AI SREs Resolve Routine Incidents, But Engineers Lose Touch With Their Systems

Sylvain Kalache — an SRE at LinkedIn back in 2012 — reflects on his early prototype for a self-healing system and sees its realization in today's AI-driven incident response tools. But in a new blog post, he warns that these "AI SREs" are eroding the hands-on experience that engineers need to handle truly novel failures.

The ironies of automation

Kalache highlights the "Ironies of Automation," a concept from Lisanne Bainbridge's 1983 paper. Automation removes routine practice opportunities while leaving humans responsible for abnormal situations. This paradox applies directly to incident response:

• AI tools handle alerts, form hypotheses, query telemetry, and even implement fixes — reducing the need for human intervention on routine incidents.
• But those routine incidents are exactly where engineers build intuition for how their systems behave and fail.
• When an ambiguous, high-severity incident arises that automation can't solve, responders are expected to step in with less practice than they would have had previously.

Aviation as a model for training on rare failures

Kalache points to aviation as a precedent. Modern turbine engines experience fewer than one in-flight shutdown per 100,000 engine flight hours — rare enough that a pilot might never face it in real life. Yet pilots must respond correctly when it happens. For example, the TransAsia Airways Flight 235 crash occurred only 117 seconds after the first warning after the crew misidentified an engine failure.

To prepare, airline pilots undergo recurrent simulator training every six months under FAA rules, including engine-failure scenarios. Kalache suggests software engineers need similar "incident simulators" to rehearse rare and complex failures.

Simulators and AI as trainers

Kalache's current employer, Rootly, partnered with Uptime Labs to build exactly that: realistic incident simulations. Engineers take the incident commander role during a simulated e-commerce outage, using observability tools and coordinating with LLM-powered stakeholders in Slack. This provides safe practice for:

• Making sense of incomplete information
• Communicating clearly
• Coordinating responders
• Running the actual response

AI can also be used as a trainer — explaining its steps and evidence — but Kalache warns that watching won't replace doing. "You might pick up a few things from watching Serena Williams play," he writes, "but you only learn tennis by getting on the court."