Invoices, applications, contracts, claims: most of it arrives as PDF. The information is in there, but a system cannot use it until someone has typed it in. PDF data extraction does that typing automatically.
What is PDF data extraction?
PDF data extraction means reading specific values from a PDF and handing them over as structured data: invoice number, amount, date, customer number, IBAN. Each value gets a name, so business software can process it directly.
Copying the text out of a PDF is not enough. The result is a block of text in which nobody knows which number is the amount. Extraction recognises what each value means. More on this in our blog post on structured data.
Not every PDF is the same
| Digital PDF | Scanned PDF | Photo or fax | |
|---|---|---|---|
| Created by | Software, e.g. an invoicing system | A scanner | A smartphone or fax machine |
| Text | Stored as text | Only an image | Only an image, often skewed or blurred |
| First step | Read the text layer | Text recognition | Text recognition with image correction |
In practice, all three arrive in the same inbox. Good extraction handles them the same way, so the process behind it does not need to know where a document came from.
How IDA extracts data from PDFs
IDA works in three steps:
- Read: IDA recognises machine print with an accuracy of more than 99 % and handwriting with more than 95 %, tables and ticked boxes included.
- Sort: IDA recognises the document type, so it knows which values to look for. More in our blog post on PDF classification.
- Extract: IDA reads the values your process needs and gives each one a name. You define them with a few marked examples per document type. More in our blog post on entity extraction.
Deterministic: recognition, not guessing. The data goes to your ERP, DMS or specialist application as JSON, if needed together with a searchable PDF for the archive.
What makes extraction difficult
Poor originals
Faint scans, faxes and skewed photos lead to reading errors. If a digit is wrong at the start, it is wrong in every system afterwards. That is why reliable text recognition comes first.
Many layouts
Every supplier designs invoices differently. Extraction must not depend on a fixed position on the page.
Tables over several pages
Line items, account statements and lists often run across page breaks. They have to be read as one table.
Handwriting
Many forms are filled in by hand. Without handwriting recognition, these values stay manual work.
Uncertain values
No reading is perfect. IDA gives every value a confidence score. Below the threshold you set, a person checks it, and only that one value.
Where PDF data extraction is used
Invoices
Invoice number, date, amounts and supplier go straight to accounting.
Identity checks
Banks and insurers read name, date of birth and address from ID documents and application forms.
Claims
Insurers take the policy number, date of loss and amount from claim reports and route the case.
Healthcare
Patient data and findings are read from reports and letters and filed in the right record.
Frequently asked questions
Can I not just copy the text out of the PDF?
Only from digital PDFs, and even then you get unsorted text. Extraction delivers named values that a system can use directly.
Does it work with scanned PDFs?
Yes. IDA first recognises the text on the image, then extracts the values.
How many examples does IDA need?
A few marked examples per document type are enough.
In what format do I get the data?
As JSON via an interface, and if needed as a searchable PDF with the data as metadata. More in our blog post on PDF and JSON.
See it with your documents
In a demo, we show you how IDA reads the values your process needs from your PDFs.