What is OCR?
A scanned invoice looks fine to a person and is invisible to every system downstream, until something reads the characters out of the pixels.
OCR (optical character recognition)
optical character recognitiontext recognition
OCR is software that detects text inside an image and converts it into machine-readable characters. It takes a scanned page, a photo or a PDF with no selectable text. It outputs a string you can search, copy or feed into another system.
A classical OCR engine works in stages. It finds the regions of an image that look like text, segments them into individual characters, and matches each shape against a trained character set. Tesseract, the open-source engine most products are built on, still works this way. It is fast, it runs offline, and it is cheap at scale.
That pipeline is also why classical OCR breaks in predictable places. Handwriting has no fixed character shapes to match. A table loses its structure the moment text gets read left to right, top to bottom, with no sense of which cell it came from. A skewed photo or a coffee stain drops accuracy fast, because the engine is matching shapes, not understanding the page.
Vision language models changed what happens after detection. A model like GPT-4o or Claude reads the image and reasons about what it means. This column is a quantity. That box is a signature. This handwritten note changes the total. It can answer a question about the page, well beyond transcribing it.
The choice is cost against what the document actually looks like
Every document pipeline starts with the same question: what do your real documents look like, not the sample ones in a sales deck? Clean, printed, single-column pages are a solved problem. Classical OCR reads them at low cost and high speed, and there is no reason to pay a model to do the same job.
The documents that break classical OCR show up every week. A handwritten field on an intake form. A table spanning two pages. A photo taken at an angle in bad light. A vision model reads those correctly more often. It reasons about layout and meaning instead of matching isolated shapes. The trade-off is cost per page and latency, and both multiply fast across a real document volume.
We built PremiumAudit.io on this split. Insurance audit documents arrive as scans, photos and handwritten worksheets, not clean PDFs. The Claude API parses them, validates the figures against each other and drafts a report. A character-matching engine cannot do that kind of reasoning. That system reports 75% faster audit cycles and 95% fewer calculation errors.
In practice, most document pipelines we run are not one or the other. A router sends clean printed pages to a cheap classical engine and routes the rest, tables, handwriting, low-quality scans, to a model. Paying model prices for every page is money spent solving a problem that engine already solved.
- IngestScan, photo or PDF arrives with no selectable text.
- ClassifyRoute by document quality and layout.
- Classical OCRClean, printed, single-column pages.
- Vision modelTables, handwriting, poor scans, mixed layouts.
- ValidateFigures cross-checked against each other.
- Structured outputFields, not a flat block of text.
The routing step is where cost gets controlled. Everything after it is where accuracy gets won or lost.
Common questions
01How accurate is OCR?
Classical OCR reads clean, printed, single-column text at well over 99% character accuracy. Accuracy drops fast on handwriting, low-resolution scans, skewed photos and complex tables, sometimes well below 90%. Vision models hold up better on those harder cases, at a higher cost per page.
02Is OCR still worth using now that vision models can read documents?
Yes, for the documents it was built for. Classical OCR is cheaper and faster than a model call, so it remains the right default for clean, printed, high-volume pages. The shift is routing the harder documents, tables, handwriting, poor scans, to a vision model instead of forcing everything through one engine.
03What is the difference between OCR and document AI?
OCR converts an image of text into characters. Document AI is the layer built on top. It works out which characters form a field, cross-checks values against each other, and produces structured output an application can use. OCR answers 'what does this say'. Document AI answers 'what does this mean'.
Related terms
- Computer vision development →Where document reading fits alongside inspection and detection systems, with a person approving anything the business cannot take back.
- PremiumAudit.io case study →The insurance audit system where document parsing and validation cut audit cycles by 75%.
- Data engineering →Where extracted document fields go next, on their way into a system of record.
- Chunking →How extracted document text gets split before it feeds a retrieval system.
