Containment rate is the share of bot conversations that end without a human taking over. That sentence sounds simple. The fights start when you define “without a human”, and when you treat containment as if it were the same as deflection, resolution, or customer success.
Stripped of deck language: containment is a conversation-level metric about routing. It does not prove the answer was correct. It does not prove a ticket would have existed. It proves the visitor left the automated channel without crossing whatever line your product draws as escalation.
Different products draw that line differently: form submit, queue join, agent message. When you compare vendors, compare definitions, not percentages.
The numerator and denominator, plainly
Numerator: conversations the bot handled start to finish by its own rules. Denominator: all conversations that started in the widget. Excluded or included depending on vendor: sessions that ended in abandon with no answer, sessions handed to a form, sessions where the human never replied in live chat.
Containment vs resolution vs deflection
| Metric | Claim | Common confusion |
|---|---|---|
| Containment | No human took the thread | Sounds like the customer is happy |
| Resolution / answer rate | Bot sent an answer | Sounds like containment |
| Deflection | Ticket volume fell vs baseline | Sounds like containment at scale |
You can contain a conversation with a wrong answer. You can resolve with an answer the customer immediately escalates. You can deflect tickets while containment looks flat because visitors skip the widget. Read resolution vs deflection before you put containment on a board slide — otherwise you will mix three stories into one green arrow.
Report containment as a secondary column, never the headline. Lead with sampled correctness or ticket volume vs baseline. Containment explains routing; it does not prove value.
When high containment is good
- Sampled correctness on contained threads is stable or improving.
- Refusal list themes shrink as content ships — containment rose because pages exist.
- Satisfaction on bot-only threads is not trailing human-handled ones by a cliff.
- Escalations that do happen arrive in one turn with full transcript.
That pattern means containment is tracking real self-service. Visitors got what they needed without paying human time.
When high containment is a warning
High containment paired with rising complaints is the classic failure mode. So is high containment with lengthening bot sessions — the customer is trapped, not helped.
- Handoff friction inflates containment by design.
- Wrong answers “contain” until the chargeback arrives.
- Forms count as containment but recreate ticket load manually.
- Visitors learn to bypass the widget; containment rises as usage falls.
Falling tickets can mean attrition, not automation success. Containment without ticket context is half a chart.
How to report containment without fooling yourself
“Containment is a routing metric. Put correctness and satisfaction beside it or do not show it to leadership.”
- Define escalation precisely — live chat pickup, ticket created, human message sent.
- Show containment with sampled audit results, even a small weekly sample.
- Track median turns and time before escalation; containment with long sessions is friction.
- Never present containment as dollars saved without a ticket baseline.
Used this way, containment explains where automation carries load. Used alone, it is a vendor metric dressed for buyer meetings. You are allowed to track it. Just do not let it become the only story.
When leadership asks for one number, give them ticket volume vs baseline with a footnote on containment — not the reverse. The question they mean to ask is “did we reduce real work without making customers give up?” Containment alone cannot answer that.
If your product lets you tag conversations that ended in abandon after a wrong answer, chart that separately from clean containment. It is the early attrition signal hiding inside a good headline number.
