Containment rate for chatbots and IVR, with benchmarks

Call centres use it to judge their automated phone menus, known as IVRs, short for interactive voice response. A 2021 report on the US Postal Service’s phone menu defines it as “the rate of calls that are not transferred to an agent”. Website chatbots are measured the same way. That is why you will see it called IVR containment rate, call containment rate or chatbot containment rate.

Calculating containment rate with hang-ups counted or left out

The usual formula takes the conversations that reached a person away from all the conversations the bot or menu started. Those handovers, as a share of the total, are the escalation rate.

A bank’s chatbot had 5,000 conversations in May. It handed 1,400 of them to a person, and in another 900 the customer left partway through. Counting only the handovers, 3,600 conversations were contained, a rate of 72%. If the 900 who left count as failures too, the rate is 2,700 ÷ 5,000, or 54%.

Tools disagree about those 900. Three of the four we checked leave an abandoned session out of the contained count. They are Genesys’s bot metrics, Kore.ai’s conversations dashboard and Dialogflow’s experiments report, though Dialogflow only counts contained sessions for phone agents. The fourth, Cognigy’s engagement report, takes away only the sessions with a handover, so a customer who gave up still counts as contained.

The bottom of the fraction moves as well. Kore.ai counts a session as contained even when the bot sent its welcome messages and the customer never typed a word.

Customers who come back the next day

Someone who gives up on the bot today and tries again tomorrow can be counted as contained twice. Parloa’s guide to containment says the rate can count one unresolved problem as two contained conversations when nobody tracks callbacks.

Some tools check for the return visit. In Dialogflow’s experiments report, a session the customer restarted is a callback, and callbacks stay out of the contained count. Containment sits next to a repeat-contact check on Genesys’s channel dashboard, the share of interactions with no further contact within seven days. If your tool has neither, pick a window and look for a second contact from the same customer inside it.

How the Postal Service’s phone menu counted hang-ups

In 2021 the US Postal Service’s inspector general reported on the phone menu behind 1-800-ASK-USPS. In the 2018 financial year the menu contained 62% of more than 72 million calls. From October 2020 to May 2021 it contained 78%. The target was 80%.

The menu counted every call it kept from a person as contained. Hang-ups counted too. Most of the 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 report said this overstated the rate, and recommended separate rates for completed calls and abandoned calls.

Self-service figures from 414 contact centres

ContactBabel, an analyst firm, surveys contact centre managers each year, and its US guide for 2024 and UK guide for 2024 give the figures below. The median is the middle centre when all of them are ranked.

A quarter of the US centres handled 44% or more of their calls entirely in self-service, and another quarter handled 5% or less.

Chatbot containment averages rarely name their sample. The one we found is 72% across more than 550 AI deployments, in Ada’s guide to automated resolution. Ada does not say whose they were. The 70 to 90% ranges on vendor pages such as Botpress’s guide and Talkative’s cite no study.

For phone menus, the Postal Service report says “there are no universal benchmarks for containment, transfer, or abandonment rates”. We read all of these on 19 September 2026.

When the only way out is to hang up

The quickest way to raise containment is to make the person hard to reach. NiCE’s glossary calls that “forced deflection”. Businesses that “hide their agents” in menu loops give all phone self-service “a bad name”, according to ContactBabel’s US guide.

Gartner surveyed 3,566 customers in early 2026. Of those, 87% said they must be able to reach a human agent when a company uses generative AI for service. Only 27% would try a chatbot again after a bad experience. Gartner’s advice is to show a clear way to a person whenever the chatbot cannot resolve the issue. The person should get everything the customer has already said.

Raising containment one topic at a time

A single rate hides very different topics. The Postal Service’s menu contained 93% of general inquiries and 91% of passport calls, but only 56% of change-of-address calls. Calls that belonged with 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.

For a chatbot, read the conversations that were handed over or abandoned and sort them by what the customer asked. A refund or a bug report needs a person, so those handovers are the right outcome. A question your help pages should answer is different. Write the missing answer, or fix the page that is out of date. If 200 of the bank’s 5,000 conversations were handovers about one missing answer, that fix is worth four points of containment.

Deacon, the chatbot we make, hands the conversation to you when it cannot answer from your documentation or the visitor asks for a person. It asks for their email address, so your reply can reach them after they have left the page.

The 20 points between containment and resolution

Containment only asks whether a person got involved. Ada’s guide to automated resolution puts containment at about 72% and automated resolution at about 52%, measured on the same conversations. The 20 points between them are conversations that stayed with the bot without solving the customer’s problem. The resolution rate counts only the solved ones, and first contact resolution asks for the problem to be solved in the first contact.

A customer who leaves the bot and phones you instead still counts as contained. The deflection rate should catch the call, because it looks across every channel you run, but only if something links the call to the chat.