The evolution of OCR: from recognition to structured data
OCR turns text in an image into machine-readable text. This January 2026 explainer looks at how that task relates to extracting fields and tables, and what multimodal language models changed.
Editorial update . Edited for clarity while preserving the announcement’s historical context.
Recognizing text
Conventional OCR pipelines can preprocess images, locate text, recognize characters, and return text with position or layout information. Tools differ: OCR is not limited to an unstructured text dump. Extracting a particular business field often adds rules or another processing stage.
Extracting with a multimodal model
Models such as GPT-4 Vision and Claude brought image input and language instructions into one workflow. They could be asked to identify invoice fields or organize line items. Context can help, but an inferred value is different from text actually visible in the source.
Comparing the approaches
The choice depends on the documents, the required output, and how results will be checked. Neither approach guarantees correct extraction from handwriting, damaged scans, or an unfamiliar layout.
| Aspect | Conventional OCR | Multimodal extraction |
|---|---|---|
| Output | Text and, depending on the tool, layout information | Fields or structured output requested through instructions |
| Preparation | Preprocessing and field-specific parsing may be needed | Instructions, schemas, and output validation may be needed |
| Errors | Recognition and layout errors need checking | Recognition errors and inferred values need checking |
Applications beyond financial documents
The original article described healthcare records, legal documents, logistics, accessibility, archives, and vehicle-related text recognition. These are examples of OCR applications, not a list of ScanToExcel product capabilities.
Why it matters
Reducing manual transcription can leave more time for reviewing and using the data. The benefit depends on the quality of the result and the effort needed to check it; automatic extraction does not remove that review step.