, ask it to mark every uncertain figure in an "Unsure" column, and check every figure that feeds an amount against the paper.

4. Arithmetic done in its head

Language models are not calculators. Ask one to "quickly total" a pasted column and it may simply produce a number, which can be off by a few rupees or a few lakhs, and looks exactly like a correct total.

What to do: never accept a figure that wasn't produced by a script or a formula. Ask for "an Excel with formulas, not values", and for control totals that tie back to the source. Scripts can be wrong too (Invoice Value used where you meant Taxable Value, 03/04/2025 read as 4 March, a Grand Total row counted as an invoice), so ask it to state which columns it used, how it read dates and which rows it dropped.

5. Overconfidence

AI sounds the same when it is right and when it is wrong. It may say "everything looks fine" after reading a sample, report "about 120 mismatches" when it estimated, or answer a slightly different question (month-wise when you asked vendor-wise) and present it as done.

What to do: judge by the evidence, never the tone. If an answer gives you nothing you can check, ask what it checked, how many rows, and what could make it wrong.

6. Context limits on long documents

AI tools have a working memory, the context window. A long loan agreement or a full-year ledger can exceed it, and when that happens earlier material is summarised to make room. Detail gets lost: a clause on page 40, or an instruction you gave an hour ago. A list of "all transfer restrictions in the SHA" can come back neat, complete in appearance, and missing the one buried in a schedule.

What to do: for data, use a script, which reads every row. For long documents, work in parts and ask it to report the page range it actually read in each part, so you can see nothing was skipped.

7. It cannot take responsibility

AI can't hold a membership, sign a report, generate a UDIN or answer to a regulator. Everything it produces is a working paper until a qualified person has reviewed it. No better model will fix that, because it is the basis of the profession.

Jobs to keep for yourself

Used this way, AI does the mechanical part of the job, and the judgement and the signature stay with the CA.


Chandan Gupta, CA, CIA, is the author of AI for CAs: A practical guide to Claude Code for tax, audit, advisory and compliance work, a hands-on guide with three tested walkthroughs on practice data and 100+ prompts. Chapter 13 covers these limits in detail. The prompts and practice files are free at topmate.io/chandan_gupta31

Independent publication; not affiliated with or endorsed by Anthropic. Not professional advice.

This guide is a short version of Chapter 13 of AI for CAs, which walks through it end to end on practice files.

Related

Go further

Get the PDF (₹499) Google Play Books Kindle on Amazon.in Read the free sample Free prompts + practice files