AI governance vs. agent governance
AI governance is the organizational work of deciding how AI may be developed, acquired, used, and reviewed. Agent governance applies that work to software that can choose and carry out actions through tools. An agent can retrieve records, submit work, or change a system. Governing it means deciding who permits those actions, who maintains its connections, and who can intervene.
The NIST AI Risk Management Framework addresses risks across the design, development, use, and evaluation of AI systems. It is voluntary guidance, not a certification for an individual agent.
Explore Dome: Dome Platform · Docs: Security tutorial.
Who is responsible for each decision?
Suppose an insurer considers an agent that assembles documents for a claims reviewer. The AI governance review asks whether this is an appropriate use of AI, how personal information may be used, and how people can challenge an outcome. Those questions still matter if the application never invokes a tool.
Agent governance adds operational decisions. Which records may the agent retrieve? Can it only prepare a draft, or can it submit a claim? Who owns its access after the pilot team moves on? What should happen if the reviewer leaves the organization?
| Decision | Primary responsibility |
|---|---|
| Approve the use case and acceptable risks | Business owner with relevant risk specialists |
| Define permitted actions and data access | Service owner and security team |
| Implement and maintain controls | Platform and application teams |
| Review outcomes and handle exceptions | Named business and operational owners |
The same person may fill several roles, but approval, maintenance, and exception handling each need a named owner.
Explore Dome: Agent Registry · Docs: Manage agents.
Be specific about what the human reviewer does
“Human-in-the-loop” generally means a person participates in a particular decision before the workflow continues. “Human-on-the-loop” commonly describes oversight with the ability to intervene. Usage varies, so specify the behavior instead of relying on either label.
For the claims agent, the reviewer needs to see the destination and attached documents before approving a submission. An operator monitoring unusual access needs the ability to suspend the agent and time to respond. The first person approves a specific action; the second oversees ongoing activity.
Define what the person approves, how long that approval lasts, and what happens if nobody responds. Put those requirements in the workflow, rather than relying on an instruction to the model.
Explore Dome: Authorization · Docs: Authorization model.
Record the decision and when to revisit it
For one use case, record its approved purpose, owner, allowed actions, escalation contact, and the results or events that should trigger a review. Revisit that record when the agent gains a new tool, changes its data source, or begins acting for a different group.
Record when the work should stop, too. A trial might end when ownership changes or a required source becomes unavailable. Without a stopping condition, a temporary exception can quietly become permanent.
Dome contributes agent records, configured call controls, and audit evidence on governed paths. It does not select the organization's acceptable risks or determine whether a business outcome was fair. Use agentic operations for the day-to-day operating discipline and agent authorization for individual action decisions.
Explore Dome: Audit · Docs: Audit events.