AI support done well deflects the routine 60% and makes human agents faster on the rest. Done badly, it traps angry customers in loops and invents policies. The difference is grounding, escape hatches, and measuring resolution instead of deflection.
Why: Airlines and retailers have been held to policies their bots invented. Every ungrounded answer is a potential commitment plus a trust hit when it's wrong.
Why: Containment-at-all-costs optimizes the metric while destroying the relationship; the repeat-yourself handoff is the single most-hated support experience in survey data.
Why: Bounded authority makes bot errors cheap and reversible, and it composes with prompt-injection defense: a manipulated bot can only do small things.
Why: Agent-assist typically shows faster, less risky ROI than bots — 20-30% handle-time reductions with zero customer-facing failure modes — and agents become your QA layer for AI answers.
Why: The worst failures look like successes in metrics: confidently wrong answers that customers didn't bother to dispute. Only transcript review finds them.
Most AI-support programs still report deflection: the share of contacts a human never touched. It’s the wrong number. A customer who gave up on the bot and churned is “deflected”. A customer forced through three loops before finding the agent link is “deflected” until the last click. Resolution rate and post-contact satisfaction — measured on bot-handled conversations specifically — are the honest versions.
The strongest 2026 pattern is the hybrid: a grounded bot with narrow superpowers (instant order status, bounded refunds, rescheduling) that hands everything else to agents who are themselves AI-accelerated. Customers get instant answers where instant is possible and competent humans where it isn’t — which is, roughly, what they always wanted.