Check an invoice for required fields with ChatGPT

Checking a source document before it goes into the books is repetitive and easy to rush: is the date there, does the amount in words match the figures, is there a signature and a title. ChatGPT reads a photo of the document and runs your checklist against it in about ten minutes, returning a table with a quote for every field. It does not replace the look at the original; it tells you exactly where to look.

Start from your own checklist

The model does not know your accounting policy. It knows what invoices usually contain, and that is not the same list. Before the first check, write the required fields for each document type you handle: delivery note, service acceptance certificate, invoice, waybill. Take them from your accounting policy. Keep the list in a note or a Project so you paste the same one every time.

Cover the parties' details before photographing if you work in a public service: bank details, tax numbers, signatures. The check is about presence and consistency of fields, and the model can confirm "a tax number is present" from a covered field as well as from a visible one.

The prompt

Role: you are an accountant accepting source documents into the books. Context: the image shows a {{type: delivery note, service acceptance certificate, invoice, waybill}}. Required fields under our accounting policy: {{list of fields}}. Task: check each field from the list. For each, state whether it is present and quote what exactly is written. Separately flag doubtful spots: illegible text, a mismatch between the amount in figures and in words, a date later than today, a missing signature or title. Format: a table field | present or missing | what is stated | remarks. At the end, a conclusion: accept into the books or return, and a list of fixes. If you cannot read a fragment, write "illegible" instead of guessing.

The "quote what exactly is written" instruction is the safeguard. A table that says "date: present" is unverifiable. A table that quotes the date exactly as printed lets you compare it with the original in a second.

What the model catches and what it misses

It is good at structural gaps: a missing position next to a signature, an amount in words that says something different from the figures, a date in the future, a field left blank. It is weak at digits on poor scans, where similar-looking digits swap. That is why every item the model marks as missing or doubtful gets checked on the paper, and why you spot-check two items it marked as fine.

What to do

1. Write your required-field checklist for this document type from your accounting policy.
2. Photograph or scan the document, covering details if you use a public service.
3. Upload the image to ChatGPT with the checklist and the prompt.
4. Read the table and check every missing or doubtful item on the original.
5. Spot-check two items the model marked as present.
6. Accept the document into the books or return it to the supplier with the list of fixes.

Done when

Every field on the checklist has a mark and a quote, your spot check of two items matched the original, and the document is either accepted or returned with a specific list. A vague return ("please correct the invoice") sends it back to you twice; the fix list from the table makes it one round.

Plans and settings

Photo uploads work on the free plan with small limits. If you check documents every day, Plus at 20 USD per month raises the limits and keeps your checklists in a Project; Business at 25 USD per seat with annual billing keeps team data out of training by default. See ChatGPT pricing. Turn off Improve the model for everyone in Data controls first, and start a new chat for each batch so details from different documents do not mix. The accountant's version is on the document check page for accountants; for long multi-page documents, see ChatGPT vs Claude.