Four configurable triggers route the conversation to a named handler — with the full conversation summary, the transcript, and any actions the AI has already taken. No cold transfers. No repeated information.
Four Triggers
Each trigger is configurable per agent. Each fire is a Timeline event. Each handoff is a warm transfer with full context.
Every conversation turn is scored against a per-task confidence threshold. Clarifying a customer's name has a different threshold to confirming a settlement value. Below the threshold, transfer.
Examples
Tone, repetition, and escalating language are scored on every turn. Sustained frustration triggers handoff before it becomes a complaint. Vulnerable-customer signals are a separate, harder trigger.
Examples
“Speak to a person.” “I want a manager.” “Get me a human.” Phrasing variations all trigger immediate transfer. The AI does not push back, attempt to deflect, or ask why.
Examples
Tool failure, integration timeout, ambiguous customer input that resists clarification, or hitting a configured policy boundary. Handed to a handler with the diagnostic captured in the case file.
Examples
The Handoff
Four steps from trigger fire to handler resolution. Every step attributed in Timeline.
One of the four triggers crosses threshold. Conversation continues until a clean handoff point — usually within one or two turns — without abandoning the customer mid-sentence.
The AI generates a structured summary: who the customer is, what they want, what has been done, what is outstanding, and why it is being escalated. Summary written to the case file.
The receiving handler picks up with the summary on screen, the conversation transcript open, and any actions the AI has already taken visible. No cold transfer, no repeated information.
Trigger reason, summary, handler ID, and timestamp logged in Timeline. The case shows a clear chain of custody from AI to handler. Auditable, exportable, attributable.
What The Handler Sees
The receiving handler picks up with everything the AI knew, in a structured summary. Conversation context survives the handoff — the customer never has to start over.
Design Principles
The receiving handler always sees what the customer said and what the AI did. The customer does not have to repeat themselves. The handler does not have to start from scratch.
Confidence cutoffs, frustration sensitivity, explicit-handoff phrase lists, and tool-failure handling are configurable per agent definition. Tune from real call data during the pilot.
Vulnerable-customer signals are a separate, more sensitive escalation path. Explicit indicators (children, medical needs, distress markers, life-safety language) trigger immediate human transfer regardless of confidence score.
Same triggers, same handoff flow, on voice, chat, WhatsApp, SMS, and email. Voice handoffs are warm transfers; async-channel handoffs route the conversation thread to a handler queue with the same context summary.
FAQ
Voice handoffs are warm transfers, typically within 5-10 seconds of the trigger firing. The handoff happens at a natural conversation break — the AI does not abandon the customer mid-sentence. Async channels (WhatsApp, SMS, email) route the conversation thread to a handler queue immediately, and the handler picks up at their queue prioritisation.
Confidence cutoffs, frustration sensitivity, and explicit-handoff phrase lists are configurable per agent definition. Implementation team configures the initial values during onboarding. Most teams adjust the values during the 30-day pilot based on real call data — too sensitive surfaces noise; too permissive risks slipping calls.
The customer summary, the conversation transcript, the actions the AI has taken, the trigger reason for the handoff, and any outstanding work flagged for resolution. The handler does not start from scratch — they pick up the conversation with full visibility of what the AI did. Voice handoffs include a brief verbal context line from the AI before the transfer completes.
Vulnerable-customer signals are a separate, harder trigger that bypasses the standard confidence/frustration thresholds. Explicit signals (life-safety language, medical disclosures, distress markers around children) route immediately to a human handler with no AI processing time. Switchboard supports your vulnerable-customer process; it does not replace it.
Yes. Routing rules per agent definition specify which queue or named handler receives a transfer based on the trigger reason and the case context. A frustration-triggered handoff on a TL conversation might route to the engineer who set the value; a vulnerable-customer signal might route to a specialist team. Routing is configurable.
Yes. Every escalation is a Timeline event with the trigger reason, the conversation summary, the handler who received it, and the timestamp. The case file shows a clean chain from AI conversation to human resolution. Auditable end-to-end.
Pilot Switchboard on your noisiest workflow. Watch the handoff metrics: transfer rate, handler-resolution time, customer-satisfaction post-transfer. Tune the thresholds. Decide.
