AI Made This Worse
The spreadsheet it misread
Written by AI Tools AcademyLast checked on 27 September 2026
Helpful first: Copilot in Excel
This is a fictional example. Fernway Group is a made-up UK company, and everyone in it, including the customers named below, is invented.
AI tools can read a pasted table and do arithmetic on it, and often they get it right. When they get it wrong, the cause is usually the spreadsheet rather than the maths: something a person would notice at a glance and a tool quietly takes at face value.
Before
Alex Doyle, Account Manager for the Midlands, needed two numbers for Tom Elliott's Monday sales meeting: the total value of orders in a short export, and how much of it was June. The export combined orders taken by phone, which Fernway's order system dates day-first (the UK way), and orders from the web shop, whose export dates month-first (the American way). Someone had also typed a TOTAL row at the bottom.
This is what Alex pasted:
Order date,Source,Customer,Value (GBP)
29/05/2026,Phone,Oakmere Lettings,1240.00
02/06/2026,Phone,Carrow Street Studio,860.00
06/05/2026,Web shop,Pellbrook Solicitors,415.00
09/06/2026,Phone,Hazelford Academy Trust,2310.00
16/06/2026,Phone,Linney Accountants,1095.50
06/18/2026,Web shop,Stannard Joinery,289.50
23/06/2026,Phone,Millrace Architects,3150.00
30/06/2026,Phone,Brookvale Veterinary,740.00
07/01/2026,Web shop,Carrow Street Studio,525.00
,,TOTAL,10625.00
The AI approach
Here's our Midlands order export. What's the total order value, and how much of it was in June? [pasted data]
Why this works: This is the version that caused the problem: it assumes the data is clean and asks for answers without asking how they were reached.
What went wrong
Total order value: £21,250.00 across 10 orders.
June 2026: £8,445.00 across 6 orders.
Orders in the export run from 7 January 2026 to 30 June 2026.
Both headline numbers are wrong. Here is what happened, row by row.
The total was doubled. The tool treated the TOTAL line as a tenth order and added its £10,625.00 to the nine real orders, which already came to £10,625.00. That gives £21,250.00, exactly twice the right answer. The "10 orders" was the clue.
One June order was read as May. Pellbrook Solicitors' order came from the web shop, so 06/05/2026 means 5 June. Every other date near the top is day-first, so the tool read it as 6 May and left it out of June. The other two web shop rows came out differently: 06/18/2026 can only mean 18 June (there is no month 18), so it was read correctly, and 07/01/2026 (1 July) was read as 7 January, which kept it out of June by luck.
| What AI said | Correct | |
|---|---|---|
| Total value | £21,250.00 (10 orders) | £10,625.00 (9 orders) |
| June value | £8,445.00 (6 orders) | £8,860.00 (7 orders) |
| Date range | 7 January to 30 June | 29 May to 1 July |
The June difference is £415.00, the Pellbrook order. The correct June figure is the seven orders dated 2 to 30 June: 860.00 + 415.00 + 2,310.00 + 1,095.50 + 289.50 + 3,150.00 + 740.00 = £8,860.00.
The third line of the output was the giveaway. An export Alex knew covered the end of May and June could not start in January. But it sat below the two numbers Alex wanted, and it was easy to skip.
Notice that the tool wasn't careless in a way a person would recognise. It made the same assumption most of us would make about the first few rows, and applied it to the rows where it was wrong. A mixed-format column is ambiguous; any reader, human or AI, has to guess.
A better approach
Fix the data before you ask the question, and make the tool show its working so you can check it.
First, clean the file. Delete any total, subtotal or blank rows. Put dates into one format, ideally by converting the web shop rows in the spreadsheet itself or by asking whoever owns the export to change its date setting. If you can't tidy it, tell the tool exactly what is messy.
Then ask for the working as well as the answer.
Weak
What's the total order value, and how much of it was in June?
Better
This export has one row per order. Ignore any row labelled TOTAL or with no date. Phone orders use DD/MM/YYYY and web shop orders use MM/DD/YYYY. First, tell me how many order rows you counted and the earliest and latest date. Then list every June order with its value, and give the June total and the overall total. Flag any row you weren't sure about.
Why it works: It removes the two traps, explains the date formats, and asks for a row count, a date range and a list you can check in under a minute.
Then check it against something you calculated yourself. A SUM of the value column, excluding the TOTAL row, takes seconds in Excel or Google Sheets. If the AI's overall total doesn't match yours, don't trust its June figure either. For a figure that matters, a formula such as SUMIFS on a clean date column is more dependable than asking a chat tool to do arithmetic, and the AI can help you write the formula if you're unsure how.
Customer names and order values are commercial information. Your organisation's policy and approved tools still take priority, and a spreadsheet tool with AI built in (such as Copilot in Excel on a work account) keeps the data where it already lives.