StackOne Launches Tool Policies for Field-Level AI Agent Permissions

Summary

StackOne has launched Tool Policies, a policy engine for governing what enterprise AI agents can do, see, and send through connected tools. The product sits above permissions in HR, CRM, and finance systems and applies rules across connectors, MCP clients, APIs, SDKs, and agent platforms, including Claude, ChatGPT, Copilot, Gemini Enterprise, Glean, and custom agents. Administrators can scope policies to users, identity-provider groups, projects, connector categories, connectors, linked accounts, or matching tool names. Five rule types cover tool blocking, input-value limits, field masking, PII redaction, and semantic checks based on plain-language questions. Policies inspect requests before a business system is called and responses before they reach an agent; conflicting rules default to deny. The product can be configured through a no-code builder, StackOne Assistant, Cedar, or JSON, and Monitor mode records affected requests before enforcement. Examples include allowing Accounts Payable agents to make payments only between $25 and $500, blocking supplier-bank-detail changes, preventing contractors from using tools matching “delete,” and masking bank details or salary data. For healthcare workflows, StackOne says redaction and field rules can keep identifiers out of agent context and logs; an Enterprise Business Associate Agreement is available when PHI is in scope. StackOne also reports 97.9% span F1 for its PII redaction model and 95.9% balanced accuracy for its Defender prompt-injection protection, compared in the article with Microsoft Presidio and Meta Prompt Guard v2. Tool Policies are available in early access on Gateway Team and Gateway Enterprise plans.