SoftTech AI Registry
Box · Box Extract

Custom Extract Agents

User-configured extraction agents that read content stored in Box, extract values defined by a metadata template, and write those values as metadata on the file, including automatic processing of files in attached source folders.

Recorded characteristics

Function
Box documents Custom Extract Agents as configurable agents that extract structured data from content at scale and automatically apply that data as metadata to files in Box. A user creates the agent in the Extract tab, selects one metadata template, enables the fields to extract, optionally adds per-field AI Instructions (prompt limit 5,000 characters), chooses the Standard or Enhanced Extract Agent, and sets an extraction policy of Keep existing metadata values or Override metadata values. After saving, the agent is activated and attached to up to ten Box source folders, each set to single-level or recursive traversal. Box states that when the agent is active and a source folder is attached, Box Extract automatically extracts data from uploaded files and applies that data as metadata, and that backfill extraction processes files that already exist in a folder when it is attached, with no file limit.
Data access
The agent reads the content of files within the Box folders attached to it, at single-level or recursive depth, and up to ten sample files selected during agent creation. Box lists supported input formats for Box Extract including images (bmp, cr2, crw, dcm, dicm, dicom, dng, gif, heic, jpeg, jpg, nef, png, raf, raw, svg, tga, tif, tiff, webp), design/CAD (ai, dwg, eps, ps, psd, xbd, xdw), documents (boxcanvas, boxnote, doc, docx, gdoc, msg, odt, pages, pdf, rtf, webdoc, wpd), presentations (gslide, gslides, key, odp, ppt, pptx), spreadsheets (csv, gsheet, numbers, ods, tsv, xls, xlsm, xlsx) and a range of code/text formats; the Box Extract limitations section separately lists a narrower set of image, document, presentation and spreadsheet formats and states extraction is supported in English, Japanese, Korean, Chinese and Cyrillic. Scope is bounded by the attached source folders and by the agent owner's access to them; there is no documented access to content outside the attached folders and selected sample files.
Actions
Can take actions
External actions
Unknown
Human confirmation
Conditional
Permission basis
User permissions
Administrative control
Enterprise admins must enable Box AI before Box Extract can be used, and then define access in the Admin Console by selecting Disable for all managed users, Enable for all managed users (documented as the default and recommended setting), Enable for select users and groups, or Enable for everyone except select users and groups, entering up to 100 names or email addresses and up to 100 groups; only groups whose Permission Setting is Admins Only can be selected. Users not enabled for Box Extract do not see the Extract tab. Metadata templates and the Taxonomy field type are created and managed by Admins or Co-Admins in the Admin Console; Box AI template generation and manual template creation are available to Admins and Co-Admins only, and managed users can use or replace an existing template but cannot create, generate or edit templates. Agent owners can Configure, Rename, Deactivate (which stops extractions) or Delete an agent, and manage source folders from the Extraction Sources tab. Only Account Admins have visibility into AI Unit consumption, through the Admin Console AI Units card and a downloadable historical AI units consumption report.
Default state
Disabled
Availability
Box states that Box Extract Agent APIs are available on Business plans and above, while creating and configuring Custom Extract Agents, and autofilling metadata templates using the Standard and Enhanced Extract Agents, are only available on Enterprise Advanced plans. Box's AI usage documentation lists Metadata Extraction (Standard and Enhanced Extract Agents) among the features unlocked at the Enterprise Advanced tier, which includes 20,000 AI Units per month.
Licensing
Box documents that extraction at scale consumes AI Units, which are allocated per entitlement period, do not roll over, and are aggregated across chargeable AI features; Standard and Enhanced Extract Agents and Box AI metadata template generation are listed as chargeable. Documented throughput limits are 500 extractions per minute per user and 700 per minute per enterprise (2.5 million pages a day), with a limit of 100 Extract agents per user, ten source folders per agent and no limit on the number of active Custom Extract Agents per enterprise.
External model or provider
Box does not name a specific model or provider for Custom Extract Agents. Box's developer documentation lists the AI models supported across Box AI, including models from OpenAI, Google (Gemini) and Anthropic (hosted on Amazon Web Services), categorised as core or customer-enabled and as standard, premium or ultra capability tiers. Box states that generating a metadata template with Box AI uses premium models only, and that the choice of Standard or Enhanced Extract Agent applies only to data extraction and not to AI-generated template creation.
Limitations and uncertainty
External action capability is recorded as unknown: Box documents the agent writing metadata inside Box and does not describe the agent itself changing state in an external system. Human confirmation is recorded as conditional because Box documents two distinct modes: inside the agent builder, Box states extraction runs only when the user selects Test Extraction or Run Extraction and is not triggered automatically by uploading a file to an agent or replacing a template, while for an active agent with attached source folders Box states data is automatically extracted from uploaded files and applied as metadata, and backfill processes files already present, with no documented per-file approval. Permission basis is recorded as user_permissions on the basis that an agent can be attached only to folders where the agent owner has Owner/Co-owner, Editor or Viewer Uploader access, that agents are private to their creator and cannot be shared, and that if the owner's folder permission is reduced to Viewer or Previewer the extraction still runs but metadata cannot be updated; Box does not describe a separate service identity. Default state is recorded as disabled because no extraction occurs until an agent is created, configured with a template and at least one field, activated and attached to a source folder, although the admin-level feature setting is documented as enabled for all managed users by default. The reviewed Box documentation does not describe confidence scores, confidence thresholds or bounding boxes for Custom Extract Agents; the documented review mechanisms are Test Extraction against sample files, Run History status logging with failure tooltips, and manual viewing and editing of extracted metadata in Box Preview or Box Apps. Box notes Run History records are stored temporarily and may not be accessible at this time, that files with no metadata found still show Success, and that some extraction processes may still occur after deactivation or deletion. Downstream use of extracted metadata by Box Relay, Box Automate or external systems is not attributed to this capability.

Evidence