Real World · For business teams

A neat AI invoice can still be wrong. Five checks before you send.

Chatbots can write a tidy invoice with wrong figures. What the makers say, why a neat layout proves nothing, and five checks to run before you send.

30 Oct 2026 · 5 min read · sources read up to 30 Oct 2026

Here is a made-up example, because we'd rather not borrow anyone's real invoice. It is nine at night, you have been on your feet all day, and a customer is waiting. You ask the chatbot on your phone to write the invoice from the items on the order. A minute later it hands you something that looks better than the ones you make yourself. You send it, a little proud and very tired, and go to bed.

In our example, your customer writes back the next morning. The GST says ₹5,700, and it should be ₹5,900. Your stomach drops, because you were so sure. A tidy page is very easy to trust, and this article is about the pause between "it looks fine" and "sent".

Why a tidy invoice can still be wrong

The people who make the big chatbots say so themselves. Anthropic (Claude), OpenAI (ChatGPT) and Google (Gemini) all say their chatbot can be wrong, and ask you to check important answers yourself ^[1][4][7]. Anthropic and OpenAI add that it can sound sure while it's wrong ^[1][4]. Their terms say the same ^[2][5], and OpenAI says to treat ChatGPT's answer as a first draft ^[4]. Some chatbots can do calculations properly ^[3][4], and we don't know how often any of them slips on an invoice, so we won't guess. A finished-looking invoice tells you nothing about whether the figures are right.

A made-up invoice, and the five checks

To show what each check can catch, we typed up an invoice ourselves. A chatbot did not write it. We put five slips into one invoice so you can see each check catch something. We are not saying a chatbot makes all five at once.

The made-up order is ten chairs at ₹4,000 each and two desks at ₹9,500 each, for a customer we'll call Example Traders.

• Totals. Redo the sum on a calculator, from the lines. Ten chairs make ₹40,000 and two desks make ₹19,000, so the lines add up to ₹59,000. Our made-up draft says ₹57,000.
• Tax. Take the price after discount, before GST is added, and apply the rate your accountant gave you for this job. Pretend that rate is 10 per cent. It is a pretend number we picked to keep the sum easy, and it isn't the rate for any item. The real rate is your accountant's to give. Ten per cent of ₹59,000 is ₹5,900. The draft shows ₹5,700, which is ten per cent of its own wrong total. That is the line your customer spotted.
• Names. Your business name and your customer's, spelled as they are in your records. The draft says "Exmaple Traders".
• Dates. The invoice date and the due date. In the draft, the due date comes before the invoice date.
• Units. The unit you sell in: kg or pieces if you sell goods, hours if you bill for time. The draft lists the chairs by the set. You sell them one at a time.

Each check is a short job with a calculator and your own records. The draft looks just as tidy with the slips in it as it would without them, so looking at it doesn't count as checking.

The tax line is the one to hand to your accountant

We won't give you a rate, and a chatbot shouldn't be your source for one either. OpenAI itself says ChatGPT doesn't replace a tax professional ^[6]. Rates also move: the GST Council reset them in September 2025, and the new rates started on 22 September 2025, with a few exceptions ^[9]. A chatbot may not know about recent changes unless it searches the web ^[4]. We are not saying any chatbot used an old rate. It is just the line where a stale number is easiest to miss, so take the rate from your accountant.

Why these five

These five are our own method, and most of them match fields the government's GST invoice rule asks for on a tax invoice ^[8]. The due date, and "hours" as a unit for people who bill for time, are our own additions that the rule doesn't mention.

Check against your own records

The makers all say to check it yourself ^[2][5][7]. For an invoice, that means your price list, your last invoice to the same customer, and the order itself. Nothing we read says whether asking the chatbot to check its own draft works, so we have no view. While you're at it, keep real customer details out of the chat and type them in yourself afterwards.

Our view

This part is opinion. A wrong GST figure on an invoice you've already sent costs more than the minutes the chatbot saved you. We can't put a number on that, and we haven't looked at what a wrong invoice might cost you in penalties, so we won't say. What we would do is simple. Redo the total, take the rate from your accountant, then read the names, dates and units against the order. Do it when the draft looks perfect, and most of all then.

How we checked

We haven't run this ourselves. Everything here comes from the pages in the source list, which we read on 30 October 2026: the help pages and terms from Anthropic ^[1][2][3], OpenAI ^[4][5][6] and Google ^[7], plus the official GST rule text and the Council's announcement from the Finance Ministry ^[8][9]. The invoice above is one we typed, slips and all. Some sites wouldn't open for our software, so we opened them ourselves ^{[4][5][6][7][8][9]}. Pages change, so read the originals again before you rely on a detail. We have no error rate for any chatbot, and nothing we read suggests one is more likely than another to slip.