Skip to content
Chatbotscape
Verified
Chatbot Containment Rate· Customer-service metric
Containment rate is the share of conversations a chatbot handles from start to finish without ever passing the user to a human. It is the self-service metric: a contained chat is one the bot 'kept' the whole way through. The number is only honest when 'kept' also means 'resolved' — a user who gave up and walked away never escalated either, so a naive containment count quietly rewards a bot that is simply hard to escape. Read it with a satisfaction signal and an abandonment check, never on its own.
By Chatbotscape Editorial· Methodology· Published 16 June 2026· Updated 16 June 2026

Chatbot Containment Rate — Definition, Formula, and Healthy Ranges (2026)

Quick answer: Containment rate is the percentage of conversations a chatbot carries to completion without handing off to a human agent. It is the cousin of deflection rate and the two are often used interchangeably, but containment is the stricter idea: it is about the bot holding the whole conversation, not just avoiding a single ticket. The trap is that a chat the user abandoned in frustration counts as "contained" unless you actively exclude it, so a high containment rate paired with high abandonment is a warning sign, not a win. Measure it with a resolution or CSAT condition attached, and read it next to the escalation rate it mirrors.

What it is

Containment rate answers one operational question: of all the conversations that came in, how many did the bot finish by itself? A conversation is contained when it reaches an end without the bot triggering a human handoff. For a customer-service chatbot this is the headline self-service number — the one operators point to when they say "the bot handles X% of our volume." The base formula is simple:

Containment rate = (conversations the bot finished without escalation) / (total conversations) × 100%

Because escalation is the only thing that breaks containment, this metric is the mathematical mirror of escalation: in the simplest accounting, containment rate + escalation rate = 100%. That tidiness is also the metric's weakness. The formula counts anything that ended without a handoff as a success, which means it cannot, by itself, tell the difference between a question the bot answered and a customer the bot exhausted.

Containment versus deflection — the distinction operators blur

Containment and deflection are the two terms vendors reach for most, and they are used loosely enough that you should always ask what a given dashboard means by each. The cleanest way to hold them apart: deflection is about avoiding a cost — a ticket that never reached a human — while containment is about the bot owning the whole conversation from first message to resolution. In a single self-contained chat widget the two numbers are nearly identical. They diverge in multi-channel setups, where a "deflected" web chat can still spawn a follow-up email ticket a day later; that conversation was deflected in the moment but not truly contained across the customer's whole journey.

The more important distinction is the one our deflection-versus-containment entry is built around: whether you attach a satisfaction condition. A raw containment count is a volume metric. The version worth reporting adds a quality gate — the bot finished the chat and the user was actually helped — which pulls the honest number down by the same 10-15 percentage points that separate naive deflection from satisfied containment. Whenever you see a containment figure with no satisfaction signal behind it, treat it as a ceiling, not a result.

The abandonment trap

The failure mode unique to containment is abandonment. A user who types one question, gets a weak answer, and closes the tab never escalated — so a naive count files that conversation under "contained." The bot looks like it handled the chat; in reality it lost the customer. This is why a containment rate read in isolation can climb precisely as a bot gets worse at the hard cases, because frustrated users leave rather than fight their way to a human.

The fix is to subtract abandonment from the numerator. A conversation should only count as contained if it reached a genuine resolution event — the user confirmed the answer helped, the flow hit a true end node, no follow-up ticket appeared within a day or two — rather than simply going quiet. Pair the metric with an abandonment check and a CSAT floor and the number starts describing resolution instead of mere silence.

What counts as healthy (2026)

There is no single published containment benchmark for chatbots — vendors define the term on their own terms and rarely separate bot-only conversations from blended ones. The ranges below are editorial working figures, kept deliberately consistent with the containment column in our deflection-versus-containment entry, and they assume the honest definition (finished and resolved, abandonment excluded). Treat them as directional:

Bot architectureHealthy containment rateReading
Rule-based FAQ bot12-18%Narrow scope; contains only the most repetitive, well-documented asks
NLU intent bot (Dialogflow-style)22-32%Handles a defined set of intents; everything else routes out
LLM with RAG, well-tuned35-50%The realistic target band for a modern SMB support bot
Premium products (Intercom Fin, Zendesk AI Agent)45-58%Tightly measured with a resolution gate, which is why the figures read lower than competitors' raw deflection claims

