Extract data from PDFs with AI instead of retyping it

The job nobody wants

A customer sends an order as a PDF, with their own product codes and their own way of writing quantities. One supplier sends delivery notes in one layout, another in a different one. Someone in the office opens the file, puts it next to the business software and retypes it line by line.

Each document takes minutes, and over time a wrong code or a skipped quantity always slips through.

How it works

I have built this before. The flow goes like this:

1. Documents arrive where they already arrive: a mailbox, a folder, an upload in your business software.
2. The text is pulled out of the PDF. Scans go through character recognition first.
3. The model fills in the fields you need: customer, date, lines, codes, quantities, prices.
4. Code checks the results: totals, codes that exist, documents already entered.
5. The good ones go into the business software or a spreadsheet; the rest land in a review list, with the PDF alongside and the doubtful fields highlighted.

Where a person comes in

The model gets fewer things wrong than people expect and more than anyone would like. It can misread a figure on a blurry scan, or take an oddly written line its own way. So the exact rules (a total, a code, a price off the list) are checked by ordinary code, and a person looks at the uncertain cases.

Usually whoever used to retype everything ends up confirming only the doubtful documents. How many there are, we find out in a trial on your own PDFs, before anything gets connected.

When you don't need AI

If documents always come from the same program, with the same layout, ordinary code reads them. If the supplier also sends a CSV or has an API, we use that. AI pays off when the documents are written by different people, each in their own way.

To connect the results to other programs, see Integrations and automation.

Frequently asked questions

How accurate is it?

On PDFs produced by software, usually very; on crooked scans or handwriting, less so. That's why there are checks: if the lines don't add up to the total, or a product code doesn't exist in your system, the document goes to a person instead of going through on its own.

Where do our documents go?

The document text goes to the model's provider, Claude or GPT for example, through their APIs: with the main providers, data sent that way isn't used to train their models. If documents can't leave the company, a model can run on your own server, tested on your cases first.

Do the PDFs have to look the same every time?

No, and that's the reason to use AI. For documents that always have the same layout, like a price list exported from a program, ordinary code is often enough: it costs less and always gives the same result.

Does it work with emails too?

Yes. An order written in the body of an email is read the same way as a PDF, and the mailbox can be checked automatically at intervals.

Read more

• The service: Integrations and automation
• More pages like this: Connections, automation and AI