Chatbot containment rate in 2026: why one deployment publishes 81%, 67% and 20%

Your vendor reports a containment rate. Your board hears “the bot solves most of our tickets”. Those are two different claims, and the gap between them is where AI customer service budgets go wrong.

The public record now holds one deployment where all three common numbers were published for the same bot. They do not agree, and the spread is wide.

Kogan.com, the Australian online retailer, says 81% of its AI-handled conversations “are not escalated to humans at all” and that it resolves 67% of them [source]. Salesforce’s own case study on the same deployment puts the “true resolution rate” at “approximately 20%” ([source]).

What chatbot containment rate means

Chatbot containment rate is the share of conversations a bot finishes without handing the customer to a human agent.

It measures where a conversation ended. It does not measure whether the customer got an answer, and a customer who gives up and closes the window counts as contained.

Is containment the same as resolution?

No. Containment counts conversations that stayed with the bot; resolution counts problems that were actually solved, and the two can sit far apart on the same deployment.

Kogan’s figures show the ladder. Not escalated: 81%. Resolved: 67%, up from the original bot’s 6.2% [source]. Truly resolved: about 20%, up from 6.2% at launch in March 2026 ([source]).

Salesforce defines “true resolution” as “a custom metric measuring whether a customer’s question was fully answered without a follow-up contact” [source]. That is the only one of the three with a published definition. Neither source reconciles the 67% with the 20%.

The satisfaction figure moves too. Kogan’s CTO said scores are “seven percentage points higher” for bot users [source]. Salesforce says agent-handled CSAT is “within four percentage points of human-handled ones” ([source]). Higher and slightly lower cannot both describe one comparison.

Is there a chatbot containment rate benchmark?

There is no chatbot containment rate benchmark that compares like with like, because every company sets its own numerator and its own denominator.

The denominator is the quiet variable. Nordea’s vendor, boost.ai, claims a 91% “in-scope resolution rate” for private customers in Nordea Sweden, where in-scope leaves out chats outside the bot’s topics [source]. Nordea’s own CIO deck says “~78% handled” across 5.2 million chats ([source]). TIN’s Nordea file is capped below green because “handled” is never defined.

A benchmark table that lines up 91%, 81% and 67% from different companies is lining up different metrics. The honest comparison is a company against its own earlier self, on a definition it published.

The proof

Three TIN case files carry the rest of this argument. Each was checked against its primary record before it went live, and each states its own limits on the page.

Klarna’s AI assistant is the containment headline everyone quotes: two-thirds of customer-service chats in its first month, 2.3 million conversations [source]. The same release reported a 25% drop in repeat inquiries ([source]), which is the closer cousin of a true resolution measure. In May 2025 the CEO said the focus on cost had produced “lower quality”, and Klarna recruited humans again [source].

Decagon at Notion reports a 2x increase in deflection and a 3.4% ask-for-human rate [source]. Every figure is vendor-published, and “2x” is relative: the case study gives no starting rate, so the deflection level itself is not on the record.

Octopus Energy’s Magic Ink is the counterexample that breaks the metric. Its CEO said the tool was answering 34% of all customer queries by April 2023 [source]. Yet a human reviews and sends every reply ([source]), so under the containment definition above, none of those replies counts as contained. Real workload absorbed, zero containment: the metric cannot see it.

For how the Agentforce figures hold up on their own, see does Agentforce work.

Which AI customer service metrics should you ask for?

The AI customer service metrics worth asking for are the ones that fail when the customer comes back: resolution without follow-up contact, repeat-contact rate, and the denominator behind each.

Kogan figures: Marketing-Interactive and Salesforce [source] [source]. Notion figure: Decagon ([source]).

Ask three questions of any figure. What is the denominator, all chats or in-scope chats? What happens to a customer who abandons? Does a repeat contact within a week undo the win?

Bottom line

The obvious reading is that containment is a weak proxy and resolution is the real number. The record says something sharper: the vendor picks which number is the headline, and it picks the highest rung of the ladder. Kogan’s own vendor printed the bottom rung in the body of its case study, and that roughly 20% is the most useful number in this whole post. The transferable rule: when a metric’s owner also writes its definition, ask for the strictest version they track internally, because a team that measures true resolution has already told you what it trusts.