What is containment rate?

Containment rate is the share of customer conversations that an automated assistant or AI agent completes without handing over to a person. A conversation counts as contained if the customer never reached a human, whether or not their problem was solved. That last clause is why containment is a weak goal for an AI support agent.

Why does containment rate matter once agents act?

Containment is the most quoted measure of automated support because it maps directly to cost: every contained conversation is one a person did not handle. Vendors report it, budgets are built on it and targets are set against it.

The trouble is what it rewards. An agent can raise containment by not offering a handover, by answering vaguely until the customer gives up, or by resolving the easy part of a request and closing the chat. Each of those looks like success on the dashboard. For a support agent in a regulated firm, a high containment rate built that way is a conduct problem waiting to be found. We argue that containment rate is the wrong goal for a support agent.

How is containment rate measured in practice?

The usual formula is the number of conversations ended without human involvement, divided by all conversations the agent started. The details matter:

• Abandonment: a customer who closes the chat in frustration counts as contained unless you exclude them.
• Repeat contact: a customer who comes back the next day by phone was contained the first time.
• Channel switch: a customer who gives up on chat and emails instead is contained in chat terms.

Without those adjustments the figure flatters. With them, it becomes a rough measure of how often customers did not need a person, which is useful but still says nothing about whether the answer was right.

An example from a retail bank

Take a retail bank whose chat agent reports high containment for card queries. Looking closer, the team finds that many "contained" customers asked about a disputed transaction, received a general explanation of the dispute process and left. Several days later, a large share of them phoned the contact centre to raise the dispute.

The chat did not fail on its own terms; it simply was not allowed to raise disputes, and it did not hand those customers to someone who could. Containment counted the deferral as a success. Adding a handover for dispute intents lowered containment and cut repeat calls.

How is containment different from resolution?

Containment asks whether a person got involved. Resolution asks whether the customer's problem was solved. A conversation can be contained and unresolved, as in the bank example, or escalated and resolved, which is often the best outcome for a complex case.

Resolution rate is harder to measure, because it needs evidence that the issue did not come back, but it is the measure that reflects customer outcomes. Treat containment as a cost indicator and resolution as the quality indicator, and never let the first rise at the expense of the second.

Which terms sit next to containment rate?

• Deflection rate: contacts that never reached an assisted channel at all.
• Escalation rate: the share of conversations handed to a person, roughly the other side of containment. How high an AI support agent's escalation rate should be depends on what is being escalated.
• Abandonment rate: conversations the customer left before an outcome.

Where Pryme fits

Pryme certifies a support agent that knows when to say no. Attest scores whether the agent escalated when the case called for it, so an agent that keeps containment high by withholding handover fails the escalation gate and stays at a lower depth.