Containment Rate vs Resolution Rate for AI Support Agents

The short answer

Containment rate tells you how often the bot kept a conversation to itself. Resolution rate tells you how often the customer left with their problem solved. The two are often reported as if they were the same, and that is where most AI support dashboards go wrong.

Why containment alone misleads

Containment is easy to compute: count the conversations with no handoff event and divide by the total. The metric has no idea why nobody escalated. A conversation counts as contained when:

• the bot answered correctly and the customer left satisfied,
• the bot answered wrongly and the customer gave up,
• the bot looped until the customer closed the tab, or
• there was no obvious way to reach a human, so nobody tried.

Only the first case is a success. The other three are failures that the metric reports as wins. Optimizing for containment can push a team to hide the “talk to a person” option, which raises the number and lowers customer satisfaction at the same time. The documented pattern chatbot loop without a human escalation path is what that looks like in production.

Why resolution rate is harder but more honest

Resolution needs evidence that the problem is gone. Common signals, from strongest to weakest:

1. No repeat contact on the same issue within a set window, for example 7 days.
2. The underlying system changed: the refund was issued, the password was reset, the order was updated.
3. An explicit customer confirmation or survey answer.
4. A closed ticket, which is weak, since tickets are sometimes closed on a polite phrase. See premature ticket closure from a politeness cue.

Resolution rate lags and needs a definition your team agrees on, but it is the number that tracks customer outcomes.

When to use which

Use containment for capacity planning: it tells you how much work reaches human agents. Use resolution to decide whether the AI agent is good. Report both together, and read the gap between them as a defect signal. A large gap means the bot is closing conversations it did not solve, which usually traces to escalation not being triggered or the agent ending the call prematurely.

A practical way to track the gap

Sample 50 contained conversations a week and label each as solved, abandoned, or wrong answer. The share labeled solved is your measured resolution among contained conversations. If that share is well below the reported containment rate, containment is overstating success. This is the broader problem of evaluation metrics that do not match real outcomes.