What is containment rate?

Why it matters for enterprise CX leaders

• Containment rate is the most-cited metric in voice AI procurement and the most loosely defined.
• Two deployments quoting the same containment rate can differ by 30 percentage points on a like-for-like measurement, depending on what the numerator and denominator include.
• For an honest internal measurement, subtract 7-day re-contact for the same intent — a call "contained" today that returns tomorrow has not been resolved.

The formula, stated precisely

Containment rate = contained calls / total in-scope calls, measured over a defined window. Every word in that sentence is a place where the number can be bent, which is why two deployments quoting "70% containment" can be describing completely different realities.

A contained call is one the automated system finished without handing to a human. Total in-scope calls is where the first argument happens: does it include callers who hung up in the first ten seconds? Callers who asked for something the AI was never configured to do? Misroutes from the IVR upstream? Each exclusion quietly raises the quoted rate.

Benchmarks by intent type

There is no honest single benchmark, because containment is dominated by intent mix. In production deployments, the bands that recur are:

• Transactional intents (balance checks, order status, appointment confirmation): 60–80%
• Mixed service intents (changes, amendments, simple troubleshooting): 35–55%
• Complex or emotionally loaded intents (complaints, disputes, vulnerability flags): 10–30%
• Blended enterprise call mix: most commonly 25–45%, and anything materially above that deserves scrutiny of the denominator

A worked example

A contact centre takes 100,000 calls a month. The voice AI engages on 60,000 of them; 30,000 finish without escalation. The vendor quotes 50% containment (30,000 / 60,000 engaged). The operations team measures 30% (30,000 / 100,000 total). Then the finance team subtracts the 6,000 callers who rang back within 7 days about the same issue and reports 24% resolved. Same deployment, three numbers — all arithmetic, no lying. The only defensible headline is the last one, which is why autonomous resolution rate exists as a separate metric.

How vendors game this metric

None of these are fabrications; all of them are choices made in the vendor's favour:

• Scope the denominator to "AI-handled calls" so callers the IVR misrouted or the AI rejected never appear in the maths
• Count silent abandonment as containment — a caller who hangs up mid-flow did not escalate, so they get counted as contained
• Quote a single champion intent (usually order status) as the deployment's containment rate
• Measure on the first month of a curated pilot cohort, before intent drift sets in
• Exclude out-of-hours and overflow calls, where containment is hardest, from the measurement window