AI-assisted incident response is moving from novelty to operating layer. Sylvain Kalache argues that modern AI SRE tools can inspect alerts, form hypotheses, query telemetry, correlate deployments, and sometimes implement fixes without waking engineers. The near-term metric win is obvious: routine incidents should resolve faster.
The risk is not that agents fail on easy incidents. The risk is that they succeed so often that humans stop practicing. Kalache ties the problem to Lisanne Bainbridge's 1983 work on the ironies of automation: the more reliable automation becomes, the more humans are left with only abnormal, high-stakes failures. In software operations, that means responders may see fewer routine degradations, then face ambiguous incidents with weaker system intuition.
Grey Haven's read: AI ops tools should be evaluated as training systems, not just resolution systems. Lower MTTR is useful, but it is not the same thing as organizational resilience. If an agent handles every low-severity alert, the team may be borrowing reliability from future incident response capacity.
Operators should ask three questions before expanding AI incident automation. Which incident classes can the agent close without eroding human practice? Which scenarios must remain in rotation for drills and shadow response? Which metrics track responder comprehension, not just response time?
The watch item is whether AI SRE vendors start shipping practice loops: replay, simulation, explanation audits, and forced human takeover drills. Without that, the product category may create a new form of comprehension debt.
Source: Sylvain Kalache, “AI handles incidents, engineers lose touch with their systems.”
