Getting structured information out of invoices, forms, contracts, reports and scanned records. It is the least glamorous category and probably the largest source of realised business value, because almost every organisation has a mountain of documents and a team of people retyping them.
Technical view
Combines layout analysis, optical character recognition, table structure recognition and extraction into a schema, increasingly done end-to-end by multimodal models reading page images. Requirements for production use are confidence scores per field, validation against expected formats and cross-field arithmetic checks, and a human review path for low-confidence extractions. The silent-plausible-value failure of generative extraction is the main risk to design around.
Related
Part of AI Terminology Index, section: Modalities and Applications.