For years, Site Reliability Engineering, from what I understand, has been built around a simple idea: automate what can be automated, but be cautious about where trust is placed. Production systems are valuable not because they’re difficult to operate, but because every decision has consequences. A small configuration mistake can become a customer-facing outage. A poorly timed deployment can impact thousands of users. Experience teaches engineers that production isn’t just another environment — it’s where judgment matters.
As AI becomes increasingly capable, many of us are beginning to ask the same question. Where does AI belong in production operations?
Some conversations immediately jump to autonomous incident response, automated remediation, or AI-driven deployments. While those ideas are exciting, I’ve found myself taking a much more conservative approach.
Instead of asking, “What can AI control?”, I started asking a different question:
“How much value can AI provide without controlling anything?”
Answer, I think, is — quite a lot.
One thing I’ve noticed over the years is that SREs rarely struggle because information is missing. They struggle because there’s too much information.
An incident begins. Alerts start firing left and right. Dashboards turn red. Logs begin filling up really fast. Slack channels become active all of sudden. People start opening Grafana, Kubernetes dashboards, Cloud Monitoring, GitLab pipelines, and documentation — all at the same time.
The challenge isn’t collecting data anymore. It’s understanding the situation quickly enough to make good decisions. This is exactly where I think AI has found its first truly valuable role.
Not replacing the engineer. Helping the engineer think faster.
If I imagine AI joining my SRE team tomorrow, the first thing I’d give it is read-only access. Read logs, Kubernetes events, monitoring dashboards, deployment history, Terraform plans, Git commits, incident timelines, and anything else that is normally involved in initial Incident troubleshooting.
Then ask it questions. Something like:
The moment AI starts making production changes, the conversation changes completely.
Restarting a deployment. Scaling workloads. Deleting resources. Changing Terraform code. Updating firewall rules.
Those aren’t information problems anymore. They’re ownership problems. Every experienced SRE knows that incidents are messy. Information is incomplete. Monitoring isn’t perfect. Sometimes the correct technical decision isn’t the correct business decision.
No AI model understands customer expectations, maintenance windows, contractual obligations, or organizational priorities unless those things have been explicitly provided.
Even then, accountability still belongs to people. That’s why I think production writes should remain human decisions.
I’ve started thinking about AI almost like an SRE intern. An incredibly fast intern. One that never gets tired. One that can read thousands of log lines in seconds. One that remembers documentation better than most humans.
But still an intern.
I’d happily ask that intern to prepare an incident timeline, summarize dashboards, highlight anomalies and suggest possible causes.
I would not ask that intern to restart production workloads without review. Not because the intern isn’t intelligent. Because the responsibility still belongs to the senior engineer who can take ownership.
I suspect AI systems will become increasingly capable over the next few years. Eventually they may earn broader operational ownership.
Perhaps they’ll restart stateless workloads under carefully defined policies and they’ll execute approved runbooks. Perhaps they’ll remediate well-understood failure scenarios. But I don’t think the journey starts there.
It starts by becoming exceptionally good at understanding production before attempting to control it.
Ironically, that’s exactly how we onboard new SREs. We don’t give production write access on day one. We let them observe, learn, dashboards and study previous incidents. Build intuition. Maybe AI deserves the same onboarding process.
Whenever people ask me whether AI will replace SREs, I think they’re asking the wrong question. The more interesting question is:
“How can AI help SREs become better at understanding production systems?”
For me, the answer begins with read-only access. Because before AI earns the right to change production, it should first prove that it truly understands it. Note: This article was written with the assistance of AI tools for structuring and drafting. The ideas, examples, and perspectives are based on the real-world experience in DevOps and cloud engineering over the years.