AI agents: operational approval before broad rollout

AI agents are no longer being introduced in organisations merely as individual chat features. On 25 August, OpenAI introduced an Admin plugin for ChatGPT Work and Codex that brings usage data, access, limits and supported administration actions into one conversation. Microsoft is also enabling Azure Copilot users to select specialised agents and turn them on individually. The shared message matters more than either product: an agent rollout needs an operational approval, not one broad switch for the entire organisation.

Administration itself becomes an agent workflow

According to OpenAI, authorised administrators can use the plugin to review usage and credit consumption, manage groups, analyse effective permissions and handle spending requests. The functions are intended to respect existing roles and approvals, and changes return structured results.

That is useful, but it does not transfer accountability. If a conversation can change access, limits or group membership, the organisation must define which action may proceed automatically and which needs human review before execution. A read-only report and lowering a limit are not the same risk class.

Individual agents instead of global activation

Microsoft describes a different but complementary mechanism for Azure Copilot: teams can select specialised agents for tasks such as incidents, deployments, optimisation or resilience directly. Administrators can enable or disable agents individually and evaluate them gradually. This is a useful pattern for every platform: define the business purpose, tool rights, data access and success criteria for each agent – not only after rollout.

Three proofs before scaling

For DACH organisations, a short operational-approval check for every agent is advisable:

• Purpose and scope: Which clearly defined work process does the agent support, and which systems may it use?
• Decision boundary: Which actions may it prepare, which may run automatically, and who approves exceptions?
• Operational evidence: Which usage, cost, permission and outcome data are reviewed regularly – and how is the agent withdrawn when deviations occur?

Start with one agent and one user group, not an organisation-wide licence release. Document the target process first, then measure actual usage, failed attempts and necessary escalations after four weeks. That turns the introduction of a new assistant into a reversible operational trial with accountable leadership.