Escalation rate for chatbots and support teams

To work it out, divide the conversations passed to a person by all the conversations, then multiply by 100. If your chatbot had 600 conversations in April and handed 90 of them to you, your escalation rate was 15%.

This page is about that support measure. Contracts, rents and construction budgets use the same words for a planned yearly price rise, which is a different thing.

Chatbot tools often call it the handoff rate. HubSpot’s customer agent report shows a human handoff rate, and Amazon Connect an AI handoff rate. Helpdesks count the tickets the first line passes on to specialists, as Geckoboard’s KPI guide does, and call centres count calls transferred to another agent or a supervisor.

For a founder, every escalation is a conversation that lands on you, and some of them should. Deacon is an AI customer support agent that answers your customers from your own docs in seconds, hands you the questions that need you, and has a free plan.

How to calculate escalation rate

The hard part is deciding what counts as passed on, and tools disagree. Say that of April’s 600 conversations, 60 visitors asked for a person. The bot passed on another 30 by itself, 20 because a rule sends every refund to you and 10 because it was still stuck after three tries. And 100 visitors left before the bot had started on anything.

Microsoft leaves out a session until the bot has started on something, such as one of its topics, a knowledge source or a handover, so the 100 visitors who left early drop out of the bottom line. Its analytics also count a session as escalated once the transfer starts, whether or not a live agent picks it up. The same April reads as 10%, 15% or 18%, so check the definition before you compare your rate with anyone else’s.

What is a good escalation rate?

There’s no standard figure, and the published ones count different things. These are the figures we found with a named sample, read on 23 September 2026.

The Taobao study is a preprint, and the AI there was only allowed standard questions, under 10% of all chats at the time. For phone menus, half the US centres had between 8% and 55% of callers zeroing out.

Vendor pages grade escalation rates without a study behind them. Decagon’s glossary calls under 20% healthy, and Kustomer’s glossary gives 30 to 50% as typical for a new AI chatbot and 5 to 15% for a mature one. Your own rate, counted the same way each month, tells you more than any of them.

Escalation rate vs containment, deflection and resolution rate

These rates describe the same conversations from different sides.

So a low escalation rate on its own proves little. It falls when the bot answers more, and it also falls when customers give up. A handover can still end in first contact resolution, if the person it reaches solves the problem there and then.

Which escalations to keep

Some handovers are the right result. Microsoft’s analytics split them three ways. One your own business rule triggers is an expected outcome, which “isn’t something needing investigation or change”. One the bot makes after failing a task several times usually means the customer was stuck. And one the customer asks for is theirs to make.

Hiding the way to a person lowers the rate and costs you the customer. The US Consumer Financial Protection Bureau’s June 2023 report on chatbots warns of bots that keep customers in “continuous loops” of unhelpful answers, “without an offramp to a human customer service representative”.

Timing matters as much as the count. Intercom’s escalation guidance says that when a customer is upset, Fin may try to resolve the issue before offering a person, because escalating too early reduces its effectiveness. The Taobao study found the cost of waiting. When its monitoring software escalated a chat because the customer was frustrated, the person who took over had a harder job repairing it, and supervisors who stepped in before frustration built up did better.

Targets bend the count. Kustomer’s glossary warns that a target to cut escalations can lead agents to resolve contacts “on paper”, closing tickets without fixing the problem or skipping a transfer the customer needed.

Pricing can bend it too. Under Intercom’s outcome billing, a handover that one of your procedures was built to make is a billable outcome, while one the customer asks for isn’t billed. Help Scout charges nothing for an AI answer after which the customer escalates. HubSpot won’t count a conversation as resolved, or bill it, if the visitor asks for a person or trips one of your handoff phrases or guidelines within 72 hours of their last message, though a teammate or workflow assigning it to someone doesn’t stop it counting. Decagon’s per-resolution model charges nothing when its AI hands a conversation to a person, as Decagon pricing explains.

Ticket escalations between support tiers

Larger support teams work in tiers, and each ticket passed up a tier is an escalation. In HDI’s 2016 survey, level 1 resolved 53% of tickets, and the rest were resolved further on, by level 2, desktop support, level 3, management or someone else.

MetricNet counts the other side as first level resolution. Its 2011 column leaves out the tickets level 1 could never resolve, such as a hardware swap, and on that basis service desks resolved about 74% at level 1. The range ran from about 38%, mostly at desks that log a ticket and pass it on, to 98%.

The ITIL glossary, from the framework many IT service desks follow, names the two directions a ticket can go. Passing it to a team with more expertise is a functional escalation. Bringing in more senior management is a hierarchic escalation.

When Deacon hands a conversation to you

A founder’s worry runs both ways, a bot that won’t let customers reach you and one that passes you everything. Deacon answers what your docs cover, in seconds and in the visitor’s own language. It hands you the rest in four cases, when your docs don’t cover the question, when it’s about the visitor’s own account or invoice, when they ask for a person, and when three clarifying questions haven’t got them an answer. Once your plan’s answers run out for the month, every new question gets a short decline and, unless the visitor has already left one, a request for their email.

It says so plainly and asks for the visitor’s email. When they leave it or ask for a person, the owners on your team get an email, unless one of you already has the dashboard open. Reply in the conversation, and if they still have the chat open, your words appear there. Until you reply, Deacon keeps answering their other questions.

Handovers your docs could have prevented become rarer. Press Answer this on a question your docs didn’t cover and write the reply, and it becomes a source Deacon can answer the next person from. The more you answer, the less it needs you. Topics sort the questions into groups with a count of the people who asked, and give most groups a type, such as how-to, account and billing, or wants a person. A how-to group is a gap you can close, and an account and billing group needs you anyway.

Deacon doesn’t report an escalation rate, but Conversations holds the counts. Set the date to This month and open the tag menu. It counts all conversations and those where a visitor asked for a person (Wants a person). Divide the second by the first for the rate as Zendesk’s BotQA counts it. Beside them, Unanswered counts the conversations holding a question your docs didn’t cover, where Deacon offered to pass the question on. Answering or dismissing a question moves it out of Unanswered, and out of Wants a person when the visitor asked for someone in the same message, so read both counts before you work through the month’s gaps.

Human handoff follows one handover minute by minute, and you can try it on the free plan, which needs no card.