Designing escalation from an AI agent to a person
The handoff is where customers decide whether an AI agent helped them. This guide sets out when to escalate, what to pass on and how to know the design is working.
Contents
An AI agent that resolves most conversations still fails a customer whenever its handoff is poor. Escalation design decides what happens to the conversations the agent should not or cannot finish, which are usually the ones that matter most to the customer.
Why the handoff decides the experience
Klarna's experience is the clearest public example. After reporting in 2024 that its assistant handled two-thirds of customer service chats1, its chief executive said in May 2025 that the company was hiring people again, because an AI-first approach had produced lower-quality service, and that customers must know "there will always be a human if you want"2. The lesson is not that the agent failed but that the route to a person mattered as much as the agent.
When to escalate
Escalation triggers fall into three groups:
- The customer asks. A request for a person should always be honoured.
- The agent is unsure. Low confidence, missing data or a request outside its permissions.
- Categories set in advance. OpenLazarus lists billing disputes, technical bugs and emotional customers as escalation cases, and its example escalation combines a billing dispute, cancellation intent and a history of repeat issues3.
The third group is a business decision, not a technical one. It should be written down, reviewed with the people who will receive the escalations, and revised monthly.
What the person should receive
A handoff should make the customer's next sentence unnecessary. OpenLazarus describes an escalation that carries the customer, the issue, prior history and sentiment, with a recommended action, and a transfer in which the colleague receives the full conversation history and company profile3. Its support page also describes the agent routing escalations to a team channel in Slack3.
Other vendors build the handoff into their pricing and tooling. Intercom lets companies configure procedures that end in a handoff to a person, and counts that as a billable outcome4. Gorgias reports handover rates by topic, so that the topics with the most handovers can be fixed first5.
An escalation checklist
- A written list of categories that always go to a person.
- A one-step route to a person for any customer who asks.
- A summary for the receiving person: who, what, history, what was tried, the customer's state.
- No repeated questions after the transfer.
- A named owner for each escalation category.
How to measure it
| Measure | Why it matters |
|---|---|
| Escalation rate by reason | Shows where the agent's knowledge or permissions are short |
| Time from escalation to a person | Shows whether the route works at all hours |
| Repeat contacts after an AI resolution | Catches "resolutions" that were not |
| Satisfaction after escalation | Shows whether the handoff worked for the customer |
| Re-escalations | OpenLazarus's example report flags conversations that customers escalated again3 |
A plan for designing escalation
- Write the always-escalate list with the team that will receive the cases.
- Define the summary format and test it with the people who will read it.
- Set the route by hour: who receives escalations at night and at weekends.
- Launch with a conservative list, then shorten it as evidence comes in.
- Review weekly for the first month: escalations by reason, repeat contacts and re-escalations.
- Feed human resolutions back into the agent's knowledge, so that the same case escalates less often.
Gartner, in predicting that agentic AI will resolve most common issues by 2029, advises service leaders to set AI interaction policies that cover privacy, security and escalation6. A company that has done this can say, for any customer, why the AI handed them over, how long they waited and whether a person solved the problem the first time. For the agents ranked side by side, see the main guide.
Questions buyers ask
When should an AI customer service agent escalate to a human?
When the customer asks for a person, when the agent is unsure, and for categories the company decides in advance, such as billing disputes, cancellation intent, legal or safety issues and repeat complaints. OpenLazarus, for example, escalates billing disputes, technical bugs and emotional customers.
What information should the AI pass to the human agent?
The customer’s identity and history, the issue in one or two sentences, what the AI already tried and the customer’s state. OpenLazarus describes an escalation summary with the customer, the issue, prior history and sentiment, plus a recommended action.
Should customers always be able to reach a person?
Klarna’s chief executive said in May 2025 that it was critical to make clear to customers that there will always be a human if they want one, after the company found an AI-first approach produced lower quality. Gartner advises setting AI interaction policies covering escalation.
Sources
- Klarna AI assistant handles two-thirds of customer service chats in its first month. Klarna, 27 Feb 2024. Accessed 8 Nov 2026.
- Klarna is hiring customer service agents after AI couldn’t cut it on calls, according to the company’s CEO. Entrepreneur, 9 May 2025. Accessed 8 Nov 2026.
- AI Support Agent: replace Zendesk, Intercom and Freshdesk. OpenLazarus. Accessed 8 Nov 2026.
- Fin pricing. Intercom. Accessed 8 Nov 2026.
- AI Agent. Gorgias. Accessed 8 Nov 2026.
- Gartner predicts that agentic AI will solve 80 percent of customer problems by 2029. CX Today, 5 Mar 2025. Accessed 8 Nov 2026.
Updated 8 Nov 2026: First published.