An AI agent handoff is the transfer of a live conversation from an AI agent to a human, carrying the full context with it. A complete handoff moves the conversation history, the customer record, what the agent already tried, and the reason it stopped.

Handoff is the seam between automation and a person, and it is where most of the customer-visible damage in AI support happens. A handoff that drops context makes the customer repeat themselves, which costs more goodwill than the original failure did. The term covers transfers to a human, to a specialist queue, and to another agent.

When should an AI agent hand off to a human?

Handoff triggers fall into two groups: capability triggers and obligation triggers. Capability triggers fire when the agent cannot finish: low confidence, a missing tool, a case type outside its scope, or a customer who simply asks for a person. Those are tuning decisions, and firms set the threshold where they want it.

Obligation triggers work differently, because the rule decides rather than the model. In UK financial services they fire on signs of vulnerability, on anything meeting the definition of a complaint, and on disclosed financial difficulty. Consumer Duty requires firms to avoid foreseeable harm, so routing those cases to a person is the rule working as intended rather than automation failing.

The distinction matters at procurement. A vendor who discusses handoff purely as a confidence threshold has not built for a regulated operation.

What makes an AI agent handoff work in financial services?

A good handoff is judged by what the human receives, not by how smoothly the conversation switched over. The person arrives with the work already done, and the customer notices nothing beyond a change of voice. What has to travel:

• Full conversation history, rather than a summary the human cannot verify
  
  

• The customer and account record the agent was working from
  
  

• Actions already taken, including any tool calls and what they returned
  
  

• The trigger reason, so the human knows whether this is a capability limit or an obligation
  
  

• Any vulnerability or complaint flag raised during the conversation
  
  

Warm transfer matters more on voice than in text. A caller put back into a queue after speaking to an agent experiences the handoff as a failure, whatever the transcript shows.

Frequently asked questions

What is the difference between handoff and escalation?

They overlap without being identical. Escalation moves a case up a level of authority, usually because it is harder or more serious. Handoff describes the transfer itself, whoever receives it, including a sideways move to a specialist queue at the same level.

Should customers be told they are talking to an AI?

Practice varies by firm and jurisdiction, and the trend is towards disclosure. Two questions are worth separating: whether you disclose at the start, and whether you confirm when a human takes over. The second causes fewer disputes, because customers notice the change and want it acknowledged.

Does a high handoff rate mean the AI is failing?

Not by itself. A deployment that routes every vulnerability and complaint case to a person shows a handoff rate that looks poor against a pure containment metric, while doing exactly what the rules require. Judge handoffs by what triggered them, separating obligation from capability before reading the number.