Agent Separation-of-Duties Assessment

Do you maintain a current inventory of every AI agent operating in production, including ones spun up informally by developers?

For each agent, can you name the specific human or team accountable for its actions?

Do you know which systems, cloud, SaaS, code, and data, each agent can reach today?

Have you checked whether any single agent holds an entitlement combination that would be flagged as a conflict if a person held it, e.g. an agent that can both initiate and approve a payment, or both write code and approve its own deployment?

Are agent entitlements reviewed and re-certified on a cycle, the way you already certify human access?

If an agent's scope creeps beyond its original grant, is that change flagged to an owner automatically, rather than discovered later?

For a specific action an agent took last month, can you produce the record of what it was authorized to do, what it actually did, and who approved that authority?

Is that record immutable, or could it be edited after the fact by someone with database access?

When an agent is allowed to exceed normal policy, is that exception formally dispositioned by a named owner, rather than just quietly accepted?

If your regulator or a major customer asked for evidence of your agentic AI controls tomorrow, would you produce a live export, or would someone need to assemble it by hand?

Are you actively tracking the AI-relevant provisions in your applicable regulatory framework or supervisory guidance, as they apply to your agents?

Does your current tooling connect agent identity, entitlements, and control ownership in one place, or do those three things live in three different systems?