Use this guide for external-facing AI agents that will be used by the public.

How to Use This Guide

Depending on the phase and maturity of the AI agent being built, the accountable owner should fill out the appropriate release documentation below, submit it through the appropriate internal channels for approval, and make the submitted documentation available to users. The accountable owner must complete certain activities before and during each phase.

Phase 1: Internal Users Pilot

This phase, meant to last no more than six months, tests the agent with 10 or fewer internal, knowledgeable, trusted users who are aware of the risks and are experts in the agent’s subject matter area.

Builders should submit the completed pre–phase 1 checklist to the responsible agentic AI lead or a centralized approval body for approval and ensure the agent is only available for use by the specified users.

Download the pre–phase 1 checklist.

Builders or evaluators should use the phase 1 checklist to document the AI tool’s performance along several dimensions. Organizations should at a minimum enforce the submission of these statistics before approving phase 2.

Download the phase 1 checklist.

Phase 2: Non-Expert External Users Pilot

This phase, also meant to last no more than six months, expands testing to a limited number of external users. Such broader testing brings additional risk. Builders or evaluators should submit the pre–phase 2 checklist to the responsible agentic AI lead, and a centralized approval body should consider it for expedited approval, before external users are provided access to the AI tool. 

Download the pre–phase 2 checklist.

Builders or evaluators should use the phase 2 checklist to document the AI tool’s performance along several dimensions.

Download the phase 2 checklist.

Phase 3: Minimum Viable Release: External Launch

Builders or evaluators should complete the phase 3 checklist before launching a full version of the agent to a broad external audience. They should then submit checklists for all three phases to the responsible agentic AI lead and a centralized approval body for full review and approval.

Download the phase 3 checklist.

Phase 4: Long-Term Goals: Continuous Improvement to Best Practices

The phase 4 checklist represents additional elements to move agentic AI tools to the ideal state over time. Organizations may wish to use the content from this section to modify the phase 2 or 3 checklists in order to mitigate organization-specific risks.

Download the phase 4 checklist.

Next section: Governing Responsible AI