A good containment rate is one that your recontact rate and customer satisfaction agree with. There is no independent industry benchmark for it, so a high rate that hides customers giving up is worse than a modest one where the problems actually get fixed.
That is not the answer most people want, so here is the longer version.
What is the difference between containment, resolution and deflection?
They get used as if they mean the same thing. They measure three different events.
Containment rate is the share of conversations the agent finished without handing to a person. It records what did not happen.
Resolution rate is the share where the customer's problem was actually fixed. You measure it by a confirmed outcome in your systems, by no repeat contact on the same issue within a set window, or by the customer saying so.
Deflection rate is usually the share of contacts that never reached your team at all, because a bot, a help article or a self-service page caught them first. It is the widest and softest of the three, because it counts people who left as well as people who were helped.
Resolution sits inside containment. Everything the agent resolves is contained, but not everything contained is resolved. The gap between the two is the number worth knowing.
Is there a benchmark for containment rate?
Not a credible one. Published containment figures almost all come from vendors describing their own customers, using their own definitions, and they disagree so widely that no range is worth quoting. A narrow bot that tracks orders and an agent that takes billing disputes are not measuring the same thing, even when both call it containment. The difference between a chatbot and an AI agent shows up here more than anywhere.
The most useful independent number points the other way. In August 2024 Gartner published a survey of 5,728 customers that found only 14% of service issues were fully resolved in self-service, and only 36% of issues the customers themselves described as very simple. That covers self-service broadly, not AI agents specifically, but it shows how far "the customer stayed in the channel" can sit from "the customer got it sorted".
So compare yourself with yourself: your rate by intent, month on month, with the other numbers beside it.
Why can a high containment rate hide failure?
Because the easiest way to raise containment is to make it hard to leave.
The usual ways it happens:
- No visible route to a person. Hide the handover and containment climbs. The customer phones instead, and that call never gets joined back to the chat.
- Sessions that time out. Someone closes the window halfway through. Unless the build marks that separately, it counts as contained.
- Answers that are correct but do not finish the job. The agent explains how to restart the router, the router still does not work, and the customer emails the next morning. Contained today, a repeat contact on a different channel tomorrow.
- Loops. The agent asks the same clarifying question twice and the customer gives up. Contained.
The opposite also happens. A well-timed handover can be the best outcome a conversation has. An agent that recognises a cancellation or a complaint early and passes it to the right person, with the full history, lowers containment and gets the customer a better result. Scoring that as a failure trains the build the wrong way. There is more on that in what happens when an AI agent does not know.
What should you measure instead?
Keep containment, but never read it alone. Satisfaction and recontact are covered in how do I know if my AI agent is working. For containment itself, these are the ones we would set up on day one:
- Contained and not back. Containment minus anyone who contacted you again, on any channel, about the same thing within a window you pick and keep fixed. This needs the chat joined to your CRM or helpdesk by customer, which is exactly the integration work that gets skipped to hit a launch date.
- Containment by intent. One rate for "where is my engineer", another for "my bill is wrong". An average across both hides that one is healthy and the other is a problem.
- Abandonment as its own outcome. Closed, timed out or went silent mid-flow. If your reporting does not separate it from contained, fix that before you report anything else.
- Outcome in the system. Did the ticket, booking, payment or account change actually happen? That is the strongest evidence of resolution, because it does not depend on anyone filling in a survey.
- Handover quality. When the agent did pass to a person, did that person have to ask the customer to start again?
Then read transcripts. A weekly sample of contained conversations, picked at random rather than from the ones that look good, will find the failures long before any aggregate does.
So what number should you aim for?
Set a target per intent once you have a month of honest data, and set it on contained and not back rather than raw containment. For simple account questions with real integration behind them, high rates are realistic. For complaints, disputes and anything emotional, a lower containment rate with fast, well-briefed handovers is often the right design.
When someone quotes you a containment rate, ask how abandonment is counted, what the recontact window is, and which intents sit behind it. If they cannot answer, it is a sales figure. Turning the honest version into money is a separate step, covered in what ROI to expect from conversational AI.
If you want a second view on how your current agent is measured, we are happy to look at it with you. Or try our agent and see where it chooses to hand over.
