Where AI gets admin work wrong

An invoice is a bill from a supplier. AI can type 200 invoice rows in seconds. It can also type a 1 as a 7 and an August date as September.

Extraction tools (AI that pulls fields out of documents) work from images and PDFs. Small print, low contrast and handwriting all cause trouble, and the model (the AI behind the tool) often fills gaps with what looks right instead of stopping.

The failures you will see most:

• Misread numbers and digits. 1 becomes 7, 0 becomes 8, 5 becomes 6. Amounts are where it hurts most.
• Date format swaps. 08/09 can be 8 September or 9 August. The model guesses the format and can flip day and month without telling you.
• Wrong column. An amount lands in the tax field, a supplier name lands in the notes field.
• Totals that do not add up. The model types a total line from the page instead of adding the rows above it.
• Duplicates. The same invoice appears twice when a page or a header repeats.
• Names spelled wrong. Priya Raghavan becomes Priya Ragavan. Small changes still break a search or a match.
• Made-up values. Where the field is blurry or missing, the model writes something that looks reasonable.
• Privacy problems. Real IDs and bank details pasted into a public tool may be stored or reviewed by people outside your company.

Ask the tool to flag, not guess

A prompt is the message you send to the AI tool. This one makes two of these failures easier to see:

Pull out the invoice number, date, supplier name and amount from the
invoice text below. The dates are DD/MM/YYYY (day first), so copy them
exactly as written. If a value is blurry or missing, write UNREADABLE.
Do not guess.

It won't stop every misread, so you still check the output.

Try it