It is the quarterly review, and the slide looks good. The new AI assistant handled 48% of contacts without a person. Average handle time is down. Somebody on the call says the word efficiency.
Then the contact center supervisor speaks up. The phones are not quieter. The front desk is busier than it was in the spring. Two of her agents have started asking callers, gently, whether they already tried the chat.
Both things are true at once. The AI is containing calls, and the work is still showing up somewhere. The gap is not a technology problem. It is a measurement problem, and it is fixable before the next contract renewal.
This post covers what two recent pieces of customer experience research say about how people actually behave with AI agents, why containment on its own is a misleading scoreboard, and the specific numbers to report beside it.
The first comes from consumer research published in October 2026 by contact center software company Five9, based on an April survey of 3,000 consumers in the United States, United Kingdom, and Germany fielded by Hanover Research. A vendor paid for it, so read it alongside independent data, but the pattern is consistent with other studies.
The headline numbers:
71% say it is very or extremely important to know they are talking to AI.
46% say the top reason they abandon AI is that it misunderstood them repeatedly.
Independent research points the same way. In August 2026, Gartner reported that 87% of the 3,566 customers it surveyed said companies using generative AI for service must provide access to a human agent.
The second piece is an argument rather than a survey. In a CX Today interview this week, Natalie Beckerman, chief business officer at iQor, made the case that containment, deflection, and shorter calls can make a contact center look efficient while customer effort simply moves somewhere else. A second call. A different channel. A customer who quietly gives up. Her recommendation is to measure resolution the first time rather than celebrating containment on its own. Worth noting that iQor runs human-staffed service operations, so it has a position in this debate.
Put the two together and you get something useful. Callers will accept AI on a short leash, and the metric most pilots are judged on does not tell you whether the leash was long enough.
Containment rate: the share of contacts the AI finishes without handing off to a person.
Deflection: steering a contact away from the phone queue, usually to chat, a web page, or a self-service form.
First call resolution: the share of contacts fully resolved on the first attempt, with no callback and no channel switch.
Warm handoff: a transfer where the caller's details and conversation history move with them, so nobody has to start over.
Containment is easy to measure, easy to chart, and easy to turn into a cost figure. That is exactly why it ends up on the executive slide. The trouble is what it quietly rewards. A system scored only on containment is being asked to keep people away from a person, and the research says that is the one thing callers will not tolerate for long.
There is a second problem. Containment counts what happened inside one interaction. The cost of a bad interaction shows up outside it, in the callback three hours later, the email to a department inbox, the walk-in at the counter, or the complaint that reaches an elected official. None of those land in the dashboard that produced the number.
For public agencies and colleges, the stakes are a little different than they are for a retailer. Residents, students, and families usually cannot take their business elsewhere. They just try again through another door, and every extra attempt costs staff time and goodwill you do not get back easily.
None of these require a new platform. Most contact center reporting tools can produce them if someone asks.
If containment rises and all six hold steady, that is a real win worth celebrating. If containment rises while repeat contacts and channel switching climb, you have a dashboard win and a service problem.
Pro: Measuring resolution gives you an honest picture and protects service levels while you expand automation.
Con: Your containment number will probably look worse once an easy path to a person exists. Explain that to leadership before the first quarterly review, not after it.
Pro: Repeat contact and channel switching data usually already exists in your reporting tools, so this costs analysis time rather than budget.
Con: Connecting a chat session to a later phone call takes some setup, and small teams have to decide what is worth the effort.
Pro: Agreeing on success measures during evaluation gives you leverage. A vendor that can report right first time is a vendor confident in its product.
Con: Custom reporting requirements can slow a procurement cycle and may carry a cost. Decide early which numbers are mandatory and which are nice to have.
The best time to define success is during evaluation, while you still have leverage. Once a platform is live and its dashboard defines what good looks like, changing the definition is a much harder conversation.
Three things to put in the requirements, not the kickoff meeting:
The measures that will be reported, including repeat contacts and right first time
The escalation rule, written down, and a demonstration of the handoff in your own call flows
Who produces the reports, how often, and in what format
Callers are willing to work with AI. They have told researchers as much. What they want to know is that a person is reachable and that their problem is actually solved. Measure those two things and the technology decision gets a lot easier.
If you are evaluating AI for your contact center and want a supplier-neutral view of how different platforms handle handoffs and reporting, Maverick helps you compare options, pilot them with your own callers, and plan the rollout, from evaluation through go-live and after.