From PDF to Excel with AI: how to automate document extraction and know when you need custom software

The problem: documents someone turns into data by hand

In most companies, someone opens a PDF, reads what it says, and types it somewhere else. Into Excel, an ERP, accounting software, a CRM. Every day. With every invoice, delivery note, form or document that comes in.

It's work that doesn't require judgment — only attention. And that's exactly the kind of work AI can do, or at least significantly reduce.

This guide explains how to do it in practice, which tools to use, the real limits of the manual method, and when it makes sense to move to real automation.

Which processes fit this approach

Not all documents are the same. Before investing time, it's worth checking whether your process fits.

Good fits

• Supplier invoices (PDF, image, scan) → Excel, ERP or accounting software
• Delivery notes → stock records or matching against orders
• Forms received by email → database or CRM
• Contracts or quotes → extracting key terms (deadlines, amounts, dates)
• Client documentation (ID, payslips, supporting documents) → case files or completeness checklists

Poor fits for the manual method

• Documents without legible text (very blurry images, undigitized handwriting)
• Documents where extraction requires case-by-case accounting or legal judgment
• Processes where the AI needs to act in real time, connected to other systems, with no human involved

If your process fits the first list, keep reading.

How to do it manually with AI: the step-by-step method

What you need

• ChatGPT (GPT-4 or later) or Claude — either works
• The original document: attach the PDF directly, or paste the text if the document allows it
• Where the data goes: an Excel or Google Sheets file to paste the result into

You don't need to install anything or have technical knowledge.

Step 1 — Attach the document and use this prompt

Open ChatGPT or Claude, attach the invoice or document, and paste this prompt:

Extract the following fields from this invoice:
- Supplier (name or legal entity)
- Supplier Tax ID (NIF/VAT)
- Issue date
- Invoice number
- Taxable base
- VAT rate (%)
- VAT amount
- Total invoice amount
- Due date (if present)
- Description of the service/product

Return the result as a table with two columns: Field and Value.
Mark as "uncertain" any field that is unclear, missing, or doesn't
add up (for example, if the total doesn't match base + VAT).

The key is in the last instruction: get the AI to flag what doesn't add up, instead of making up a plausible value. A field marked "uncertain" is useful information. A made-up field is a silent error.

Step 2 — Validate what's marked as uncertain

The AI returns a table with the extracted fields. Only review what's marked as uncertain:

• Total doesn't match base + VAT — could be a supplier error, a rounding issue, or a misread field. Check the original before paying.
• Tax ID illegible or incomplete — ask the supplier for a corrected invoice.
• Due date missing — add it yourself based on the agreed payment terms with that supplier.
• Generic description — expand it with the order reference if you have one.

Anything not marked as uncertain doesn't need your attention — just copy it to the destination. The goal is to focus your time only where it's actually needed.

Step 3 — Move it to Excel

Copy the table the AI returns and paste it into your Excel sheet. If you use Google Sheets, you can ask the AI to return the result as CSV to make pasting easier.

For repeat invoices from the same supplier, once the prompt works well for that format, processing each invoice comes down to three actions: attach, copy, paste.

The limits of the manual method

This method works. But it has a clear ceiling worth knowing before you rely on it.

• The time is still yours. Even if each invoice takes 2 minutes instead of 10, at 100 invoices a month you're still spending over 3 hours on work that requires no judgment — only attention.
• There's no cross-checking between documents. If you need to match an invoice against its delivery note, or detect that a Tax ID already billed that number before, you have to do it yourself. The AI extracts from one document at a time — it doesn't connect data across several.
• It doesn't integrate with your tools. The result is a table you copy by hand. It doesn't go straight into the ERP, accounting software or CRM.
• Quality depends on the document. A low-resolution scan or a crooked photo can generate more uncertain fields than usual. It works well with most standard documents — but not all of them.

When it's worth moving to automated software

The manual method makes sense as a starting point: it lets you confirm the process is automatable and quantify the time you're spending. But there comes a point where it stops being the solution — and that point depends on your volume and where the data needs to go.

• More than 50 documents a month processed by hand Below that volume, the savings from the manual method may be enough. Above it, the time spent starts to represent a cost that already justifies real automation.
• The destination is an ERP, CRM or accounting software Direct integration — no copy-pasting — is what actually makes the operational difference: data arrives validated and correctly formatted, with no human intervention.
• You need cross-checking between documents Matching invoices against delivery notes, detecting duplicates, verifying amounts match across different systems — that can't be done document by document. It requires a flow that connects all the data.

If you recognize your situation in any of these three points, the next step isn't optimizing the prompt — it's automating the workflow.

Frequently asked questions

Is it safe to upload invoices to ChatGPT or Claude?

Don't upload documents with third-party personal data without anonymizing them first. To test the method, use dummy documents or check the privacy policy of the tool you're using — this technique is for validating the process, not as a production method for sensitive data at volume.

How many invoices a month before automating is worth it instead of manual AI?

As a rule of thumb: below 50 documents a month, manual AI plus Excel is usually enough. Above that, or if the data needs to reach an ERP, CRM or accounting software, the time spent already represents a cost that justifies real automation.

Does it work equally well with ChatGPT, Claude or Gemini?

Yes. Any AI that lets you attach a document and read its content works for this method. The quality of the result depends more on the document's quality — resolution, orientation, sharpness — than on which specific AI you use.

What if the invoice is scanned or poorly photographed?

The AI may fail or flag more fields as "uncertain" than usual. With low-quality scans, review the result more carefully before trusting it, or re-scan the document if you can.

How much does it actually cost to automate the process, beyond testing it with AI?

It depends on volume and where the data needs to go. A document automation sprint usually starts with a single document type and a specific flow, not a full platform. In the free 20-30 minute assessment we'll tell you what it would cost in your case.

Does this method work for documents other than invoices?

Yes. The same prompt, adapting the fields you ask for, works with delivery notes, contracts, forms, or any document you need to extract structured data from.