What 150+ deployments taught us about voice AI containment rates
The real number is lower than the vendor decks say — and that's fine, if you design the handoff right.
Lead image — What 150+ deployments taught us about voice AI containment ratesREF: INS-ART-01JPG, PNG or WebP · 16:9 · min 1920×1080pxImage placeholder — supply asset with this REF
Editorial note: this piece is still an outline — the section list below is what it will cover. The full article is being drafted.
Vendor decks quote containment — the share of calls an AI agent resolves without a human — somewhere north of 80%. Across our own deployments the honest figure sits well below that, and it varies more by call type than by model.
This piece will set out what we actually measure, why a lower containment rate is not a failure, and how the design of the handoff to a human determines whether callers experience the gap as competence or as a dead end.
What this article will cover
- How containment is defined, and the three ways the definition gets gamed.
- Containment by intent class: balance checks versus disputes versus retention.
- Designing the escalation so context travels with the caller.
- The metrics we report to clients instead of a single containment number.