How is it different from deflection?

Deflection asks whether an enquiry reached a person at all; containment asks whether a conversation that began in the automation stayed there. They answer different questions about the same week, and can move in opposite directions without either being wrong.

What second number should we pair it with?

Anything the same team cannot quietly move: repeat contacts within a few days, satisfaction measured on contained conversations, or whether the customer took the action they came to take. The pairing is what makes containment readable at all, because on its own it cannot distinguish success from surrender.

Is it a bad measure then?

No, it is a good measure of the thing it measures. Knowing what share of volume never reaches the queue is how a team gets sized, and that is genuinely useful. The error is reading it as a statement about customers rather than about workload.

Glossary

Containment rate

The share of conversations an automated system saw through to the end on its own, whether or not the person on the other side got what they came for.

In plain terms

How often the software kept the conversation to itself. It is a measure of the system's reach, not of whether anybody was helped, so a customer who gave up in frustration is counted exactly the same as one who left satisfied.

Why it matters

Because it is the number an operations team can most easily improve without improving anything. Making it harder to reach a person raises containment immediately, and every part of that change looks like progress in the report while the customers experience the opposite.

How it works

It measures where a conversation ended, not how. Staying inside the automation is the whole test, so the same result covers a question answered well, an answer that was wrong, and somebody who stopped trying.

It is the most directly gameable number in support. Hiding the route to a person raises it within a day, requires no improvement to anything, and produces a chart that points the right way for as long as anybody is looking only at this.

Reading it requires a second number that is not under the same control. Repeat contacts, satisfaction on contained conversations or what customers did next all work, and the point of choosing one is that it moves the other way when containment is bought rather than earned.

It is genuinely useful for capacity planning, which is what it was for. Knowing what share of volume never reaches the queue is how a team is sized, and that use does not depend on the conversations having gone well.

It differs from deflection in what it counts rather than in how well it does it. Deflection asks whether an enquiry reached a person at all; containment asks whether a conversation that started in the automation stayed there, so the two can move in different directions on the same week.

The failure mode it hides is the frustrated customer who says nothing. They do not escalate, they do not complain, and they do not come back, so the only trace they leave is in numbers nobody in support is looking at.

Three endings, one number

The second row is the one worth dwelling on, because it is the case nobody is looking for. A confidently wrong answer produces a contained conversation, a satisfied-looking ending and a customer who acts on bad information somewhere else entirely, which means the cost lands in a different part of the business from the one reporting the success. Refunds, complaints and churn all arrive weeks later with no obvious connection to a support metric that was climbing at the time. The third row is more familiar and easier to accept, and the honest thing to say about it is that it is not a measurement problem: containment is doing exactly what it says. The problem is that a number describing workload gets read in a room where people are trying to decide whether customers are being served, and it has no opinion at all about that. Pairing it with one measure outside the same team's control costs almost nothing and is the difference between a metric and a story.

Seen in the wild

A rate that jumped after the route to a person was moved further down the menu.
Sierra
Contained conversations that end without the customer ever confirming anything was solved.
Tidio Lyro
An automated flow finishing a case itself with nobody checking the outcome.
Relevance AI

Common misconceptions

People assume

A contained conversation is a resolved one.

In fact

Containment records where the conversation ended, not whether the person got what they needed. Somebody who gave up halfway is contained, and in most reporting looks identical to somebody who left satisfied.

People assume

Rising containment means the system is getting better.

In fact

It also rises when reaching a person becomes harder, which takes a configuration change and no improvement at all. Without a second number that is not under the same control, you cannot tell the two apart.

Questions

How is it different from deflection?: Deflection asks whether an enquiry reached a person at all; containment asks whether a conversation that began in the automation stayed there. They answer different questions about the same week, and can move in opposite directions without either being wrong.
What second number should we pair it with?: Anything the same team cannot quietly move: repeat contacts within a few days, satisfaction measured on contained conversations, or whether the customer took the action they came to take. The pairing is what makes containment readable at all, because on its own it cannot distinguish success from surrender.
Is it a bad measure then?: No, it is a good measure of the thing it measures. Knowing what share of volume never reaches the queue is how a team gets sized, and that is genuinely useful. The error is reading it as a statement about customers rather than about workload.