What changes when AI can take action?
A tool that drafts an answer and an agent that changes a business record create different review needs. An agent can cross systems, pass data to another tool, and repeat a mistake. Governance needs to describe actions and resources, not just the model in use.
For each proposed agent, name the task it supports, the person responsible for that work, the allowed systems, and the conditions that require a human decision.
Use six checks before approval
Purpose: Is the requested task clear, with an input and a useful output? Ownership: Is there a named human responsible for it? Scope: Are actions limited to the resources needed for that task?
Separation of duties: Could the proposed combination let one agent both prepare and approve a sensitive change? Evidence: Can a reviewer trace the request to the work description and relevant policy? Lifecycle: What happens when the owner changes jobs, leaves, or no longer needs the agent?
Example: draft a supplier record
The request is to turn an approved source packet into a draft supplier record. The allowed actions might be to read a designated folder and create a draft in a staging area. The agent should not infer that this includes bank-account changes, supplier activation, or payment release.
A reviewer should inspect the exact resources and actions. A broad “write suppliers” permission may exceed the job even if the human has that permission. The proposed scope should be the intersection of what the task needs, what the human may delegate, and what company policy allows.
A sensible request-to-review flow
Start with the person’s request and current task description. Read identity, HR, and access records from their actual sources. Keep source dates and distinguish verified facts from missing information.
Draft a manifest, show the requested and excluded actions, and route it to the responsible reviewer. Approval should identify the exact version reviewed. Live issuance, runtime enforcement, and revocation then need working integrations and evidence that each action succeeded.
Questions to ask a governance vendor
Can we trace an agent back to its owner and business task? Can you show the source of each permission? What happens if a source is stale or unavailable? Can a reviewer reduce scope? How do you detect conflicting combinations of permissions?
Ask for a demonstration of a role change and a failed revocation. A dashboard that records an intention is different from proof that the downstream system changed. Microsoft Entra and Okta document agent identity governance capabilities; evaluate their current requirements alongside your own systems.
Try the work-context layer
DutyGraph’s fictional company sample shows a request being evaluated against vendor-shaped records and turned into a proposed manifest. It is a demonstration of the workflow, with simulated approval and issuance.
The pilot starts earlier: leadership kickoff, team responses, participant-reviewed task cards, and advisor review. That work record gives a future agent request a concrete purpose. Download the manifest worksheet to use the same questions in your own review.
Sources
Primary references used in this guide.