At some point in the last few years, containment became the headline metric of conversation automation. Boards see it. Investors ask about it. Vendors lead with it. The logic sounds reasonable on the surface: if the AI handled the call without a human, that’s a success. But often, the data is lying to you.
Unfortunately, the organizations that figure this out late are paying for it in churn they can’t easily trace back to the source. To understand why, we need to review what containment actually measures, how it hides the true cost of churn, and why resolution is a better metric.
What Containment Actually Measures
Containment measures one thing: whether a human agent got involved. It says nothing about whether the customer’s question was answered, whether their problem was solved, or whether they’ll be back tomorrow, frustrated, on a different channel, ready to cancel.
Consider these scenarios, which would all check the box for containment:
- A customer who couldn’t get the AI agent to understand their billing dispute and finally gave up.
- A customer who repeated their email address four times and hung up rather than sit through another retry.
- A customer who navigated a deflection loop designed intentionally or not, to make escalating more trouble than it was worth.
These are not success stories. They are quiet failures that don’t announce themselves in your metrics. Containment captures the absence of escalation, it doesn’t capture the absence of frustration. Those are very different things, and confusing them is expensive.
How Containment Hides Customer Churn
Here’s what makes containment a particularly dangerous north star: the damage it hides doesn’t show up immediately.
A customer who leaves a contained interaction unsatisfied doesn’t typically cancel on the spot. They go about their day. They try again later through a different channel. They start paying slightly less attention to your promotional emails. They begin, almost imperceptibly, to disengage. Weeks or months later, when renewal comes around, the decision to leave feels like it came from nowhere. It didn’t. It came from that conversation, and probably the one before it, and the one before that.
The churn signal is real. It’s just separated from the containment metric by enough time that the connection is easy to miss. Containment goes up. Churn goes up. The teams responsible for each number are looking at different dashboards, and nobody draws the line between them until it’s too late.
This is the false sense of security that containment creates. It lets you believe the automation is working while the underlying customer relationship quietly erodes. By the time the churn data is loud enough to force a conversation, you’ve lost customers you didn’t know you were losing.
Engineered Containment Doesn’t Help Customers
There’s a blunter version of this worth saying plainly: some containment is engineered.
Deflection flows that make reaching a human cumbersome. Menu structures that bury the “speak to an agent” option. Retry prompts that keep looping rather than gracefully routing. These designs can drive containment rates up without solving a single customer problem. They are, by any honest definition, traps, and customers recognize them even if they can’t articulate exactly what happened.
The frontline worker industry knows this. The best operators have always known that a customer who gives up is not a customer who was served. Conversation automation doesn’t change that principle. It just gives you new and more sophisticated ways to obscure it in your reporting if you’re not careful about what you measure.
Resolution Is the Honest Metric
The alternative to containment isn’t a softer version of the same idea. It’s a fundamentally different question which is why resolution is the only honest metric.
Resolution asks: did the customer get what they came for? Not, “Did the AI finish the conversation?” or “Did the customer hang up before reaching an agent?” Did the customer’s specific intent, the reason they contacted you in the first place, get addressed, accurately and completely, in a way they’d accept again?
That’s a harder number to hit. It’s also the only one that tells you something true about what your AI agent is actually doing for your business.
Resolution data, measured honestly, gives you two things containment never can.
First, it gives you real signal. When resolution rates are high, you know that the automation is working. When they’re low, you know exactly where to look. Second, it gives you granular visibility. Which intents are resolving cleanly? Which ones are consistently failing? Where in the conversation is the customer’s need getting lost?
The organizations getting the most durable value from conversation automation are the ones who made this trade deliberately. They accepted that their “automation rate” would look smaller on a slide, and they got an honest picture of what their AI agent was actually achieving in return. Then they improved it systematically, because they could see what was working and what wasn’t.
A few recommendations to balance containment and resolution that have worked over the past few years:
- When a user asks for a human agent, provide a “what’s in it for me?” response. Tell them how long it generally takes to reach a human agent and what the agent can truly solve to provide them confidence in the AI’s ability to assist them.
- Set retry limits for specific collections at no more than 2 to avoid trapping users that aren’t having success.
- Set global retry limits no more than 4 to ensure the user isn’t hitting retries over-and-over throughout the conversation.
The Metric Shapes the Behavior
This is ultimately a question of organizational incentives. The metric you optimize for determines the decisions you make and the designs you build.
Optimize for containment, and you will, consciously or not, make design choices that contain customers. Loops that run a little longer before offering a human. Escalation paths that require a little more effort to reach. These aren’t always deliberate. But they are predictable consequences of asking the wrong question.
Optimize for resolution, and the incentives invert. Every unnecessary step in the conversation becomes a problem to eliminate. Every retry is a signal that something in the design is off. The entire operation starts pulling in the direction of actually serving the customer, because that’s what the metric rewards.
You can’t measure your way to a good customer experience with a metric designed to hide a bad one.
If you want to talk about what honest resolution management looks like in practice, talk to the Zenarate team.
Featured Resources
Press Release
