Human Handoff

A human handoff (also called escalation to a human agent or human escalation) is the moment a conversation handled by a chatbot or AI agent is transferred to a person on the team. It happens when the case goes beyond what the automated system can resolve with quality: an ambiguous question, an upset customer, a commercial exception or an issue that calls for human judgment and accountability.

The key to a good handoff is not just passing the conversation along but passing it with full context: the chat history, the customer's data and what the AI already tried travel with the conversation, so the person does not start from scratch or ask the customer to repeat everything. In a sales operation, this clean pass between automated service and the team is part of what Sellium orchestrates on top of the CRM platform.

At the design level, the handoff defines when it is triggered (rules, detected intent, sentiment, an explicit request to "talk to a person"), who it is routed to (a queue, a skill, an available agent) and how control comes back: many modern flows let the human resolve the issue and then hand the conversation back to the automated agent.

The human handoff is one of the concepts that separates a good conversational system from a frustrating one. No matter how capable an AI agent is, there will always be cases where the right move is for a person to step in: a sensitive complaint, a price negotiation, a decision with financial or legal consequences, or simply a customer who asks to talk to someone. The handoff is the mechanism that makes that transition smooth instead of a dead end.

Why it matters

A chatbot that does not know how to escalate does more harm than good. The customer gets stuck in "I didn't understand your question" loops, repeats the problem three times and ends up abandoning the conversation or complaining on social media. A well-designed handoff avoids that ending: when the system recognizes it has reached its limit, it passes the case to a person without friction and without losing context. This raises satisfaction (CSAT) and, in sales, keeps a hot opportunity from going cold because of poor automated service.

How it works

The handoff rests on three design decisions:

• Trigger (when to escalate): it can be explicit (the customer writes "I want to talk to a human"), rule-based (high amounts, certain products, keywords), driven by sentiment detection (frustration or anger) or by low model confidence when it cannot find a solid answer. In healthcare, the handoff is mandatory at any clinical signal: the agent handles the appointment or the refill and hands off to the healthcare professional.
• Routing (to whom): the conversation is assigned to a queue, to a team by skill (sales, support, collections) or to an available agent, ideally meeting a response-time SLA.
• Context transfer (how): the full history, the customer's data and a summary of what the AI tried go along with the transfer. Without this, the handoff is just a "let me pass you to someone else" that forces the customer to start over.

A key practice is human-in-the-loop: in many designs the AI does not switch off when it escalates, but stays on as the human agent's copilot (it suggests replies, looks up data) and can take the conversation back once the hard part is resolved. Human-in-the-loop is precisely the pattern where a person supervises or approves what the system proposes.

A concrete example (Latin America)

A wholesale distributor in Argentina takes orders over WhatsApp with an automated agent. The bot handles the bulk of routine questions (order status, price list, opening hours). When a customer writes "I need a special volume discount", the agent detects that it is a commercial exception outside its scope and triggers the handoff: it routes the chat to a sales rep for that territory, passes along the history and the account data, and the rep continues the negotiation with all the context in view. The customer never felt a break; for them it was the same conversation.

Common mistakes

• Escalating too late: leaving the customer going around in circles with useless answers before passing them to a person. A trigger based on frustration or failed attempts works better.
• Escalating without context: transferring the conversation "cold" and forcing the customer to repeat everything, the mistake that breaks the experience the most.
• Having no one to escalate to: promising "I'll pass you to an agent" outside business hours or without a staffed queue. The handoff needs a real destination and an SLA.
• Not closing the loop: the human resolves the case but the conversation never returns to the system or gets logged, so traceability is lost.

Modern handoff vs traditional transfer

In short, a human handoff is not a sign that the AI failed but a deliberate design feature: it sets the healthy boundary between what is worth automating and what calls for human judgment. The most mature conversational systems are measured as much by what they resolve on their own as by how well they know when to ask for help.