Containment rate
Containment rate is the share of eligible automated interactions completed without a transfer to, or intervention from, a human agent during that interaction.
The numerator is the number of interactions the automated system completes on its own. The denominator is the set of interactions the organization has defined as eligible for automation. That eligibility rule is essential. Including calls that policy requires a person to handle will lower the rate for reasons unrelated to the automation's quality, while excluding difficult but intended cases can make the rate look stronger than it is.
A team should also define what “completed” means. A call that ends without a transfer is not necessarily contained successfully. The caller may hang up, receive an incomplete answer, or call again. A useful definition ties containment to a verified outcome, such as receiving the requested information, completing an allowed action, or accepting a clearly stated next step.
For AI phone agents, containment depends on more than conversational ability. The agent needs access to the approved information and actions required for the task, a call flow that handles common variations, and escalation rules for cases it should not complete. Missing data, unavailable systems, unclear caller intent, or policy boundaries can all produce legitimate handoffs.
Containment rate should therefore be read with other measures. First-call resolution helps show whether the issue stayed resolved. Repeat contacts can reveal callers who appeared contained but still needed help. Abandonment and customer-effort measures can expose friction. Reviewing escalation reasons shows whether transfers come from proper boundaries, missing capabilities, or avoidable failures.
The metric is most useful when segmented by call reason rather than treated as one target for all traffic. A routine information request and a sensitive exception should not have the same expected outcome. Raising containment by suppressing appropriate handoffs can make the service worse. The operational goal is to automate the calls that can be completed reliably while making the transition to a person straightforward when human judgment is needed.