Support metrics glossary
Containment rate. The bot metric everyone quotes.
Containment rate is the share of bot conversations that never reach a human agent. Simple to say, easy to inflate, and the first number a bot vendor will quote you.
Last updated 14 August 2026
What containment rate means
Containment rate is the percentage of bot conversations that end without escalating to a human. The formula: contained conversations divided by total bot conversations, times 100. A bot that handles 800 conversations and closes 560 without a handoff has a rate of 70%.
A conversation is contained when the bot closes it and no agent ever touches it. That is the whole definition. What the definition never asks is whether the customer got an answer.
A bot metric, not a queue metric
Containment measures the bot: of the conversations the bot saw, how many stayed inside it. Deflection measures the whole queue: of every issue customers had, how many were resolved without an agent, by the bot, the help center, anything.
The denominators differ, so the numbers answer different questions. Route fewer conversation types to the bot and containment rises while deflection falls: the bot now only sees the easy work. The deflection rate page covers the queue side and the ways that number gets gamed.
| Metric | Formula | Worked example | What it tells you |
|---|---|---|---|
| Containment rate | Contained bot conversations / total bot conversations | 560 contained of 800 bot conversations = 70% | How often the bot finishes what it starts |
| Deflection rate | Issues resolved with no agent / total issues | 900 resolved of 2,000 total issues = 45% | How much of the whole queue never needs a person |
The lazy version counts abandonment as success
Here is the gimmick. A customer asks, the bot answers badly, the customer closes the chat. They never escalate; there is no button, or they are out of patience. The transcript just stops. Measured by silence, that is contained. The customer was not helped. They left.
Hence the most flattering number in a bot vendor's deck. Count a conversation as contained only when an outcome confirms the customer was helped: a completed action, a helpful vote, or no return with the same issue inside a re-contact window of 24 to 72 hours.
How to measure it honestly
Three rules. Define contained by an outcome, never by the absence of an escalation. Set a re-contact window, and a customer who returns with the same problem inside it counts as not contained. Segment by intent: password resets at 90% will drag a 30% rate on billing disputes up to a respectable blended average.
What a good number looks like
There is no honest universal benchmark, and anyone who quotes one without asking what you route to your bot is selling. A bot that only sees order status questions should contain almost all of them; our WISMO page shows the math on that category. A bot that also sees complaints will sit lower and can still be excellent.
Judge the trend month over month, next to resolution or satisfaction. A number that climbs while satisfaction falls is not a bot getting better. It is a bot learning to stonewall.
One boundary worth drawing. The metric says nothing about agent assist, where AI drafts and suggests for a human who was always in the loop. Nothing to contain. Different category, different scoreboard.
Containment rate questions, answered first
What is containment rate in customer support?
The percentage of bot conversations that end without escalating to a human agent. Divide contained conversations by total bot conversations and multiply by 100. A bot that closes 560 of 800 conversations without a handoff has a containment rate of 70%.
How is containment different from deflection?
Containment is a bot metric: of the conversations the bot handled, how many never escalated. Deflection rate is a queue metric: of every issue customers had, how many were resolved without an agent by any self-service path. The denominators differ, so the two can move in opposite directions.
How do you calculate containment rate?
Contained conversations divided by total bot conversations, times 100. The formula is the easy part. The hard part is what counts as contained: count it only when an outcome confirms the customer was helped, not when the transcript goes quiet. Silence is abandonment.
What is a good containment rate for a chatbot?
There is no honest universal number; it depends on what you route to the bot. A narrow bot handling simple intents should contain almost everything. A broad one seeing complaints and edge cases will sit lower and can still perform well. Track your own trend, next to customer satisfaction.
Why can a high containment rate be a bad sign?
Because the laziest version counts abandonment as success. A customer who gives up and closes the chat never escalates, so a dashboard that measures silence calls it contained. If the number rises while satisfaction or repeat contact worsens, the bot is not resolving more. It is wearing people down.
Does containment rate apply to agent assist?
No. Agent assist works inside conversations a human is already handling, drafting replies and suggesting answers, so there is nothing to contain. It applies to customer-facing bots that own a conversation end to end. Mixing the two produces numbers that mean nothing.
The queue side of the story
Containment tells you what the bot did. Deflection tells you what customers finished without a person. Read both before you believe either.
The self-service math, worked through on ecommerce's biggest ticket category.