Containment rate is the share of conversations your AI agent or bot handles without escalating to a human. It is the most quoted measure of an automated channel because it maps directly onto cost, and it is also one of the easiest to misread, because it counts the absence of a human rather than the presence of a resolution. A customer who gives up and closes the window is counted as contained, exactly like one whose problem was solved. Measuring it honestly means pairing the containment number with a resolution check that reads what the AI agent actually said, so you can tell real containment from abandonment wearing the same badge.
In short
- Containment rate = conversations completed without human escalation, divided by total conversations the automation handled.
- On its own it measures deflection, not resolution. Abandonments, wrong answers, and dead ends all raise the number.
- The fastest way to inflate containment is to hide the path to a human, which is the one thing that most damages the customer relationship.
- The fix is a paired metric: of the conversations you contained, what share genuinely resolved the customer’s issue?
- Producing that share means scoring the conversations, not the routing log, against a rubric that checks whether the answer was correct, complete, and appropriate.
- Score every contained conversation rather than a sample, because false containment hides in the ones nobody reads.
- Because Kaizo does not sell its own AI agents, it can grade those conversations and report true containment against abandonment with nothing to protect.
What containment rate actually measures
Containment rate is the percentage of conversations your automated channel finishes without handing off to a person. The calculation is simple: take every conversation that entered the AI agent over a period, count the ones that ended without an escalation, and divide.
Containment rate = conversations ended without human escalation / total conversations handled by the automation.
It became the headline metric for bots and AI agents for a good reason: it is the number that maps most directly onto cost. A contained conversation carries no agent handling time, so a higher containment rate reads as a lower cost to serve. The problem is not the arithmetic. It is what the arithmetic quietly assumes.
The word doing the damage is “resolved”
Everyone describes containment as conversations the bot “resolved” without a human. But almost no containment measurement checks for resolution. It checks for the absence of a human, and treats that absence as success. Those two things come apart constantly.
A conversation is counted as contained when the customer got a correct answer and left satisfied. It is counted exactly the same way when the customer could not find the escalation button and gave up, when they abandoned the chat halfway through, when the bot gave a confident wrong answer they did not catch, or when it answered a question they were not asking and they closed the window. All four are failures. All four improve the containment rate.
The perverse incentive
Once containment is the target, the cheapest way to move it is to make escalation harder. A bot that buries the path to a human posts a better number than one that offers help the instant it is stuck. So the metric, read alone, rewards precisely the behavior that most frustrates customers. That is not an argument against measuring containment. It is an argument against ever reading it by itself.
The paired metric: containment plus resolution
The fix is not a better containment formula. It is a second number read alongside it. Containment tells you how many conversations stayed away from a human. Resolution tells you how many of those actually solved the customer’s problem. You need both, and the gap between them is the most useful thing on the dashboard.
| Outcome | Counts as contained? | Counts as resolved? | What it really is |
|---|---|---|---|
| Customer got a correct answer and left | Yes | Yes | True containment |
| Customer abandoned the chat mid-way | Yes | No | False containment |
| Bot gave a wrong answer, customer left unaware | Yes | No | False containment, and a trust risk |
| Customer could not reach a human and gave up | Yes | No | False containment caused by design |
| Bot escalated cleanly to a human who resolved it | No | Yes | A working safety net, not a failure |
How to measure resolution honestly
Resolution cannot be read off the routing log, because the routing log only knows whether a human got involved. It has to come from the conversation itself. That is a quality-assurance task, and it is the same task you already run on human agents, pointed at the bot.
Score the conversation, not the outcome flag
Take a contained conversation and evaluate what the AI agent said against a rubric: was the answer factually correct, was it complete, was it appropriate to what the customer actually asked, and did the conversation end with the issue resolved rather than abandoned. Writing a rubric an AI can score covers how to phrase those criteria so they are checkable rather than vague.
Score all of it, not a sample
False containment hides in the conversations nobody reads. A 2% sample of contained conversations will show you a handful of clean resolutions and tell you almost nothing about the abandonment rate in the other 98%. Scoring 100% of conversations is what turns resolution from an anecdote into a rate you can trust. At UiPath, Kaizo automated 100% of QA with 200% ROI and an 8% lift in quality score, which is the kind of coverage that makes a resolution number real.
Watch the escalation path, not just the escalation count
If containment is rising while resolution is flat or falling, check whether the path to a human got harder to find. A containment gain that comes from hidden escalation is a customer-experience loss disguised as an efficiency win.
The neutrality problem in who reports the number
There is a structural reason containment rates tend to look good in vendor dashboards. The platform reporting the number is often the same platform selling the AI agent whose containment is being measured. Containment is how those agents are sold, so the incentive runs toward the flattering reading of it, and the resolution check that would puncture it is the one measurement that never quite ships.
Kaizo does not sell its own AI agents. It scores the conversations they produce, whoever built them, which means it has no number of its own to protect when it separates true containment from abandonment. Every score traces back to the specific lines in the transcript that produced it, so a resolution rate is something you can audit rather than take on faith. That is the difference between a metric marketed to you and a metric you can defend to your own leadership. Our guide to QA for AI agents and chatbots covers the full scoring approach, and measuring AI agent performance puts containment in the context of the other signals that matter.
Common mistakes with containment rate
- Reading it alone: containment without a resolution check tells you cost avoided, not jobs done, and rewards hiding the escalation path.
- Counting abandonment as success: a customer who gave up is not a customer you served. Separate abandonments out explicitly.
- Sampling instead of covering: false containment lives in the unread conversations, so a small sample systematically flatters the number.
- Trusting a self-reported figure: if the tool that built the bot also grades it, ask to see the evidence behind each resolved flag.
- Optimizing containment as a target in itself: once it is the goal rather than a signal, the cheapest way to move it is to make help harder to reach.
Frequently asked questions
What is containment rate?
Containment rate is the share of conversations an AI agent or bot handles without escalating to a human. It is calculated as conversations ended without human escalation divided by total conversations the automation handled. It is widely used because it maps closely onto cost, but on its own it measures the absence of a human rather than whether the customer’s problem was solved.
What is a good containment rate?
There is no benchmark worth chasing, because containment on its own does not tell you whether customers were helped. A high number produced by hiding the escalation path is worse than a lower one where every contained conversation genuinely resolved. Judge containment against a resolution check on the conversations themselves, not against an industry average.
How is containment rate different from resolution rate?
Containment counts conversations that ended without a human. Resolution counts conversations where the customer’s issue was actually solved. A customer who abandons the chat raises containment but not resolution. Reading the two together, and watching the gap, is the honest way to measure an automated channel.
Why can a high containment rate be a bad sign?
Because the fastest way to raise containment is to make escalation harder. A bot that buries the path to a human contains more conversations and resolves fewer of them. Without a resolution check alongside it, a rising containment rate can mean customers are giving up rather than being helped.
How do you measure whether contained conversations were resolved?
By scoring the conversations themselves against a rubric, not by reading the routing log. Evaluate whether the AI agent’s answer was correct, complete, and appropriate, and whether the issue ended resolved rather than abandoned. Doing this across every contained conversation rather than a sample is what makes the resolution rate trustworthy.
Related terms
Find out how much of your containment is real
Bring the conversations your AI agent contained last month. We will score them against your own rubric across every conversation, not a sample, and show you how many genuinely resolved the customer’s issue and how many were abandonments counted as wins. Because Kaizo does not sell its own AI agents, every resolution score traces back to the evidence in the transcript.
