Containment rate is the wrong goal for a support agent

Containment rate counts how many conversations an agent kept away from people. Made into a goal, it rewards exactly the wrong behaviour: holding on to cases that needed a person. The cases that should have escalated disappear into the contained figure, and the metric looks best at the moment the agent is failing the customers who most needed help.

What does containment actually measure?

Containment rate is the share of conversations that start with an agent and end without a person being involved. It is a measure of where a conversation finished, not of how it finished. A customer whose problem was solved counts as contained. So does a customer who gave up, one who was given a wrong answer and believed it, and one who left to phone instead.

That is why containment is a poor proxy for good service. It treats every conversation that did not reach a person as a success, regardless of what happened to the customer. As a diagnostic it has some uses. As a target it is actively harmful.

How does a containment target change agent behaviour?

Targets shape design. Once containment is the goal, the people building and tuning the agent face steady pressure to keep conversations inside it. That pressure shows up in predictable ways:

• Escalation triggers get narrower. Conditions that used to send a customer to a person are tightened or removed.
• The route to a person gets harder. "Talk to a person" is buried, needs to be asked for several times, or is offered only after the agent has tried again.
• The agent tries one more time. Briefs grow instructions to attempt another answer before handing over.
• Conversations end ambiguously. The agent closes with a link or a suggestion, and a conversation the customer simply abandons counts as contained.

None of these needs bad intent. Each is a reasonable-sounding response to a number that someone is asked to improve.

Which cases hide inside a high containment rate?

The ones that matter most. A high containment rate can conceal:

1. Customers in vulnerable circumstances who needed a person early and did not get one.
2. Complaints the agent answered as routine questions, which never entered the complaints process.
3. Wrong answers the customer accepted, discovered later through a fee, a missed deadline or a refused claim.
4. Customers who abandoned the chat and called, so the contact cost moved rather than fell.
5. Requests that needed authority the agent lacked, answered with a vague holding reply.

Each of these is a case where the right outcome was escalation. The agent's escalation behaviour should be judged on whether it hands these over. A containment target pushes the other way. The question of how high an AI support agent's escalation rate should be has no fixed answer, but it is certainly not "as low as possible".

Who carries the cost of a wrongly contained case?

The customer first, then the firm. A customer who received wrong information and acted on it may be out of pocket. The firm may then face a complaint, redress and, in regulated markets, questions about whether it treated the customer fairly. The analysis of who is responsible when a support agent gives a customer wrong information is clear that the firm does not escape responsibility because an agent gave the answer.

Containment counts none of this. It records the conversation as a saving at the moment it ends and never revisits it.

What should replace it?

Measures that reward the agent for doing the right thing, including handing over when it should. Three are especially useful:

• Correct escalation: of conversations reviewed, the share where the agent escalated when it should have and did not when it should not.
• Resolution without repeat contact: conversations where the customer did not come back about the same issue within a set period, on any channel.
• Handover quality: when the agent hands over, whether the person could pick up without asking the customer to repeat themselves.

These are harder to collect than containment, which is precisely why containment became popular. A fuller set is discussed in measuring a support agent beyond resolution rate.

Is containment useful at all?

As a diagnostic, yes, read alongside the measures above. A sudden rise in containment with no change in repeat contacts may mean the agent has improved. A sudden rise with a jump in repeat contacts or complaints means the agent is holding on to cases it should release. Containment tells you where to look; it should never tell you whether things are good. Used that way, it prompts questions rather than answering them: why did it move, which request types drove the change, and did those customers come back?

The same caution applies to its close cousin, deflection rate, which counts customers sent to self-service content. Both measure where the customer went, not whether they were helped.

What should you do differently?

Take containment off any target or incentive linked to the support agent. Keep it on a dashboard as one signal among several. Set targets instead on correct escalation and on resolution without repeat contact. Then review a sample of contained conversations each week, looking specifically for the cases that should have reached a person. If you find them, the containment figure was hiding them. This is the stance behind the AI support agent that knows when to say no.