How we extract receipts with AI without sending them to the cloud

Our bookkeeping service receives earnings statements, fuel receipts, and phone bills every month, as PDFs and as phone photos. Typing every number in by hand was slow and error-prone. Sending the documents to a cloud AI service would have been fast, but these are financial records of real people. So we built a pipeline that keeps them on our own infrastructure.

1. Read digital PDFs directly

Many documents, such as platform earnings statements, are digital PDFs with a real text layer. For those, no AI is needed at all: a PDF parser reads the text exactly, instantly, and for free. This is the fastest and most accurate path, so it always goes first.

2. Use OCR for photos and scans

A photo of a crumpled receipt has no text layer. Here the pipeline falls back to optical character recognition (OCR), which turns the image into text. OCR is not perfect, which is exactly why the next steps exist.

3. Rules first, AI second

Each document type has its own extractor. A fuel receipt, a phone bill, and an earnings statement all look different, and known formats can be parsed with simple, predictable rules. Only when the rules are not confident does the text go to a language model, which runs locally on our own hardware and returns structured fields: vendor, date, total, VAT, and category.

4. A human confirms

The result is never saved automatically. The bookkeeper sees a pre-filled form next to the original document, corrects anything that looks wrong, and saves. The AI removes the typing; a person stays responsible for the numbers.

Why local instead of a cloud AI API?

• Privacy: the documents never leave infrastructure we control, which makes GDPR compliance much simpler.
• Predictable cost: no per-document API fees that grow with volume.
• No training on your data: nothing is sent to a third party that could reuse it.

The trade-off is that local models are smaller than the largest cloud models. That is why the pipeline leans on exact parsing and rules first, and why a human reviews every result.

Where else this works

The same pattern fits any business that re-types information from documents: purchase invoices, delivery notes, order forms, contracts, or application forms. If your team copies numbers from PDFs into another system, it can very likely be automated, without handing your documents to anyone else.