Reading Data Out of PDFs and Forms Automatically: How It Works

Key Takeaways

• Text recognition makes a document readable, and an AI model puts the details you need into fixed fields.
• Unlike older tools, AI doesn't need a template per form and copes with changing layouts.
• It only becomes reliable with clear field definitions, plausibility checks and a person for uncertain cases.
• Measure quality per field, and relax the checks only once a field has proved itself.
• Start with one common type of document that someone types up by hand today.

Data in PDFs and forms can now be read out automatically: text recognition (OCR) makes the document readable, and an AI model puts the details you need into defined fields. Unlike older tools, this doesn't need a template for every form and copes with changing layouts. What makes it reliable is clear field definitions, automatic checks and a person who reviews the uncertain cases.

Which documents are suitable for automatic extraction?

The best candidates come up often, contain similar details and are typed up by hand today. Typical examples in smaller businesses:

• Registration and application forms, such as damage reports, course sign-ups or forms from prospective tenants.
• Delivery notes, order confirmations and purchase orders from suppliers.
• Contracts, where you copy the parties, term, notice period and amounts into an overview.
• Inspection reports, certificates and data sheets in the trades and construction.
• Scanned letters from authorities and insurers, where the deadline and reference number matter.

Accounting receipts work on the same principle, but your accounting software usually has specialised features for those already, so this guide is about everything else.

How does AI read a document?

In five steps, from the incoming document to the finished record:

1. Capture the document.The PDF, scan or phone photo arrives by email, upload or in a watched folder.
2. Recognise the text.For scanned documents, text recognition turns the image into text, including tables and checkboxes. In PDFs created digitally, the text is already there.
3. Fill the fields.A language model looks for the details you've defined, such as name, date, property and amount, and returns them in a fixed structure. It understands that “date of signature” and “signed on” mean the same thing.
4. Check.Rules test whether the details are plausible. Does the total add up? Does this customer number exist? Is the date in the future?
5. Pass it on or flag it.Anything that passes every check goes into your system. Uncertain cases go to a person, with the doubtful fields marked.

How is this different from traditional text recognition?

Traditional tools read fields at fixed positions and need a template for each form. AI recognises the details by what they say. That helps when a supplier changes its layout, when documents arrive in German, French and Italian, or when the same detail has a different label every time. The downside is that a language model can fill a gap with a plausible but wrong value. Why AI makes things up explains why that happens. For a single form that never changes, a rule-based approach is often enough, as AI vs. rule-based automation shows.

What if different documents land in the same inbox?

Then there's one more step before extraction: sorting. The system first works out whether it's looking at a damage report, a delivery note or a notice of termination, and then applies the right field definition. AI handles this well too, as long as each type of document is clearly described. If one scan contains several documents, it splits them first. Anything that fits no category needs a fixed route to a person, so nothing gets stuck.

How do you make sure the data is right?

With several checks that back each other up:

• Define the fields precisely: format, allowed values and what happens when a detail is missing. An empty field is better than a guessed one.
• Check against your master data: match customer numbers, properties, suppliers and items against your system.
• Apply arithmetic rules: totals, quantity times price, dates in the right order.
• Mind Swiss formats: amounts like 1'250.50, dates like 01.10.2026, documents in several national languages.
• Take samples: at the start, regularly check cases that pass every test as well.

Measure quality per field, not per document. A name that's almost always right and an amount that's wrong one time in ten need different amounts of checking. Relax the checks field by field once a field has proved itself over a while.

Where are the usual stumbling blocks?

Rarely in the reading itself, mostly in documents that are messier in everyday work than in the test:

• Tables running over several pages, where rows come apart at the page break.
• Stamps, signatures and handwritten notes on top of the printed text.
• Checkboxes ticked crookedly, half-ticked or ticked twice.
• Photos with shadows, taken at an angle or out of focus.
• Old versions of a form, where a field has a different name or is missing altogether.

None of this is a reason not to start. But these cases belong in your test documents, so you know before you go live how the system deals with them.

What about data protection?

Forms almost always contain personal data, and sometimes sensitive data, such as health details in an accident or damage report (Art. 5 FADP). Find out where the processing happens, whether the provider uses your documents for training and how long copies stay with them. What to look for when you choose a provider is covered in AI tools and where your data goes.

How do you start a first project?

Small, with a single type of document and full checks at the start:

1. Pick one type of document.Choose one that comes up often and is typed up by hand today. One type is enough to begin with.
2. Set the fields and the destination.Decide which details you need and where they go: into your industry software, your ERP or, for now, a spreadsheet.
3. Collect real examples.Gather a few dozen documents from everyday work, including poor scans, handwriting and odd cases. With only clean examples, every test looks better than real life.
4. Measure your baseline.Note how long one document takes today and how often typing errors creep in. Measuring time saved by AI shows a simple way to do it.
5. Run a pilot with full checks.At first, a person checks every result until you know which fields are reliable. Then you reduce the checks step by step.

An example: a construction company receives delivery notes as PDFs, as scans and as phone photos from the site. Until now, someone in the office has typed up the supplier, date, site, items and quantities to match them against orders. In the pilot, the system reads out these fields, matches them against open orders and only puts the differences in front of a person. The person in the office can then focus on what really matters: a missing delivery, a wrong quantity or an item that was never ordered.

The data entry page shows what this looks like in different industries. How the extracted data gets into your systems without anyone retyping it is what our automation and integration service is for.