AI-driven operations fail when the estate has no reliable source of truth. Models cannot reason over missing CMDB data, unnamed subscriptions, or networks that exist only in a Visio from 2019. Preparation is inventory, telemetry and access design.
Inventory: what runs, where, and who owns it. Even a maintained spreadsheet plus tags is better than five conflicting CMDBs. Telemetry: metrics and logs that actually arrive, with resource IDs that match the inventory. Identity: workload identities for automation, not shared admin passwords in a vault nobody rotates.
Change records matter. If production changes happen outside a recorded path, an agent will propose work that collides with a human who already did it. Put a thin, honest change process in place before you add a planner.
Environment access should be least privilege and break-glassed. An operations agent in production with Owner is a future incident. Start in read-only investigation, then gated write actions, then a small set of well-tested remediations.
This work looks like infrastructure hygiene because it is. Organisations that skip it buy a demo. Organisations that do it can later add agents without gambling the estate.
All technical perspectives