Containment Rate: Definition, Measurement and Benchmarks

Containment is a volume metric. A bot pointed at order-status and OTP-resend questions will hold most of its traffic. Point the same bot at billing disputes or a failed UPI mandate and a far larger share goes to people. The number moves with contact mix far more than with model quality, which is why comparing containment across two companies tells you almost nothing.

The definitions underneath the two numbers decide everything. Scope a conversation properly: one session, one customer, one contact reason. Stamp the contained outcome at close, from the system, not from a human review sample. Then declare what the denominator admits: does an abandoned session two turns in count as a bot conversation? Most reporting arguments about automation turn out to be definitional rather than technical.

Vendor-reported benchmarks put mature deployments around 70–90% and early rule-based bots closer to 20–40%. Treat those as directional. A BFSI bot handling loan collections and a D2C bot handling WISMO queries should not be held to the same target.

Never report containment alone. On its own it rewards a bot for stalling a customer into giving up. Pair it with resolution rate and CSAT, or the number is a cost proxy pretending to be a quality signal.

Often confused with: Deflection rate counts contacts that never reached a human queue at all, including help-centre visits where no bot session ever opened. Containment is narrower: the session opened in the bot and closed there.