Why most call-center AI pilots stall at 30% containment

Most call-center AI pilots get to about 30% containment and then quietly plateau. Leadership assumes the model needs to be smarter. In our experience across a dozen production deployments, the model is almost never the problem. The integration layer is.

Containment stalls when the agent can answer a question but cannot complete the task. It can tell a member their booking exists, but it cannot move the date, because it has read-only access to the reservation system. The caller still ends up with a human. On paper the AI 'handled' the call; in reality it deflected nothing.

The five fixes that move the needle are unglamorous: give the agent write access to the systems of record behind a permissions layer; model the top twenty intents as complete transactions, not Q&A; wire live CRM context into every turn so the agent knows who is calling; build escalation as a first-class path with full context handoff; and instrument every call so you can see exactly where containment breaks.

When we rebuilt one travel operator's flow around transactions instead of answers, containment moved from the low thirties into the seventies, and the calls that did escalate arrived with a human-readable summary instead of a cold transfer. The lesson holds: buy outcomes, not demos, and judge a voice agent by what it can finish, not what it can say.