The number that matters is not the containment rate
Contact centre vendors report containment rate: the share of contacts an AI assistant handles without a human. It is the easiest number to publish and the least informative about what the customer actually experienced. J.D. Power's 2026 U.S. Wireless Carrier Satisfaction Study, Volume 2, based on 57,339 customer responses fielded December 2025 through May 2026, measured something different: what happens to satisfaction depending on whether the AI assistant actually solved the problem (J.D. Power, 2026).
Among wireless customers who called in for phone-based support, 45% of interactions involved an AI assistant, and that assistant resolved the request in 48% of those cases (J.D. Power, 2026). Split the outcome and the picture changes completely. When the AI assistant resolved the issue, overall satisfaction rose 83 points on J.D. Power's 1,000-point scale. When it did not, and the customer was pushed to a live agent anyway, overall satisfaction across all AI-involved calls came in 20 points lower than calls that never touched an AI assistant at all (J.D. Power, 2026).
That 20-point gap is the actual cost of a failed handoff. It is not neutral, it is worse than never deploying the assistant in the first place. A customer who gets to a human by dialing straight through starts the interaction fresh. A customer who gets to the same human after an AI assistant already failed to help arrives already frustrated, and the final satisfaction score reflects both experiences combined.
The same pattern shows up in banking
J.D. Power's 2026 U.S. Digital Banking and Credit Card Mobile App Satisfaction Studies, covering both U.S. and Canadian markets and released in May and June 2026, found a comparable split in a completely different channel: in-app virtual assistants rather than phone support (J.D. Power, 2026). Virtual assistants handle everyday questions and simple transactions well, and customers use them for exactly that. But on fraud alerts and complex account issues, the same assistants consistently fall short, and limited escalation paths trap customers in self-service loops instead of routing them to a person who can actually act (J.D. Power, 2026). Jennifer White, J.D. Power's managing director of financial services intelligence, put the fix in design terms rather than model terms: banks need virtual assistants built to recognize complexity and escalate seamlessly, not just to automate the easy cases (J.D. Power, 2026).
Two different industries, two different channels, phone and in-app chat, and the same shape of result. That consistency is what makes the finding worth building a policy around rather than treating as one vendor's isolated numbers.
Why this is an escalation design problem, not a model quality problem
Neither study attributes the failure to weak language models. The mechanism is structural: an AI assistant that cannot resolve a request has two paths, hand off cleanly with context intact, or let the customer discover the failure themselves and start over with a human. Most deployments still do the second. The customer repeats information, re-explains the problem, and now measures the whole interaction, AI included, as one experience. The satisfaction hit is not really a verdict on the AI's competence. It is a verdict on how the failure was handled after competence ran out.
This is why containment rate alone is a misleading north star for CX Tech buyers. A high containment rate paired with a bad escalation path can be dragging satisfaction down even while the reported automation number looks like a win. The number worth asking a vendor for is not "what percentage did the AI handle," but "what happens to satisfaction on the contacts it didn't resolve, and how does the handoff work."
What this means for CX teams evaluating AI support
Before expanding AI support, ask for outcome-split satisfaction data, not blended averages: what does satisfaction look like on resolved AI contacts versus failed ones that escalated. Ask how the assistant hands off context to a human agent, since a clean handoff appears to be the variable separating the wins from the losses in both datasets. And treat containment rate as a cost metric, not a quality metric. It tells a team how much human labor was avoided. It says nothing about whether the customer who didn't get resolved left worse off than if no AI had answered at all.
Sources: J.D. Power. "2026 U.S. Wireless Carrier Satisfaction Study — Volume 2." 2026. · J.D. Power. "AI-Powered Virtual Assistants Struggle with Complex Tasks in Bank and Credit Card Apps and Websites." 2026. · J.D. Power. "2026 Canada Digital Banking and Credit Card Mobile App Satisfaction Study." 2026.