"Have you tried turning it off and on again?" is the oldest joke in ops. Now an agent can try the restart, watch what happens, and record whether it helped. During an incident, another dashboard rarely solves the problem. Someone still has to pull together the alerts, logs, recent deployments, old incidents, and half-forgotten runbooks quickly enough to make a safe call.

I will look at the jobs where AI can help. An assistant can summarize noisy alerts and help with triage. An agent can test a possible cause against telemetry and, within carefully set limits, take action. It can also reach the wrong conclusion, so we will cover the safeguards that matter in production: useful context, limited permissions, human approval, audit trails, rollback, and a record of what worked before.

Azure SRE Agent is one implementation of these ideas. I will also share lessons from building AI tools that help engineers investigate outages across Azure. If you keep a service running, this talk will help you decide which parts of incident response are safe to automate, where a human still needs to make the call, and whether you should build or buy.

**Takeaways:**
- Match AI and agent capabilities to specific jobs during an incident.
- Set boundaries around context, permissions, approvals, and rollback.
- Make a practical build-versus-buy decision for AI-assisted reliability.