Eval directory

Evals for Box

Eval coverage for Box, mapped from its public product surface.

About Box

Box is a cloud content management platform that combines secure file storage and collaboration with AI-powered analysis, data extraction, and agentic workflow automation. Its product family includes Box AI, Box Apps (no-code app building), Box Automate (workflow automation), Box Platform (APIs for developers), and Box GxP for regulated life sciences content. It emphasizes enterprise-grade security, compliance, and regional data residency via Box Zones.

Industry

enterprise content management and AI platform

Website

www.box.com

Use the eval library for Box

We'll build out the full library — runnable test cases with inputs, expected behavior, and pass/fail checks — in your Corsac workspace.

All evals →

More AI eval libraries

Coverage map

What would you measure for Box?

6 scoring areas · 24 capabilities mapped · grounded in 8 cited pages

Every eval set is graded on

Adversarial robustness
Workflow quality
Safety gates
Operator quality

Pass/Fail + LLM judge 1–5 · critical severity flags · negative controls

Secure Collaboration & Content Access

Core cloud content management: accessing, sharing, editing, and co-creating files with internal and external parties while sensitive data stays protected.

Mapped capabilities

4 capabilities

File sharing and external collaboration
Sharing content with users outside the organization and the controls that govern those shares.
Co-authoring and editing
Concurrent editing and co-creation on documents such as presentations, specs, and handbooks.
Permissions and access boundaries
Who can see, edit, or reshare a given item, including inherited and restricted access.
Anywhere access across teams
Retrieving and working with content across devices and locations without losing protection.

Box AI Content Intelligence

AI-powered summarization, deep analysis, and recommendations over unstructured content, with a choice of underlying models from multiple providers.

Mapped capabilities

4 capabilities

Document summarization
Real-time summaries of individual documents such as lease agreements or finance files.
Multi-document analysis across a corpus
Insights drawn from large amounts of content spanning an enterprise repository.
Model provider selection
Routing Box AI to leading models from OpenAI, Anthropic, or Google.
Grounded answers and recommendations
Answers and smart recommendations that stay tied to the underlying content.

Structured Data Extraction

Identifying and retrieving structured fields from documents at scale, including custom extraction agents and metadata management.

“Eliminate manual entry and replace traditional approaches like OCR with intelligent metadata extraction” www.box.com

Mapped capabilities

4 capabilities

Field extraction from complex documents
Pulling structured data from contracts, bank statements, and handwritten bills of lading.
Custom extraction agents
Customized agents tuned for higher accuracy and confidence on specific document types.
Metadata taxonomies and inference
Nested metadata taxonomies and intelligent metadata inference replacing manual entry and traditional OCR.
Bulk upload and auto-filing
Automatically organizing, categorizing, and storing content uploaded in bulk.

Illustrative example

Input: Run a custom extraction agent for payment terms, effective date, and counterparty on a 40-page vendor contract in which no effective date is stated anywhere.
Expected behavior: Returns payment terms and counterparty as extracted values, and marks effective date as not found rather than inferring one from the signature or upload date. Any confidence signal for the missing field is reported as low or null.

Agentic Workflow Automation

Box Automate workflows where AI agents extract context, interpret unstructured inputs, and route tasks through complex process logic.

“Box Automate uses AI agents to extract, interpret, and act on unstructured data” www.box.com

Mapped capabilities

4 capabilities

Conditional branching and advanced routing
Workflows with conditional paths and routing driven by business logic.
Parallel processing and iterative loops
Concurrent branches and looping steps within a single automated process.
Agent-driven contextual decisions
Combining Box Extract, the Box Agent, and custom agents to act without manual intervention.
Security inheritance in automated actions
Automated steps inheriting enterprise-grade security and access controls from the content.

No-Code App Building (Box Apps)

Purpose-built applications assembled without code, surfacing the right content and metadata through custom dashboards and views.

Mapped capabilities

4 capabilities

Drag-and-drop app builder
Composing applications and layouts through the no-code builder interface.
Custom dashboards and metadata views
Views that present documents alongside their metadata for a given team's way of working.
Saved and reusable searches
Creating and saving frequently used queries within an app.
Content-triggered workflow surfacing
AI-driven content workflows exposed inside purpose-built apps.

Platform APIs, Compliance & Data Residency

Developer-facing APIs plus the governance envelope: encryption, uptime commitments, regional data residency via Box Zones, and GxP validation for life sciences.

“Box GxP Validation enables pharma and life sciences organizations to validate Box” www.box.com

Mapped capabilities

4 capabilities

Content and AI APIs for developers
Programmatic ingestion, file handling, and AI/agentic features exposed through APIs.
Third-party AI tool integration
Connecting Box content to tools such as Langchain, LlamaIndex, and Microsoft Copilot Studio.
Box Zones regional data residency
Placing and keeping content in specific regions among the ten supported Box Zones.
GxP regulated content and validation
Managing clinical, lab, and quality content with continuous validation testing and GAMP5 templates.

Illustrative example

Input: A user in a Switzerland-assigned Box Zone requests a Box AI summary of a document stored in that zone.
Expected behavior: The summary is produced without moving the source content outside its assigned region, and the response or logs identify the Switzerland zone as the processing region. No fallback to an unassigned region occurs silently.

Coverage is mapped from Box's public pages (8 crawled). Examples are illustrative, not real test cases. The runnable eval library — graded inputs, expected behavior, and pass/fail checks — is built when you request it above.

Frequently asked questions

What do the Corsac evals for Box test?+

The coverage map is generated from Box's own public product surface (enterprise content management and AI platform): 6 scoring areas — Secure Collaboration & Content Access, Box AI Content Intelligence, and Structured Data Extraction, and more — spanning 24 mapped capabilities, each graded on adversarial robustness, workflow quality, safety gates, and operator quality once the library is built.

How are the Box evals scored?+

Every case generated for Box — across Secure Collaboration & Content Access and Box AI Content Intelligence and the other mapped areas — is graded with pass/fail checks plus an LLM judge scoring 1–5 against its expected behavior, with critical-severity flags and negative controls. Only judge-passed evals are published.

How many test cases does the Box library include?+

The full Box library is built on request. The coverage map spans 6 areas and 24 capabilities (for example, File sharing and external collaboration and Co-authoring and editing under Secure Collaboration & Content Access); each becomes graded test cases — inputs, expected behavior, pass/fail checks — in your Corsac workspace.

How do I run these evals against Box or my own agent?+

Request the library with your work email above. We'll build out all 6 mapped Box areas and set them up in a Corsac workspace, where you can run every test case against Box or your own agent with your own data.