IT operations already have automation: scripts, runbooks, RMM, orchestration. Agents add a planner that can choose tools from a prompt. That is powerful for investigation. It is reckless as an unsupervised executor on production identity and firewalls.
Use traditional automation when the path is known: patch rings, certificate renewal, account disable on leavers, backup verification. Those jobs need determinism, logs and a predictable schedule. An agent that “decides” to skip a ring is not a feature.
On one operations team, a deterministic rule closed a known noisy disk-space alert, which was reliable and easy to audit. Where an agent helped was the messy alerts: it pulled the last change record, recent deployments and related alerts into one summary and proposed a next step, and an engineer approved anything that touched production.
Use agents when the path is not known yet: correlating an incident across Azure Monitor, a firewall log and a change ticket; proposing a remediation; drafting a plan a human approves. The value is compressed investigation time, not removing operators.
The control plane matters more than the model. Tool permissions, environment isolation, approval gates, and an audit record of every call. If you cannot say which identity the agent used to reset a password, you do not have an operations system. You have a liability.
Where agents help, they sit against existing operational systems, not as a replacement for them, and not as a product in their own right. If a workflow is already reliable in code, leave it in code. Put the model on the messy edges, with a human on the approval path.
Want AI running on infrastructure you can trust? AI Infrastructure & Automation
All resources