Containment Rate for a Voice Agent — and Why It's Easy to Game

The number the business case rests on

Containment rate — the share of calls a voice agent handles end to end without a human — is the headline metric, because it's the main input to the cost case. It's also the easiest metric to inflate without improving anything, which is why voice AI containment rate should never be read on its own.

Three ways it lies

1. The agent dead-ends the hard calls. It can't help, doesn't hand off, and the call ends. That "contains," and the customer is worse off than if they'd waited in a queue.

2. Confident wrong answers. The agent gives a wrong answer with conviction, the call ends, and the customer calls back tomorrow — or acts on the wrong information and it's a bigger problem.

3. The denominator is gamed. Route hard call types away from the agent before they're counted, and the containment rate on what's left looks great.

The metrics that keep it honest

Measure these alongside containment, always:

• Escalation reason breakdown. Caller asked, low confidence, sentiment, restricted topic. A pile-up of "low confidence" means thin intent coverage; a pile-up of "caller asked" means callers don't trust the agent yet.
• CSAT split by contained vs escalated. If contained calls score lower than escalated ones, the agent is holding calls it shouldn't.
• Wrong-information rate. Spot-check transcripts for any factual claim the agent made that wasn't backed by a tool result. This should be near zero by design; anything else is a grounding bug.
• Repeat-contact rate. Callers back within a day about the same thing — a contained call that didn't actually resolve.

Setting the target the right way

The goal is the lowest containment rate at which contained-call CSAT and repeat-contact stay healthy — not the highest number you can hit. A high containment rate with unhappy contained callers is worse than a lower, honest one.

The honest ceiling

In lending, only calls that don't touch a decision can be contained, so the ceiling is set by your call mix, not your model. On our voice AI engagement the automation rate is bounded by the rule that anything adverse-action-adjacent goes to a human — the 35% call-centre load reduction is what that mix allows, not a number the model could push higher on its own.

How to read a containment number you're handed

When a vendor or a team reports "70% containment," it's not an answer, it's the start of four questions: 70% of what denominator — all inbound calls, or only the call types routed to the agent? What's the CSAT split between contained and escalated calls? What's the repeat-contact rate on the contained ones? And how many of the "contained" calls actually resolved versus dead-ended? A number without those four is a headline, not a metric.

The trend matters more than the level

A containment rate rising over months with flat or improving contained-call CSAT is real progress — the agent is genuinely handling more. A containment rate rising with falling contained-call CSAT is the agent being pushed to hold calls it shouldn't. Watch the two lines together; the second one is the honesty check on the first.

Where this stops being right

• A young deployment will show a low containment rate that rises as intent coverage grows — don't over-read the early figure.
• A simple status-only line can legitimately hit a high rate; the gaming risk is lower when the calls are all easy.
• Chasing containment past the point where contained CSAT drops is optimizing the wrong thing.

FAQ

What's a good containment rate? The lowest one at which contained-call satisfaction and repeat-contact stay where you want them. A high number with unhappy contained callers is worse than a lower honest one.

How do I know if my containment number is gamed? Check the escalation-reason mix, the CSAT split, and the repeat-contact rate. If contained calls score badly or come back, the rate is fiction.

Does auto-containing 60% of calls mean a 60% cost saving? No. Subtract the per-minute run cost and the escalation staffing that never goes to zero, and confirm the contained calls actually resolved rather than dead-ended. The saving is on the calls genuinely handled end to end, net of what it costs to handle them.

ISTRALLEN builds voice agents instrumented so containment is read alongside CSAT, repeat contacts, and escalation reasons; see AI for Fintech.