Box posts Extract API deep dive

Box is pitching its Extract API as a way to turn files stored in Box into structured data that outside AI agents can actually use. (blog.box.com ) (developer.box.com) The company’s April 15, 2026 blog post and a scheduled April 30 technical webinar walk through how developers can ingest documents, extract fields, and expose that data to third-party agents through Box’s Model Context Protocol, or MCP, server. (blog.box.com) (training.box.com) At the center is a simple idea: most enterprise information sits inside contracts, invoices, claims forms, and policy documents that are readable by humans but hard for software to query. Box says Extract converts that unstructured content into metadata saved back into Box. (blog.box.com 1) (blog.box.com 2) Box’s developer docs split the product into two endpoints. Structured extraction returns predefined key-value fields from a file, while freeform extraction lets developers ask for output in natural language or formats like JSON or XML. (developer.box.com 1) (developer.box.com 2) Structured extraction is the more enterprise-shaped path because it can use metadata templates and optical character recognition, the text-reading layer that handles scans and images. Box says the freeform endpoint does not support optical character recognition and is less suited to image-heavy or multilingual document sets. (developer.box.com 1) (developer.box.com 2) The company also draws a line between a standard and enhanced extract agent. In Box’s launch materials, the standard agent is aimed at files under 50 pages and fewer than 20 fields, while the enhanced version is meant for longer and more complex documents. (blog.box.com) (developer.box.com) The security pitch is that the extraction happens inside Box’s existing control plane rather than by copying whole document collections into a separate ad hoc pipeline. Box’s architecture write-up says the system checks file permissions before processing and keeps extracted data tied to the original file as metadata. (japan.box.com) (blog.box.com) That matters for companies trying to let tools like ChatGPT, Claude, or custom agents work with internal records without flattening access rules. Box’s webinar description says the goal is to make content available to third-party systems and agents with MCP after extraction and processing. (training.box.com) (blog.box.com) Box’s January 15 general-availability announcement framed the product around workflows like loan documents, public-sector forms, insurance claims, and legal contracts. Customer quotes in that release came from Valmark Financial Group and the Texas Department of Motor Vehicles. (blog.box.com) The thread running through Box’s recent AI product push is that retrieval alone is not enough. The company is arguing that agents need governed, searchable fields attached to files, not just a chatbot that can read a PDF once and forget it. (japan.box.com) (developer.box.com) So the new deep dive is less a fresh launch than a map of how Box wants enterprise AI to be wired: files stay in Box, permissions stay in place, and agents get structured answers instead of raw document dumps. (blog.box.com) (training.box.com)