OCR and scanned documents: what works in 2026
How well text and data can now be extracted from scans, photos and PDFs, where it still fails, and how to build a document capture flow a business can rely on.
The short answer
Optical character recognition, OCR, the reading of text from images, has moved from a technology that needed clean scans and careful setup to one that reads most business documents well. Printed text on a clean, straight, well-lit scan is a solved problem. Phone photos of crumpled receipts, faded thermal paper, handwriting, dense tables, stamps and annotations are much better handled than a few years ago and still imperfect. The larger change is that modern document models go beyond characters: they understand layout, find the fields a process needs, the total, the date, the supplier, the line items, and return them as structured data with confidence scores. Reliability in a business comes from the flow around the model as much as the model itself: checking image quality before extraction, validating results against known values, routing low-confidence fields to a quick review, and keeping an audit trail from every value to the original image. And the cheapest improvement available is scan quality, because most failures begin with a bad picture.
What works and what needs care
| Document | Extraction today | What to add |
|---|---|---|
| Printed invoices and statements, clean scan or digital PDF | Reliable, with structured fields | Validation against supplier list and expected totals |
| Phone photos of receipts | Good in decent light; weak on faded, crumpled or angled | Quality check with retake prompt; review for low confidence |
| Forms with printed fields and typed entries | Reliable | Field mapping once per form type |
| Forms with handwriting | Usable for block capitals and numbers; weaker on cursive | Review step; constrain fields where possible |
| Tables and line items | Good on clear tables; errors on merged cells and wrapped text | Row-count and total checks |
| Contracts and long documents | Text reliable; clause identification good with a model | Human review of anything acted on |
| Identity documents | Specialised models do well; regulatory constraints apply | Purpose-built verification services; strict data handling |
| Stamps, annotations, signatures | Detection improving; interpretation uncertain | Flag for a person |
| Multi-page scans in the wrong order or orientation | Handled by pre-processing | Automatic rotation and page ordering |
Building a reliable capture flow
- Capture well: guide the user to a straight, well-lit photo or a proper scan; check quality automatically and ask for a retake.
- Pre-process: rotate, deskew, crop, order pages.
- Extract structured fields with a document model appropriate to the document type.
- Validate: formats, ranges, arithmetic, known suppliers and references.
- Threshold: high-confidence, valid fields flow; the rest go to a review screen with the image beside the fields.
- Record: every value linked to the image and the confidence; every correction logged.
- Monitor: extraction rate, review rate, correction rate per document type and per source.
- Improve: the corrections show which fields and sources need better capture or mapping.
Where the value is
Receipts and expenses captured at the moment of purchase instead of at month-end. Supplier invoices read and matched without retyping. Forms from customers and the field turned into records the same day. Archives of paper made searchable. In each case the extraction is the engine and the flow is the vehicle, and businesses that build the vehicle get years of value from an engine that keeps improving underneath.
What this means for you
Text extraction from documents is reliable enough to build on, and document models now return the structured fields a process needs. Build the flow around it: good capture with quality checks, validation, per-field confidence with a review step, an audit trail to the image, and monitoring per source. Do that, and the paper and photos that used to be retyped become data the same day, with the few hard cases handled by a person in seconds.
Frequently asked questions
Is OCR accurate enough to trust now?
For printed text on a clean, straight, well-lit scan, yes, at accuracy levels where errors are rare and caught by validation. For phone photos taken in poor light at an angle, faded thermal receipts, handwriting, dense tables and documents with stamps or annotations, the results are usable but need a review step. The question is less whether OCR is accurate than whether the flow around it catches the cases where it is not.
What is the difference between OCR and document AI?
Classic OCR turns pixels into characters and leaves you with a block of text. Document models understand the document: they recognise that this number is the total, that block is the supplier address, these rows are line items, that date is the due date, and they return structured fields with confidence scores. For business processes, invoices, receipts, forms, contracts, the structured output is what matters, and it is what has improved most.
How do we handle the documents it cannot read?
By design rather than surprise: an image quality check that asks the user to retake a blurry photo; confidence scores per field so uncertain values go to a review queue; validation against known values such as supplier lists and expected totals; and a person who confirms or corrects in a purpose-built screen with the image alongside. The unreadable few are handled quickly, and the readable many flow through.