Feature-flagging an AI capability behaves nothing like feature-flagging a button. The cost spike, the latency tail, the silent quality regression — none of them play by the rules your team learned shipping classical features. The discipline that separates safe AI rollouts from expensive surprises.
Feature-flagging an AI capability behaves nothing like feature-flagging a button. Cost can spike 10× in fifteen minutes, latency tails do not behave at scale the way they did at 5%, and quality regressions surface only through customer complaints — weeks late. A safe AI rollout instruments four dimensions, not one, and gates ramp on each.
Classical flags toggle deterministic behaviour with known cost. AI flags toggle probabilistic systems whose cost depends on input distribution, latency depends on provider load, and quality depends on factors the team does not fully control. Every dimension of risk classical flags let teams ignore — cost, latency, drift — AI flags actively introduce.
Five-stage with explicit pass criteria: 5% for cohort-shape validation, 25% for first real signal, 50% with mandatory observation pause, 100% only after checkpoint review. Every step has a defined backout. The cadence is slower than classical feature rollout because the dimensions to watch are wider.
AI workloads can spike 10× in fifteen minutes. Hourly cost dashboards are useless for incidents at that timescale — by the time the alert fires, the damage is done. Sub-five-minute monitoring on a proxy metric (token count rather than billed dollars) is the difference between catching the spike and reading about it in tomorrow's post-mortem.