Sylvain Kalache argues AI on-call tools will slow hard incident fixes

Sylvain Kalache published an argument on 4 September about what happens to on-call engineers when AI tools take over routine incidents. His prediction is specific: average time to resolution will fall across most incidents, and resolution time will rise sharply on complex ones, because responders have lost practical knowledge of their systems.

Kalache works at Rootly, an incident management company, and the post describes a simulation product Rootly built with Uptime Labs. That commercial interest is behind the argument and is worth stating first. The reasoning should be judged on its own.

The mechanism he names

The claim is not that AI incident tools work badly. He calls them fantastic, and says the value is real when one resolves a capacity issue at night without waking anybody.

The problem he names is what those routine incidents were doing for the team. They are how responders safely build an intuition for how a system behaves and how it fails. Remove them and the practice disappears too, while the responders are still responsible for the incident nobody has seen before.

He attributes the pattern to Lisanne Bainbridge's 1983 paper, The Ironies of Automation, which argued that automation cuts an operator's chances to practise routine work while leaving them accountable for abnormal situations, and concluded that such operators need more training than before automation rather than less.

The comparison with aviation

Kalache draws the parallel with flying, and the numbers are the useful part of it.

Modern turbine engines experience fewer than one in-flight shutdown per 100,000 engine flight hours, rare enough that a commercial pilot may finish a career without seeing one outside a simulator. When one does happen the response has to be immediate and correct. He cites TransAsia Airways Flight 235, where the right engine's propeller autofeathered after takeoff, the crew misidentified which engine had failed, and the aircraft stalled and crashed 117 seconds after the first warning.

Aviation's answer is scheduled practice. Under US FAA rules, captains complete recurrent training or a proficiency check every six months, including engine failure during takeoff.

His proposal for software follows the same pattern: incident simulation as a permanent part of on-call preparation rather than an occasional exercise. He also suggests using the agent itself as a teacher, asking it to explain the steps it took, the signals it read and the evidence behind its diagnosis, so a responder who did not run the incident still learns from it.

This is a prediction, and it is testable

Nothing here is a measured result about software teams. The engine and aviation figures are real and sourced, the Bainbridge paper is real, and the extension to on-call engineering is Kalache's own argument.

What makes it worth acting on is that it is cheap to check. A team automating its routine alerts already has the two numbers needed: median resolution time, and resolution time on the incidents that got escalated to a human. If the first falls while the second rises, the effect he describes is happening, and it will be visible in data the team already collects.

The practical step is to decide what replaces the practice before the automation removes it, because the gap only becomes visible during an incident the agent could not solve, which is the worst moment to discover that nobody on the rota has debugged this system by hand in a year.

Source

AI handles incidents, engineers lose touch with their systems, Sylvain Kalache, 4 September 2026.