Escalation rate, from chatbot handoffs to support tiers

HubSpot’s customer agent report calls it the human handoff rate, and Amazon Connect’s metrics the AI handoff rate. Helpdesk software files each customer request as a ticket, and there the rate counts the tickets your first line of support passes to a specialist. Call centres also count calls passed to a supervisor. SQM Group’s list of call centre metrics files those under call escalation.

The escalation rate formula for a chatbot

June brought 1,000 conversations to the chatbot on a tax-filing app’s website. In 180 of them the visitor asked for a person. The bot passed another 60 to the team without being asked, because the app sends every refund question to a person. Of the rest, 200 visitors left before asking anything.

Zendesk, HubSpot and Microsoft each count a different part of that month.

Microsoft counts a session as engaged once the bot starts on one of the topics it was built for, or hands over. The 200 visitors who left early never got that far, so its rate for June is 240 ÷ 800, or 30%.

Some tools also count a handover nobody answered. In Microsoft’s analytics, a session ends as escalated once the bot starts the transfer, whether or not a live agent picks up. A handover counts in Gorgias’s AI metrics only when the shopper reached a person. Freshdesk’s AI agent report counts conversations transferred to a live agent.

Deacon tags a message Wants a person when the visitor asks for someone from your team. Filter the conversations list by that tag and by This month, then divide by all of this month’s conversations. Zendesk’s BotQA counts the same way. Answering or dismissing that message later can change its tag.

Published figures for calls and chats passed to a person or another agent

None of the helpdesk or chatbot vendors we read publishes an escalation rate from its own customers. The figures below come from contact centre surveys, a call centre research firm and one field experiment, read on 19 September 2026.

Half the US centres fell between 8% and 55%. The US average was 13% in 2013. In the UK survey, 18% of web chats were handled by a virtual agent alone, up from 4% in 2019.

The Taobao study, a preprint, covers only the chats marked as eligible for the AI, under 10% of all chat volume at the time. SQM calls a transfer rate of 15% or less good.

Vendor pages that grade escalation rates disagree. A good rate is under 5% in KPI Depot’s entry and under 20% in Decagon’s glossary. For handovers from AI to a person, Qiscus’s guide calls 20 to 35% healthy. None of the three cites a study.

The handovers a low rate should keep

Some handovers are the right result. A handover set off by one of your own rules is expected, and Microsoft’s guide to bot analytics says it needs no investigation. The tax-filing app’s 60 refund handovers are of that kind.

Hiding the way to a person lowers it too. The US Consumer Financial Protection Bureau looked at chatbots in a June 2023 report on consumer finance. It says a chatbot can lead customers into “continuous loops of repetitive, unhelpful jargon or legalese without an offramp to a human”.

AI vendors choose how early to hand over. Intercom’s escalation guidance says that when a customer is upset, Fin may try to resolve the issue before it suggests handing over. Escalating too early makes Fin less effective, it says.

A target does the same to a support team. Kustomer’s glossary says a target to cut escalations can lead agents to resolve contacts “on paper”. They close tickets without fixing the problem, or skip a transfer the customer needed.

Vendors that charge per resolution bill handovers differently. Under Intercom’s outcome pricing, a handover that one of your written procedures asked for costs 99 cents, and other escalations cost nothing. A customer who asks for a person costs nothing under Salesforce’s pricing for its Help Agent or Zendesk’s automated resolutions.

Ticket escalations between support tiers and to managers

Larger support teams work in tiers. The first line takes every ticket and passes the hard ones up to specialists, and each pass is an escalation. In HDI’s 2016 survey of 506 IT support centres, 53% of tickets were resolved at level 1. The rest went further up, and 4% were resolved by management.

MetricNet’s 2011 column on first level resolution counts the other side, the tickets resolved at level 1 without an escalation. Its net version leaves out tickets level 1 could never have solved. That version averaged about 74% at IT service desks. The range ran from about 38% to 98%.

The ITIL framework for IT services names two kinds. In the ITIL glossary, passing work to a team with more expertise is a functional escalation, and bringing in senior management is a hierarchic escalation.

Complaints that reach a manager are rarer. John Goodman of the research firm CCMC puts the share at 1 to 5% of customers who had a problem. His 2025 article on the complaint multiplier gives no sample.

Freshdesk uses the word for an alert. An SLA, or service level agreement, sets a deadline for replying, and Freshdesk’s SLA policies can email a manager when a ticket misses one. That needs the Pro plan or above.

June’s conversations counted as other rates

The tax-filing app’s bot solved 430 of June’s conversations. That leaves 330 with no handover and no solution, including the 200 visitors who left early.

Counted this way, containment and deflection are both 100 minus the escalation rate. The 330 conversations nobody solved count in their favour. Taking the resolution rate from 100 gives 57%, and the escalation rate is 24%.

A handover can still end in first contact resolution. In MetricNet’s column, a ticket a specialist solves during the customer’s first call counts as solved on first contact, though not at the first level.