Containment rate formula and chatbot benchmarks

To work it out, take the conversations handed to a person away from all the conversations the bot started, divide what’s left by the total and multiply by 100. If your chatbot started 800 conversations in May and passed 120 of them to you, it contained 680, a containment rate of 85%.

Contact centres measure their automated phone menus the same way. Those menus are called IVRs, for interactive voice response, and a 2021 report by the US Postal Service’s inspector general defines their containment rate as “the rate of calls that are not transferred to an agent”. Chatbot tools kept the word, so you’ll also see it called IVR containment rate, call containment rate or chatbot containment rate.

For a founder who answers support alone, a contained conversation is one that didn’t need you, and it only counts if the customer left with an answer. Deacon is an AI customer support agent that answers from your own docs in seconds, tells your customers when it doesn’t know, and has a free plan.

How to calculate containment rate

The usual formula counts every conversation that never reached a person. A stricter one also takes out the conversations the customer left without an answer, which is the split the Postal Service’s inspector general recommended for its phone menu.

Back to May. In 160 of the 800 conversations, the visitor left before getting an answer and never asked for anyone. The usual formula counts them as contained, which gives 85%. Take them out as well and 520 conversations were contained, a rate of 65%. It’s the same month read two ways, 20 points apart.

Which reading you get depends on your tool. These are the rules in four bot platforms’ documentation and in the Postal Service report, read on 23 September 2026.

So a rate from Cognigy and one from Genesys can describe the same month very differently. Before you compare your figure with anyone else’s, find out whether theirs counts the customers who left.

What is a good containment rate?

There’s no standard figure. The Postal Service report says “there are no universal benchmarks for containment, transfer, or abandonment rates”, because every phone menu is different and they’re used across many industries. These are the published figures we found with a named sample, read on 23 September 2026.

In the contact centre surveys, bots finished about 1 in 9 web chats alone in the US and 1 in 5 in the UK. ContactBabel also reports 27% and 33% of chats shared between AI and a person, but it labels those figures two ways, as chats a bot handed to an agent and as chats where AI suggested replies to one, so they can’t be turned into a containment rate. Ada’s 72% is far higher. Its guide counts a conversation as contained when it “never reached a human”, and doesn’t say whose deployments it measured.

Vendor guides often give 70 to 90% as the target, as Botpress’s guide does for enterprises, without a study behind it. The comparison that means something is your own rate, month by month, counted the same way each time.

Containment rate vs escalation, deflection and resolution rate

These rates are often quoted together, and each one counts something different.

Containment and escalation are two sides of one count. When every conversation either stays with the bot or goes to a person, they add up to 100%. Containment and resolution don’t. Ada measured about 72% containment and 52% automated resolution on the same conversations, so 20 in every 100 stayed with the AI without being resolved.

Deflection looks further than the bot. A customer who gives up on your chat and emails you instead was contained, but the question still reached you, so it wasn’t deflected.

Why a containment rate can read too high

A containment rate can look better than the support behind it, in five ways.

Leaving counts as success. The Postal Service’s menu counted every call that wasn’t transferred as contained, and most of its contained calls ended with the caller hanging up before the closing thank-you message. Some of those callers had what they needed and some had given up on the menus. The inspector general said this overstated the rate, and asked for separate rates for complete calls and abandoned ones.

Silence counts as success. Kore.ai counts a session as contained when the bot sends its welcome messages and the customer never types a word.

Coming back counts twice. A customer who calls back about the same issue within a day wasn’t helped the first time. Parloa’s guide to containment warns that without callback tracking, the rate can count that one problem as two contained conversations.

Hiding the exit counts as success. A bot that makes a person hard to reach contains more conversations, and NiCE’s glossary calls that “forced deflection”. ContactBabel’s US guide says a caller who agrees to try self-service has an “implied contract” to reach an agent when it doesn’t suit them.

Customers see it the same way. In Gartner’s survey of 3,566 customers in February and March 2026, 87% said access to a human agent is essential when a company uses generative AI for service, and only 27% would try a chatbot again after a bad experience. Gartner’s analyst says a chatbot should connect customers to human support, “not containment traps”.

Staying isn’t the same as being helped. The Postal Service’s menu contained 91% of passport calls, though the report notes that on its own, the menu can only help those callers find a post office that accepts passport applications.

Raising containment one question at a time

A single rate hides very different questions. From October 2020 to May 2021, the Postal Service’s menu contained 93% of general enquiries and 56% of change-of-address calls. Calls meant for a specialist team, such as the Inspection Service, were often contained less than 5% of the time. The report called those high transfer rates a good sign, because the callers reached someone more likely to solve their problem.

The biggest rise came in package tracking, the kind of call the Postal Service had worked on in its 2019 financial year. Its menu contained 48% of tracking calls in the 2018 financial year and 72% from October 2020 to May 2021.

For a chatbot, read the conversations that went to a person or ended without an answer, and sort them by what the customer asked. A refund, a bug report or a question about someone’s own account needs a person, so those handovers are the right result. A question your docs should have answered is a gap. Write the answer where the bot reads it, and the next customer who asks can get it without you.

What Deacon answers and what it hands to you

A bot can raise its containment rate by guessing, because a customer who leaves with a made-up answer counts as contained. Deacon answers your customers in seconds, in their own language, from your own docs. When your docs don’t cover a question, it says it doesn’t know instead of guessing and invites the visitor to leave their email, so you can reply.

The way to a person stays open. When a visitor asks for one, Deacon tells them to leave their email and that your team will be in touch, and the owners on your team get an email. Reply from the dashboard, and if they still have the chat open, your reply appears there under your first name.

Every question your docs didn’t cover waits in Conversations with an Answer this button. The answer you write becomes part of what Deacon knows, so it can answer the next person who asks, and the more you answer, the less it needs you. Topics group similar questions, count how many people asked and label most of them as a how-to, a bug report, a feature ask and so on. You see which gaps keep sending people to you, and what customers want you to build next.

Deacon doesn’t report a containment rate, but Conversations holds the counts for one. Set the date to This month and open the tag menu. It counts all conversations, and those holding a request for a person (Wants a person), a question your docs didn’t cover (Unanswered), an error (Failed) or a question that arrived after your plan’s answers ran out (Over limit). Taking Wants a person away from all conversations gives the loose rate, which counts a visitor Deacon couldn’t help as contained. Take away the other three as well, counting a conversation once if it carries more than one tag, for the strict rate, which leaves out every conversation with a question Deacon couldn’t answer. Answering or dismissing a question moves it out of Unanswered, so count before you work through the month’s gaps. The Answered figure on Analytics › Widget is a different measure. It counts questions rather than conversations, and a request for a person counts as answered there.

Human handoff walks through a handover from both sides, and you can try Deacon on the free plan, which needs no card.