AI Agent Governance: A Simple Starting Policy
AI agent governance is the set of rules for what your agents may do, who is responsible for them and how you find out when something goes wrong. It matters because agents act with real permissions, and it's easy for them to spread across a business before anyone has counted them.
You don't need a committee to begin. A short policy and a simple list will put you ahead of most teams that are still improvising.
Start with a register
You can't govern agents you don't know about. Create one list of every agent in use, including tools that staff turned on themselves. Record:
- Name and purpose: what it does
- Owner: the person accountable for it
- Systems and data: what it can reach
- Actions: what it can change
- Risk tier: see below
- Last reviewed: the date someone checked it
A spreadsheet is enough to start.
Sort agents into risk tiers
| Tier | What the agent does | Controls |
|---|---|---|
| Low | Reads non-sensitive data, drafts for a person | Owner, logging |
| Medium | Takes reversible actions, sees internal data | Plus limits and regular sampling of its work |
| High | Moves money, contacts customers, touches personal data or takes irreversible actions | Plus human approval, security review and testing before each change |
Adapt these tiers to your business. The principle is that the controls should match the potential damage, so that low-risk tools don't get buried in process.
A one-page starting policy
Copy this and edit it:
- Every agent is registered and has a named owner. No owner, no agent.
- Agents have their own identities and minimum permissions. Shared admin accounts aren't allowed.
- Limits are enforced in code. A rule that exists only in a prompt doesn't count as a control.
- Humans approve high-impact actions. The tier table defines these.
- All actions are logged. Logs are kept for [your chosen period].
- Changes get tested. That includes changes to the model, prompts and tools.
- Any agent can be shut off quickly. The owner knows how.
- Incidents are reported to [named person], who decides on the response.
- Agents are reviewed every [period], and unused ones are removed.
Have your legal or compliance adviser review it, especially if you work in a regulated industry.
Use existing frameworks for depth
When you need more, you don't have to start from nothing. The NIST AI Risk Management Framework organizes this work into four functions: Govern, Map, Measure and Manage. ISO/IEC 42001 is a standard for AI management systems. If you operate in Europe, check how the EU AI Act applies to you.
How to tell whether it works
A policy nobody follows gives you false comfort. A few simple checks will tell you how it's going:
- Can you produce a full list of your agents today?
- Does every one have an owner who knows they're the owner?
- Could you explain why an agent took a specific action last week?
- How long would it take you to shut one off? Try it.
- Have any agents been found that weren't on the list?
Good governance should also make it easier to say yes. When teams know the rules, low-risk agents can move quickly. For the technical side, see the AI agent security checklist.
Talk to us
Need agents with these controls built in from the start? DPR Software Labs builds agents with clear limits, logs and approval steps.