Databricks releases real-time document APIs for apps and AI agents

A newly uploaded contract or invoice can become input for an application’s next step rather than wait for a batch-processing job. Databricks on October 14, 2026, made its Document Intelligence REST APIs for parsing, classification and extraction generally available. The interfaces are designed for real-time applications and agents; a separate document-routing API launched in beta.

Processing at upload, not in a later batch

Databricks frames the release around a timing problem. Traditional document workflows use asynchronous batch pipelines, delaying information needed for immediate decisions or interactive review. Its synchronous APIs are designed to run when a file arrives, letting an application identify the document, extract its contents and put the result before a person or agent.

The company’s examples span legal, finance and insurance work. A contract needs parties and renewal terms entered into a tracking system. An invoice needs vendor, purchase-order and total information for approval. An insurance claim’s attachments need reading before routing. Databricks identifies that initial reading and data-entry step as the shared bottleneck.

The interfaces divide that work into distinct functions. Parsing makes the file’s contents usable; classification labels what arrived; extraction returns requested fields. Decision-making is a separate function that can select a next step. Databricks describes the four tools this way:

• ai_parse_document reads the file and turns content, including tables and figures, into structured text. It also calculates citation bounding boxes—the locations of source material within the document—for downstream processing.
• ai_classify assigns labels to documents or extracted entities using a business’s own categories. Optional rationales and confidence scores accompany those labels.
• ai_extract returns the fields a developer specifies in a structured result. Optional citations and confidence scores let humans or agents check the extracted values against the source.
• ai_decide uses document content to choose what happens next. That can mean sending the document to a particular workflow or flagging it for review.

Different availability for processing and routing

01Generally available

Parsing, classification and extraction

Databricks released the REST APIs for these three document-processing functions into general availability.

02Beta

Document decisions and routing

The ai_decide REST API launched in beta, rather than sharing the other interfaces’ general-availability status.

A prefilled contract still has a reviewer

Databricks’ contract-intake example shows how these calls fit together inside an application. A lawyer uploads an agreement to a workflow built with Databricks Apps. Parsing produces structured content the app can save for other uses. Classification then identifies a nondisclosure agreement, master services agreement or statement of work, determining which review form the app opens.

Extraction fills in the parties, effective date, renewal terms and notice period. Citations let the lawyer compare each value with the agreement. The lawyer then corrects or approves the fields before the app writes the reviewed record and source reference to Delta. In this example, automation prepares the review rather than replacing the person who approves the record.

The same interfaces can feed an agent

A review form is not the only intended destination. Databricks says a custom agent built on Agent Bricks can call the APIs when it needs to understand or analyze incoming documents before making its next decision. That route places document processing inside the agent’s workflow, instead of using the results solely to populate a human-facing application.

For developers getting started, Databricks points to Genie Code in their workspace, its REST API reference and guides for parsing, extraction and classification. Those three processing functions can also be tried interactively from the workspace’s Agents page through the Agent Bricks interface.