A client asked me last week, half-joking, “Sir, ab toh ChatGPT se hi file kar lenge — aapki kya zaroorat?” I laughed, because honestly, it’s a fair question. I use AI constantly in my own practice. I’ve built reconciliation tools with it, drafted notes with it, and used it to make sense of messy Tally exports running into thousands of lines. So I’m the last person who’ll tell you AI has no place in tax work.
And yet, my answer to that client was an easy one: please don’t let a chatbot file your return on its own.
That instinct just got some company. A recent round of reporting — the LiveMint piece that’s been doing the rounds, echoing what tax leaders told Money control — has tax professionals openly cautioning people against blindly trusting ChatGPT, Claude or Gemini for ITR filing. After an incident where a general-purpose AI got someone’s tax treatment wrong, the question moved from “can it?” to “should you?” I want to add a practitioner’s perspective to that, because I think the nuance matters.
The problem isn’t that AI is dumb. It’s that tax isn’t arithmetic.
Here’s the thing people get wrong about return filing: they think it’s a calculation problem. Add the income, subtract the deductions, apply the slab. If that were all it was, a spreadsheet would have replaced us decades ago.
The real work is judgment. Which ITR form actually applies to your facts? Is that ₹40,000 of FD interest something you “feel” you’ve already paid tax on through TDS, or does it still need to be declared in full? Does that one foreign brokerage account trigger a schedule you’ve never heard of? Is your crypto gain business income or capital gains? Are you sure the small consultancy you did on the side doesn’t change your form altogether?
A general AI model is genuinely good at the arithmetic. It is far less reliable the moment a situation needs interpretation — and as the experts quoted in that coverage pointed out, these models aren’t built on a fixed compliance framework. They’re trained on a vast, general slice of the internet, which means they can confidently hand you a slab rate, a rebate threshold, or a form number that was correct two years ago and is wrong today. With the shift to the new Income-tax Act and the “tax year” framing now in play, “confidently outdated” is a very real failure mode.
AI only knows what you tell it — and that’s exactly the trap.
There’s a comment in that reporting I keep coming back to: AI works well if all the relevant information is given accurately, but taxpayers often leave things out simply because they don’t understand what needs to be reported in the first place.
Sit with that for a second. The danger isn’t that the AI lies to you. The danger is that it answers the question you asked — beautifully, fluently, persuasively — while staying completely silent about the question you didn’t know to ask. It won’t volunteer that your savings interest above the threshold is taxable if you never mentioned the savings account. It can’t reconcile your numbers against your AIS, your TIS or Form 26AS, because it has never seen them. It doesn’t know that the rent in your AIS implies a TDS trail that has to be matched. A model can’t flag an omission it was never shown.
A return prepared by someone who knows what should be there is a fundamentally different thing from a return prepared by a tool that only sees what you remembered to type.
Two risks people underestimate
The first is data. People are casually uploading PAN, salary breakups, bank statements and full financial pictures into general-purpose chatbots. Treat your tax documents like you’d treat your net-banking password — be very deliberate about where they go and what happens to them afterwards.
The second is accountability, and this is the one that should stop you cold. When you click “verify” on that return, your name is on it. Not OpenAI’s, not Anthropic’s, not Google’s. If the calculation is wrong and a demand or a notice follows, you are the one holding the bag — and the Income Tax Department has never once accepted “the chatbot told me so” as a defence. AI carries none of the legal responsibility. You carry all of it.
So where does AI actually belong?
Right beside the professional, not in place of one.
In my own work, AI is a brilliant junior — fast, tireless, great at first drafts, summarising a long statement, spotting a pattern in data, explaining a concept in plain language. What it never does is sign off. Every output gets checked against the law as it stands today and against the client’s actual facts before it goes anywhere. That review is the job. The Income Tax Department itself now offers an AI assistant, “Kar Saathi,” to help taxpayers understand obligations — and even that is positioned as guidance, not as something that quietly files on your behalf.
Think of it like a calculator. A calculator made us faster; it didn’t make accountancy obsolete, because the skill was never the multiplication — it was knowing which numbers to multiply and why. AI is the most powerful calculator we’ve ever had. It’s still a calculator.
So, use it. Ask it to explain a provision. Ask it to help you organise your documents. Ask it what questions you should be asking. Just don’t ask it to take responsibility for a return it doesn’t fully understand, on behalf of a financial life it has only half-seen. The convenience is real. So is the cost of getting it wrong.
For anything with capital gains, foreign assets, crypto, business income or more than one source of income, get a human pair of eyes on it before you file. Not because the technology is useless — I’d be a hypocrite to say that — but because, for now, judgment and accountability are still very human things.
*****
Disclaimer :- The above mentioned information is for general purpose only. Views are my own. Always confirm your specific tax position against the current law and your own facts with tax professonals.
Author Bio
