Forty invoices a month. Each one — open the PDF, copy down the supplier, number, date, amount, VAT, IBAN. Into the spreadsheet. Next.

That’s roughly four hours nobody remembers spending. And it’s one of those jobs AI genuinely does well — not because it’s clever, but because it doesn’t get tired at invoice number thirty-seven.

This lesson isn’t about AI replacing the accountant. It’s about how to stop retyping numbers by hand — and how to get a check along the way that you didn’t have before.

What you need to start

Nothing to install and nothing to pay for.

1. An AI chat you can upload a file to — Claude, ChatGPT or Gemini. The free versions are enough. If you have an AI app on your computer that you can let read one folder — even better: you won’t need to upload files one by one.
2. A few real invoices as PDFs. Take last month’s.
3. The spreadsheet you use now — Excel or Google Sheets.

Important before you start. Invoices contain your suppliers’ data. Before uploading them to AI, read How to start using AI at work (safely) — it explains which tools train models on your data and which don’t. Two minutes of reading that decides whether you can do this at all.

First attempt: one invoice

Don’t automate straight away. First find out whether the AI reads your invoices correctly at all — suppliers’ layouts can be strange.

Upload one PDF and send this prompt:

A supplier invoice is attached to this message.

Read these fields from it and present them as a table:
supplier name | invoice number | issue date | payment due date |
amount excl. VAT | VAT amount | amount incl. VAT | supplier IBAN | payment reference

RULES
- Don't invent anything. If a field isn't on the invoice or you can't read it,
  write "MISSING" — don't guess.
- Copy numbers exactly as written; don't change the format or round.
- Give dates in YYYY-MM-DD format.
- Answer in English.

AT THE END, tell me separately:
1. Which fields you weren't sure about, and why.
2. Whether anything on the invoice is unusual — handwritten additions, unclear
   printing, totals that don't add up, or VAT that doesn't match the amount excl. VAT.

Now check every field with the same invoice in front of you. This is the only time you’ll check everything — but you can’t skip it.

The most common mistakes you’ll see: the issue date and the due date mixed up, an order number read instead of the invoice number, or the last character of the IBAN lost when it wrapped onto the next line.

Found a mistake? Don’t be cross with the AI — fix the prompt. For example: “This supplier’s invoices show the due date at the bottom, under the table, not at the top.” From now on the AI will know.

When one works — the whole month

Now the same for all of the month’s files — and not tips on how to do it, but a finished file. This prompt also compares with the previous month straight away:

You are my invoice entry assistant. Turn supplier invoices into a tidy table
and show what has changed since last month.

SOURCES
- This month's invoices: [a folder, e.g. Documents/Invoices/2026-10,
  or attached to this chat]
- Last month's table: [attached / I don't have one — then skip the comparison]

PROCESS
0. TOOLS. Tell me whether you can read the files in that folder yourself and
   create an Excel or CSV file. If not, ask me to upload the invoices in batches.
1. ASK. If an invoice is unusual or it's unclear where a field goes — ask.
2. PLAN. Tell me how many files you found and whether they are all invoices
   (not reminders, not pro-forma invoices).
3. DO. Read them all.

FIELDS — one row per invoice
supplier | invoice number | issue date | payment due date |
amount excl. VAT | VAT | amount incl. VAT | supplier IBAN | file name

GOAL — a finished file, not advice
- An Excel or CSV file: each invoice a row, a grand total at the bottom.
- A separate list: suppliers whose IBAN changed; new suppliers; amounts that
  changed by more than 30 %.
- The job is finished when the file exists and has as many rows as there are invoices.

RULES
- Don't invent anything. Whatever you can't read — write "MISSING".
- Copy numbers exactly; don't round. Dates YYYY-MM-DD. Make the CSV with
  semicolons — my Excel uses Lithuanian regional settings. [If yours uses
  English settings, change this to commas.]
- Never pay, never change bank details and never write to suppliers. You read
  and flag — a person decides.
- Don't decide whether it's fraud. Just show what has changed.
- Answer in English.

AT THE END
How many invoices, the grand total, the list of changes and separately — WHAT YOU DID NOT CHECK.

The sum of the amounts must match what the bank or your accounting software shows. If it doesn’t, the mistake is in one row, and you’ll find it in a minute. So always ask for a grand total at the bottom of the table: one number that checks all the others.

The same job that catches a scammer

This is the part you won’t find in other lessons.

When you type invoices in by hand, a changed supplier IBAN looks like nothing at all. Just another number. That’s exactly why invoice fraud works — and exactly how an accountant at one Lithuanian company transferred €400,000 in reply to a genuine email with a fake account number.

But when AI reads the data, the comparison costs it nothing. The main prompt above already does it. If you only have last month’s table and a few new invoices, a shorter one is enough:

Attached are last month's invoice table and this month's invoices.

Read this month's invoices as before, then SEPARATELY, clearly marked, list:
- suppliers whose IBAN differs from last month;
- new suppliers who weren't on the list before;
- invoices whose amount changed by more than 30 % since last month;
- invoices that stress urgency ("urgent", "today", "late-payment interest" —
  or the Lithuanian "skubu", "šiandien", "delspinigiai").

Don't decide whether it's fraud. Just show what has changed. Answer in English.

That last line matters. The AI shouldn’t decide — it should notice. The decision is made by a person, by phone, on a number they already knew.

The line the AI never crosses: it never pays, never changes bank details and never writes to suppliers. It reads and flags — a person makes the call.

What not to do

The internet is full of tutorials suggesting you connect AI to your whole mailbox: let it read the emails itself, pick out the invoices itself, enter them into the accounts itself. Technically it works. But the tutorials almost never say what that means.

Permanent access to a mailbox means the AI sees everything — payslips, contracts, personal correspondence. And every email that arrives in that mailbox is text the AI will read. Including a scammer’s email that says “new payment instructions”. An automated system has no gut feeling.

That’s why in this lesson the AI reads a folder you saved the invoices into, not the mailbox. It’s more tedious and safer, and you save almost as much time. Come back to permanent mailbox access only when:

• you know where the AI makes mistakes with your invoices;
• you have a separate mailbox only for invoices (e.g. invoices@company.lt) that nothing else goes into;
• someone in the company is responsible for reviewing what the system did.

Until then it isn’t automation; it’s an unexamined risk.

What to do today

1. Take one invoice from last month and run the first prompt.
2. Check every field. Found a mistake — add the fix to the prompt.
3. Save the corrected prompt somewhere you’ll find it — notes, a Word file, anywhere.
4. Next month, save the invoices into a folder and run the month prompt.

Step three is the real result. Not one invoice read, but a prompt that works with your suppliers.

Check yourself

Four questions on what you’ve just read. At the end you’ll get a certificate with your score.