Deflection rate is the wrong number to buy automated support on. AI escalation to human agents is the part that decides whether a customer leaves satisfied, and it is the part almost nobody writes down before signing.
The contact center market has half-admitted the problem. Through 2026 its pitch moved from how much volume the AI resolves alone to how well it hands off: vendor documentation now treats escalation as a designed mechanism, with explicit triggers (the customer asks, the customer is frustrated, the request is outside what the bot may handle), a summary that travels with the case, and routing to a matched agent, not as proof the automation failed. Spanish-language guides put it more bluntly. If the customer asks for a person, or arrives already angry, escalate immediately and send the full history. The vocabulary changed faster than the buying criteria did.
We built that handoff before the bots existed. For Tigerspike we designed and launched a crowdsourced customer support platform for a Fortune 500 company, where customers of major household brands were answered in live chat by brand advocates, not call center agents. We could not assume competence on the receiving end, so the routing weighed each helper's expertise, current availability and capacity, performance history and ratings, and language and time zone, and complex conversations escalated to professional support. The published results were "Real-time chat at scale" and "Community-driven support model validated".
The escalation trigger is the easy half. Deciding that a conversation should move is a rule you write in an afternoon. Deciding who receives it is a ranking problem that never stops being tuned, because the cheap answer is always whoever is free. Our own write-up records the lesson as matching quality beating matching speed, because customers preferred real help over fast non-answers. A fast connection to someone who cannot help does not read as a rescue. It reads as the second failure.
The work that held the Tigerspike platform together sat behind the queue. Customers rated every conversation, low-rated interactions went into review queues, and new helpers went through onboarding and certification before they could take anything. None of that is interesting and all of it carries weight. Without the review queue, quality drifts and nobody notices until the ratings arrive. An AI stack depends on the same scaffolding, and it is the first thing cut once the automation number looks good. Volume you no longer handle does not lower the standard for the volume you still handle. It concentrates the hard cases into fewer, longer conversations reaching a smaller bench.
We put the QA on every interaction rather than on the automation. Our delivery runs through partner centers we select and manage, nearshore in Mexico and Colombia or offshore in the Philippines, India and Egypt, and the AI sits in triage and agent-assist in front of the people who take over when a bot hands off, not as a wall between the customer and them. Cost still runs onshore, then nearshore, then offshore, and the honest version of that trade is hours and language, not a discount. An escalation that arrives in the customer's own time zone, in their own language, is a different event from one that waits overnight.
If you are buying automated support this year, specify the escalation path instead of the deflection target. The mechanics of triggers and context are a solved design problem, and our chatbot to human handoff guide walks through them. The buying decision is different, and four questions do most of the work. What travels with the conversation when it moves. How the receiving queue is ranked, and whether availability can outvote expertise. Who reads the badly rated transcripts, and how often. What the staffing commitment is in the week the deflection rate drops. A vendor who can answer those is describing an operation. One who answers with a percentage is describing a demo.
The customer never experiences your deflection rate. They experience the one conversation that needed a person.