Getting Structured Data Out of PDFs, Emails and Scans
Last updated:
What it is good at now
Structured and semi-structured business documents — invoices, purchase orders, delivery notes, statements, application forms — extract well. Layouts vary, but the fields are named and the values are in predictable formats.
Handwriting and photographs of documents are workable but noticeably less accurate, and should always route through review rather than straight into a system of record.
Realistic accuracy by document type
| Document | Typical field accuracy | Main difficulty |
|---|---|---|
| Standard supplier invoice | 92–97% | Line items on multi-page documents |
| Purchase order | 93–97% | Non-standard product codes |
| Bank statement | 95%+ | Merchant names and categorisation |
| Handwritten form | 70–90% | Legibility, obviously |
| Photographed document | 80–92% | Angle, lighting, shadows |
| Contract or long-form | Varies | Meaning rather than fields |
Confidence scores are the design centre
A useful extraction system does not just return values, it returns how sure it is about each one. That is what lets you automate the confident majority and review the rest.
Set thresholds per field by consequence. A supplier name being wrong is recoverable; a bank account number being wrong is not. Different fields deserve different confidence bars.
Design the review step properly, because it is the product
The review interface determines whether the system saves time. Done well, a person confirms twenty documents in the time it took to key one. Done badly, review is slower than typing and people abandon it.
- Show the original document beside the extracted fields, with the source region highlighted
- Focus the cursor on the first uncertain field, not the top of the form
- Allow keyboard-only completion — reviewers should never need the mouse
- Learn from corrections so the same supplier is not queried repeatedly
Validate against what you already know
The cheapest accuracy improvement is not a better model, it is checking against your own data. Does the supplier exist? Does the total equal the sum of the lines? Does the purchase order number match an open order? Is the date plausible?
These rules catch errors extraction cannot detect on its own and are quick to implement. They also make the failure modes intelligible, which matters for trust.
Measure the right thing
Field-level accuracy is the vanity metric. What matters operationally is straight-through processing rate: the proportion of documents needing no human touch at all, and time per document overall including review.
A system at 94% field accuracy might have a 60% straight-through rate, because one uncertain field sends the whole document to review. Improving the reviewable design often beats improving the model.
Frequently asked questions
Do we need to train it on our documents?
What about documents in other languages?
How much does a document extraction project cost?
Can it handle documents that arrive as email attachments?
Drowning in PDFs that become spreadsheet rows?
Send us five representative documents, including two awkward ones. We will tell you the realistic accuracy and what the review step would look like.