A policy for autonomous agents should be specific enough to guide real decisions. Generic responsible AI language is useful as a principle, but teams building agents need operational rules: what the agent may do, what it may access, when review is required, and what evidence must be kept.
The template should be short enough to use and complete enough to enforce.
Purpose and scope
Start with scope. Define what counts as an AI agent: software that uses an AI model to plan, decide, retrieve information, call tools, or take actions in pursuit of a goal. Include internally built agents, vendor agents, workflow automations, browser agents, and product features.
State that experiments, pilots, and production systems may have different requirements, but agents with sensitive data or action-taking authority require review before use.
Allowed and prohibited uses
List allowed uses such as drafting, summarization, analysis, triage, internal productivity, and supervised workflow support. Then list prohibited or restricted uses, such as unsupervised decisions affecting legal rights, unauthorized processing of personal data, credential handling, security bypass, deceptive impersonation, or actions outside approved tools.
Restrictions should be tied to risk, not fear. The point is to make boundaries clear.
Autonomy limits
Define autonomy levels: suggest, draft, recommend, act with approval, act within limits, and act without review. For each level, specify required approvals and controls.
For example, an agent may draft customer messages but must not send them without human approval until the system is separately reviewed.
Data and tool rules
Require data classification, least-privilege access, approved retrieval sources, vendor review, retention rules, and restrictions on confidential or regulated data. Require a tool inventory for every agent, including permissions, action types, and approval points.
Operations
Add requirements for logging, monitoring, incident reporting, change review, periodic review, and evidence retention. The policy should say what happens when prompts, models, tools, or user populations change.
A good policy template gives teams a path to launch responsibly. It does not make every use case high risk, but it does make hidden autonomy and undocumented data access unacceptable.