📊Evaluationalso: deflection, self-service rate
Deflection rate
In support agents: the percentage of customer contacts the agent resolves fully without escalating to a human.
Deflection rate is the single most-watched metric for AI customer support agents. A deployment at 65–75% deflection on tier-1 chat is doing real work; below 50% usually means a knowledge-base or routing problem rather than an agent problem.
Net deflection (what actually reduces human workload) is the metric to optimize. Raw deflection counts can be gamed — an agent that resolves easy contacts but escalates everything hard hasn't moved the needle.
Where this shows up
Frequently asked
What deflection rate should I expect?+
For tier-1 chat, 55–75% sustained is realistic in 2026. Voice lags chat by 10–15 percentage points. Anything claiming 90%+ is likely a metric trick or a very narrow domain.