Announcing the next step in Box's agentic AI framework

Today, Box is taking a major step forward by delivering a new set of agentic AI capabilities designed to give enterprises the confidence and control they need to scale AI. We are announcing the general availability of our new Enhanced Extract Agent to deliver the accuracy that builds confidence, and the beta launch of our remote Box Model Context Protocol (MCP) Server to give you secure control over your entire AI ecosystem. These releases, along with key updates to Box AI Studio and AI Units, and Box AI Admin Console, empower you to operationalize AI with the confidence and control your business demands.

For any enterprise, moving from scattered AI experiments to system-wide deployment hinges on solving this dual challenge. You need confidence that an AI's results are accurate, transparent, and compliant, and you need control over how AI agents securely interact with your proprietary data. 

Let's break down how today's announcements deliver on these fronts: 

With chain-of-thought reasoning, extract data with confidence

Enterprises are fighting a battle for confidence in their own content. Critical data is locked in documents, creating uncertainty across the business. Knowledge workers lose nearly 30 percent of their day searching for or recreating data already on hand, costing companies millions annually in lost productivity. This manual work introduces a crisis of confidence in core operations, leading to costly errors like unintentionally renewing contracts or overpaying suppliers due to misread decimals. While traditional extraction tools have provided valuable automation, they often reach around 80% accuracy which is not sufficient for important use cases and can face challenges with very large or complex documents, sometimes requiring teams to supplement automated results with manual review for complete confidence.

The Box AI Enhanced Extract Agent is designed to close this gap. It is powered by Gemini 2.5 Pro, but raw model power is only half the story. What truly differentiates this service is its use of a "chain-of-thought" approach. This technique prompts the model to provide clear reasoning behind each extracted value, improving both accuracy and transparency.

This "chain-of-thought" reasoning is surfaced directly in the API output, so developers can see not just the extracted value but why it was chosen. For example, a response might include "reasoning": "The document contains signatures with dates..." alongside the extracted date. This transparent method delivers high accuracy, with reliable handling of files exceeding 50 pages, and extraction of more than 20 fields —all accessible by specifying the enhanced_extract_agent ID in an API call or by using the "Autofill" click in Box. We will soon support rich elements like large taxonomies, tables, and checkboxes.

To learn more about how to use the Enhanced Extract Agent within the Box AI API, head over to our developer blog. 

Putting agentic extraction to work across your organization

• For Legal teams: Instantly surface exclusivity or change-of-control clauses across thousands of contracts, turning days of manual review into seconds of automated analysis.
• For Finance teams: Automate data entry from hundreds of invoices or earnings reports directly into financial systems, eliminating re-keying errors and ensuring compliance.
• For HR departments: Process new hire paperwork automatically, populating systems from intake forms and insurance cards without re-typing.

Control your AI ecosystem with the Box MCP Server

Enterprises want to leverage best-of-breed AI agents from platforms like Microsoft Copilot Studio and Salesforce’s Agentforce, but IT leaders face a critical challenge: How do you grant access to your sensitive enterprise content without creating dozens of complex, hard-to-manage custom integrations? The answer is centralized control. Our new remote Box MCP Server, now in beta, provides this control. It acts as a single, secure, and governable bridge that ensures any external AI agent must respect your existing Box security permissions and access policies. By building on the emerging Model Context Protocol open standard, you can connect your entire AI ecosystem, including Microsoft Copilot Studio, Microsoft Azure, and Salesforce Agentforce, while maintaining full control over your content.

"The future of enterprise AI relies on a securely connected and open ecosystem. Box's launch of their remote Model Context Protocol server is a key milestone, enabling our mutual customers to connect Azure services with their content in Box. As organizations scale AI across the enterprise, the ability to govern, manage and apply role-based access controls to Box's MCP Server with Azure API Center represent a meaningful step forward..." Balan Subramanian, General Manager Azure App Platform and Integration Services.

"An AI model's ability to reason effectively is fundamentally linked to the quality of the data it can access. Box's new remote MCP Server provides a secure and standardized bridge to the wealth of contextual knowledge stored in an organization's content. This allows models like Claude to deliver more accurate, relevant, and trustworthy responses, which is essential for tackling the most complex enterprise use cases. MCP provides a universal, open standard for connecting AI systems with data sources and we're excited to see Box bring enterprise content into that ecosystem." Michael Gerstenhaber, VP of Product Management at Anthropic

“We're entering a new era of enterprise AI and productivity, and our partnership with Box is driving secure agentic AI interoperability for human and digital workforce collaboration," said Gary Lerhaupt, VP, Product Architecture at Salesforce. "By bringing Box’s intelligent content solutions onto AgentExchange, we’re not just connecting Agentforce and Box's trusted MCP server, we’re also empowering businesses to gain new levels of productivity and insights from their content. AI agents can efficiently and securely access information needed to help employees work smarter.”

Manage AI costs and custom agent creation with AI Units and AI Studio

Systematically deploying AI requires not just technical capabilities, but also strong financial and development controls. We are rolling out key updates to provide exactly that.

Cost management with AI Units

To ensure you can scale AI usage with predictable costs, our refined AI Units system provides transparent insight into usage. By introducing new Standard and Premium model tiers, we are empowering you to manage your AI spend confidently and align costs with compute-intensive workloads.

Custom agent control with AI Studio

For organizations building their own custom solutions, Box AI Studio now provides admins with deeper control over the agents being created for their organization. New enhancements empower you to build and manage these custom agents more safely and effectively than ever, with features like:

• Real-time notifications about expiring or inactive AI models directly within the UI.
• Support for image and spreadsheet querying to unlock deeper insights from more data types
• Improved Model selection to make it easier to build agents.
• A new debug mode with a handy "Restore Configuration" option to quickly identify and fix issues.

A toolkit for enterprise AI governance and adoption

Systematically deploying AI requires robust governance controls to drive adoption. To empower administrators to manage and scale AI with confidence, we are rolling out significant updates to the Box AI Admin Console. The Admin Console now provides deeper, more precise control through a new, dedicated “Box AI” tab, which centralizes permissions for users and groups. This is enhanced by granular controls that allow admins to enable AI for specific services ensuring features are deployed where they add the most value.

Start your AI journey today

To learn more about how agentic AI can transform your business, watch our “What's Next in Box Al” webinar, now available on-demand.