Containment rate is the share of conversations an AI agent or bot finishes without a human handoff. It counts conversations that never reached a person, so a customer who gave up counts the same as one whose problem was solved.
In short
- Containment rate measures deflection, not resolution.
- A high containment rate can hide a bad customer experience.
- The honest way to read it is next to a check on how many contained conversations were actually resolved.
- That check is a quality-assurance task: read the conversation and score it against a rubric.
How containment rate is calculated
Containment rate is calculated by dividing the conversations the automation finished without escalating to a human by all conversations that entered the automated channel in the same period. The result is a percentage. It is simple to pull from a routing log, which is why it is often the first number a bot deployment reports.
Containment rate = conversations resolved without human escalation / total conversations handled by the automation.
So if 1,000 conversations reach an AI agent in a week and 650 of them never touch a human, containment is 65%. The appeal is obvious: it maps almost directly onto cost, because a contained conversation carries no agent handling time. That is also why it tends to become the number a bot deployment is judged on, and why it deserves more scrutiny than it usually gets.
Why containment rate can mislead
Containment rate can mislead because most measurements do not check for resolution at all. They check for the absence of a human. A customer who gave up, abandoned the chat or left with a wrong answer is counted as contained, exactly like a customer whose problem was solved. The gap between those two outcomes is where the metric goes wrong.
The weakness is in the word “resolved” in that formula. A conversation counts as contained if the customer got what they needed and left satisfied. It also counts as contained if the customer could not find an escalation path and gave up, if they abandoned the chat halfway through, if they got a wrong answer and did not realize it, or if the bot answered a question they were not asking and they closed the window. Every one of those is a failure, and every one of them improves the containment rate.
This creates a perverse incentive. The fastest way to raise containment is to make escalation harder, and a bot that hides the path to a human will post a better number than one that offers help the moment it is stuck. The metric rewards exactly the behavior that most damages the customer relationship. The same blind spot affects deflection rate, which counts conversations kept away from agents without asking whether they ended well.
Containment rate vs resolution rate
Containment rate counts conversations that ended without a human, while resolution rate counts conversations where the customer’s issue was actually solved. The first is a proxy for cost avoided, the second for the job done. A customer who gives up is a success for containment and a failure for resolution, which is why the two numbers need to be read together.
| Containment rate | Resolution rate | |
|---|---|---|
| What it counts | Conversations that ended without a human | Conversations where the customer’s issue was actually solved |
| A customer who gives up | Counted as a success | Counted as a failure |
| What it is a proxy for | Cost avoided | Job done |
| How to measure it | Read the routing log | Read the conversation |
How to use containment rate honestly
Use containment rate as an efficiency signal, never on its own. Pair it with a resolution check: of the conversations you contained, how many actually resolved the customer’s problem, and how many were abandonments or wrong answers wearing a containment badge. Report both numbers side by side so a rise in containment cannot hide a fall in resolution.
Answering that means looking at what the AI agent said and whether it was correct, complete, and appropriate, which is a quality-assurance question, not an analytics one. It is the same discipline you would apply to a human agent, applied to a bot: read the interaction, score it against a rubric, and trace the score back to the evidence. Our guide to measuring containment rate without fooling yourself walks through how to do that, and how to QA AI agents and chatbots covers the scoring method underneath it.
A worked example of the rubric: add one criterion to the scorecard you already use for AI-agent conversations, “Did the conversation end with the customer’s issue resolved?”, with three outcomes: resolved, abandoned, or wrong answer. Every contained conversation gets that score, and each score points to the line in the conversation that justifies it. Containment then splits into the part that saved cost and the part that only moved the problem somewhere else.
Where contained conversations go wrong: the handoff
Contained conversations most often go wrong at the point where the AI agent should have handed over to a person and did not. A late handoff, a missing escalation path or a handoff that loses the context all show up as containment or as a frustrated customer, never as an obvious error in the routing log.
That is why checking containment and checking the handoff are the same job. Our guide to scoring the AI-to-human handoff covers what to score in both directions of the transition, and AI customer service metrics shows which numbers to report next to containment.
Kaizo scores 100% of human and AI-agent conversations against your own scorecard, and every score links to the evidence in the conversation, so contained conversations can be checked for true resolution against abandonment. At UiPath, Kaizo automated 100% of QA with 200% ROI and an 8% lift in quality score.
See it on your own conversations. Kaizo scores 100% of your support conversations and turns the findings into coaching. Let’s talk.
When a containment check is not for you
A resolution check on contained conversations is not worth building when automation handles only a handful of conversations a week: at that volume, someone can read them all. It is also premature if you do not yet have a scorecard for your human agents, because the AI agent should be judged against the same standard. Start there, then extend it to the bot.
Frequently asked questions
What is a good containment rate?
There is no single benchmark worth quoting, because containment rate 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 number where every contained conversation was genuinely resolved. Judge containment against a resolution check rather than against an industry average.
What is the difference between containment rate and deflection rate?
They are often used interchangeably. Both describe conversations kept away from a human agent. Deflection sometimes refers specifically to steering a customer to self-service before a conversation starts, while containment usually refers to a conversation the automation handles to completion. Both share the same blind spot: neither, by default, checks whether the customer’s problem was solved.
Does a high containment rate mean the AI agent is working well?
Not by itself. Containment counts the absence of a human, not the presence of a resolution, so a customer who abandons the chat is counted as contained. To know whether the AI agent is working, pair containment with a resolution check that reads the actual conversations and scores whether each one solved the issue.
How do you measure whether a contained conversation was actually resolved?
By scoring the conversation itself, not the routing outcome. That means evaluating what the AI agent said against a rubric: was the answer correct, complete, and appropriate, and did the customer’s issue end resolved rather than abandoned. It is the same quality-assurance approach used on human agents, applied to automated conversations.
What does containment rate mean in a call center?
In a call center, containment rate usually describes the share of contacts that an automated system, such as a phone menu (IVR) or a chatbot, completes without transferring the customer to a live agent. The calculation and the blind spot are the same: a customer who hangs up or gives up counts as contained.
Is a low containment rate always a bad sign?
No. A lower containment rate can mean the AI agent hands over quickly when it is stuck, which is often the better customer experience. What matters is whether the contained conversations were resolved and whether the escalated ones reached a person in time with the context intact.
Should AI-agent conversations use the same scorecard as human agents?
Mostly, yes. Accuracy, tone and resolution apply to both, so one standard makes the numbers comparable. Bots usually need a few extra criteria, such as staying within policy limits and handing over at the right moment, which our guide to one scorecard for human and AI agents covers.
