Databricks Document Intelligence's real-time REST APIs became generally available on October 14, 2026, according to the company's official blog: the parsing, classification, and extraction APIs are out of preview, while the decisions API for document routing launched in beta. It sounds like a status change, but it alters the timing of how enterprises handle contracts, invoices, and claims — a file can now be understood and enter a workflow the moment it arrives, instead of waiting for an overnight batch run.
Four APIs, each with one job
The GA capabilities are split into four separately callable functions. ai_parse_document reads a file into structured content, converting tables and figures into text and calculating citation bounding boxes. ai_classify labels a document against a company's own taxonomy, with optional rationales and confidence scores. ai_extract pulls specified fields into a structured result, again with optional citations and confidence scores so a human or an agent can check values against the source. The still-beta ai_decide uses the document's content to choose what happens next, such as routing it to a workflow or flagging it for review. Applications and agents can call these synchronously when a document arrives.
The contract review example Databricks chose
The worked example is a contract intake flow: a lawyer uploads an agreement, the system parses it, identifies whether it is an NDA, an MSA, or a statement of work, then prefills the review form with the parties, effective date, renewal terms, and notice period. Each value carries a citation back to the source text; the lawyer corrects or approves, and the reviewed record is written to Delta. Insurance claim routing and invoice entry follow the same pattern. The point is less about extraction accuracy ticking up than about moving a step that used to live in batch pipelines into the interface where work happens, as a real-time, human-in-the-loop moment.
Where the GA boundary actually sits
The boundary matters: parsing, classification, and extraction are GA, while ai_decide remains beta, so fully automated routing of critical documents should keep a human fallback for now. The announcement gives no pricing or accuracy benchmarks either, and extraction quality still depends on how messy a company's own documents are — testing on real contracts and invoices before rollout is not optional. The direction, though, is clear. From putting a new model in front of all 12,000 employees on launch day ({LINK['en']}) to making document processing real-time, Databricks is competing on how fast enterprise data enters business workflows. Teams sitting on piles of unstructured documents can pilot one low-risk flow first, then decide how far to migrate.