Two cautions move these bands more than the architecture does. First, scope: a bot pointed only at order-status lookups will post a high containment rate because it declined the hard work, not because it is better built — a narrow bot with high containment can be doing less for customers than a broad bot with lower containment. Second, the satisfaction gate: a containment figure quoted without a resolution or CSAT condition is the easy version of the metric and will sit 10-15 points above the number you can defend. A high containment rate next to a low CSAT or a high abandonment rate is the classic signature of a bot that is hard to escape rather than genuinely self-sufficient.

How platforms expose it

Where the number lives depends on the platform class. Support-desk products such as Intercom and Tidio report a resolution-gated containment figure natively — Intercom's "resolution rate" for its Fin agent requires a user-confirmed resolution before a conversation counts, which is why its published numbers read lower than rivals' headline deflection claims and are methodologically tighter, not worse. Flow-first builders like Manychat and SendPulse usually express containment implicitly: every flow that reaches an end node without hitting a "talk to a human" branch is contained, and you assemble the rate yourself from flow analytics and handoff tags. Developer-grade builders such as Botpress let you mark resolution events explicitly in the conversation, which is what makes a clean, abandonment-aware containment rate possible without exporting transcripts.

Whatever the surface, the question to ask a platform is not "what is your containment rate" but "what does a contained conversation have to clear before you count it." A tool that counts any non-escalated chat is handing you the flattering version. A tool that requires a resolution signal — and lets you slice containment by intent so you can see which topics the bot actually finishes — is handing you something you can act on.

  • Chatbot deflection rate — the cost-avoidance sibling; near-identical in a single channel, divergent across a multi-channel journey.
  • Deflection vs containment — the full comparison, and why the satisfaction gate is what separates the honest number from the easy one.
  • Chatbot escalation rate — the mathematical mirror; containment and escalation sum to the whole.
  • Chatbot CSAT — the satisfaction floor that turns a raw containment count into a resolution metric.
  • Human handoff — the event that breaks containment.

FAQ

Is containment rate the same as deflection rate?

Not quite, though the terms are often swapped. Deflection is about avoiding a cost — a ticket that never reached a human — while containment is about the bot owning a conversation end to end. In a single chat widget the two numbers are nearly identical; they diverge across a multi-channel journey, where a chat can be deflected in the moment yet spawn a follow-up email ticket that means it was never truly contained. The bigger practical difference is whether a satisfaction condition is attached, which our deflection-versus-containment entry covers in full.

What is a good chatbot containment rate?

As a directional target, a well-tuned LLM support bot lands around 35-50% when containment is measured honestly — finished and resolved, with abandoned chats excluded. Rule-based bots sit lower (12-18%) because they cover a narrow set of asks. There is no universal benchmark, because the figure depends on the bot's scope and on how strict the resolution gate is. A narrow bot can post a high rate simply by declining hard questions, so read the number next to scope, not in isolation.

Why can a high containment rate be a bad sign?

Because abandonment counts as containment unless you exclude it. A user who gives up and closes the chat never escalated, so a naive count files them as "contained" — which means the rate can rise precisely as the bot gets worse at hard cases and more people quietly leave. A high containment rate next to a high abandonment rate or a low CSAT usually means the bot is hard to escape, not genuinely self-sufficient.

How is containment related to escalation rate?

They are mirror images. In the simplest accounting, every conversation either escalates or it does not, so containment rate + escalation rate = 100%. Reading the two together is the quickest way to sanity-check a containment figure: if containment is high and the escalation rate is near zero, confirm that the missing conversations were resolved rather than abandoned.

Does a low containment rate mean my platform is weak?

Usually not by itself. Containment is driven mostly by operator-owned factors — knowledge coverage, scope, tone, and how the bot is tuned — plus the genuine difficulty of your support mix. The platform matters at the margins: whether it can measure containment with a resolution gate, slice it by intent, and carry context across a handoff. Widen coverage and tighten the failing intents first; the metrics guide shows where containment sits in the full KPI stack.

Sources