AI's 95% incident fix may deskill SREs, SigNoz's Mathew warns

The debate over AI in operations has largely focused on how much work it can take off engineers' plates. Mathew's argument flips that: the risk is what engineers stop being able to do. She draws on a 1983 paper that described the same paradox in industrial systems, where automation reduced operators' chances to intervene in rare failures until their intervention skills eroded. The SRE version is concrete. If AI summarizes logs, engineers read fewer raw log streams. If AI proposes root causes, engineers form fewer hypotheses. If AI maps dependencies, engineers build fewer mental models of their systems.

The concern sharpens at the career level. Under PagerDuty's tier model, Tier 1 incidents (known problem, known fix) are fully automated and Tier 2 involves AI recommendations with human verification. If humans are left only with Tier 3 — novel, complex, cascading failures — Mathew asks how they are supposed to develop the intuition and feel for the system that such incidents demand without years of practice. Younger SREs who begin their careers in heavily automated environments face a more basic problem: not deskilling, but never skilling at all.

Mathew's proposed response is not rejection but deeper integration. She points to aviation, where the FAA requires pilots to log manual flying hours, and suggests SRE teams deliberately reserve some AI-capable incidents for human handling as a long-term investment in skill. The goal, she argues, is AI that strengthens SREs rather than replaces them, with humans deliberately designed into the most important 5% of tasks.