Don’t let AI do your sums
Key takeaways
- An AI can sound certain about a number it got wrong, so never take a figure from the model’s head.
- Prices, stock and totals come from your records and from code or a spreadsheet, looked up fresh every time.
- Let AI decide what to work out and explain the result, and make every number show where it came from.
Short version: an AI chatbot can explain a sum and set one out, but a number it gives from its own head can be wrong while sounding certain. For anything that matters, such as prices, totals, discounts and stock, have it fetch the figures from your records and work them out with a calculator, a spreadsheet or code, then check the answer.
Can AI do maths?
Sort of, and that’s the trouble. A chatbot is very good at producing text that looks like a correct answer. For simple sums it usually is correct. For longer ones, with several steps, awkward numbers or a rounding rule, it can slip a digit or drop a step, and it gives the wrong answer in exactly the same confident tone as the right one.
Newer AI tools often hand the sum to a built-in calculator or run a small piece of code behind the scenes. That helps a great deal. But you can’t always tell from the reply whether it calculated the number or just wrote one that looked plausible.
Why AI gets sums wrong
A language model works by predicting what text comes next. It has seen a vast number of sums, so it has a good feel for what an answer should look like. A feel isn’t the same as working it out, in the same way that a confident guess at a restaurant bill isn’t the same as adding it up.
It also doesn’t know your numbers. Ask it what a trade customer’s order comes to and, without your price list in front of it, it will happily use a price it made up.
Where it goes wrong in a small business
Say you sell garden planters online. The sums that cause real damage tend to be ordinary ones:
- A quote for forty planters with a trade discount and a delivery charge.
- A reply telling a customer how much they’ll get back for a part return.
- A stock answer, “yes, there are twelve left”, when the AI is guessing.
- Adding up a month of orders for a quick report.
Each one looks fine on the screen. The mistake only shows when the customer pays the wrong amount or the planters aren’t on the shelf.
Numbers come from tools, not from the model’s head
The rule is simple: the AI may decide what to work out, but the number itself comes from something that actually calculates.
- Facts from your records. Prices, stock and order totals come from your shop, stock or accounts system, looked up fresh, never remembered from earlier in the conversation.
- Sums from code or a spreadsheet. Discounts, totals and rounding are done by a formula you’ve checked once and trust. If the AI is connected to tools, give it a calculator tool and tell it to use it for every figure.
- Rules written down once. Your rounding, minimum order and discount rules live in one place the tools read, not scattered through the AI’s instructions. There’s more on that in One home per fact.
- Show the working. Any reply with a number in it should say where the number came from, so whoever checks it can see at a glance.
What AI is good at with numbers
None of this means keeping AI away from figures altogether. It’s useful for the jobs around the sum:
- Turning a messy email into a clear list of what the customer wants to order.
- Writing the spreadsheet formula you couldn’t remember, which you then test on a few rows.
- Explaining what a report shows, once your own system has produced the report.
- Noticing that a number looks odd and flagging it for a person.
Can AI do accounting?
It can help with the admin around it: sorting receipts, drafting payment reminders, explaining a report in plain English. It shouldn’t be the thing that produces your figures from memory. Use proper accounts software for the numbers, and a person, or your bookkeeper, for the judgement.
And keep money itself well away from it. There’s more on that in Why your AI should never issue a refund.
A quick test you can run
Give your AI tool a realistic sum from your own business, such as a quote with a discount and delivery, and check its answer against your spreadsheet. Then change one number and ask again, and ask the same question twice. If it ever gets one wrong, or gives two different answers to the same question, you have your reason to keep the numbers in the tools.
Want to see which jobs AI could take off your hands safely? Grafto’s AI Tools Report works it out from a short questionnaire.
Frequently asked questions
Why is AI bad at maths?
A chatbot predicts what text should come next, so it produces answers that look right rather than working them out. Simple sums are usually fine; long ones with several steps or rounding can go wrong with no change in its confident tone.
Can AI do calculations I can rely on?
Only when it hands the sum to a calculator, a spreadsheet or code, and uses figures looked up from your own records. Ask it to show where each number came from, and check anything a customer will see.
Can AI do accounting for me?
It can help with the admin, such as sorting receipts and drafting reminders, but your figures should come from accounts software and the judgement from a person. Don’t let it produce numbers from memory.
Questions, or want a tool we don't have yet? Email hello@grafto.co.uk — a real person replies.