Confidence threshold, not vibes
Sonar scores every customer signal. If the score is below your configured threshold, the AI offers to find someone — it doesn't fake the answer.
Feature · Handoff
The AI that knows when to hand off.
A chatbot answers everything until it embarrasses you. OrcaLinq pages a person before that happens — based on confidence, policy, and customer intent.
The problem
Most chatbots fail by trying too hard.
AI overconfidence is the most common reason customers escalate to "I want to speak to a human." A semantic PBX flips the default: AI is conservative, escalation is fast, and the audit trail records why.
Policy-driven escalation
Refund language, regulatory keywords, VIP tags, and high-cart-value triggers force handoff regardless of confidence.
When handoff fires
Six default triggers, plus your own
The routing policy ships with sensible defaults. Adjust them per intent, per channel, per business hours, or per customer segment.
Customer asks for a person
"Can I talk to a human?", "Get me a manager", or sentiment that says the visitor is done with automation. Handoff fires immediately, no questions asked.
Confidence is too low
Sonar scores intent; if confidence drops below the routing threshold, the AI offers to find someone instead of guessing.
Policy says escalate
Refunds, complaints, legal, medical, account-sensitive, or VIP-tagged conversations always page a person, regardless of confidence.
Buying intent
High-value purchase signals or qualified prospect cues route to sales rather than support. Owner-defined.
Tool failed
If a knowledge lookup or live tool fails (CRM down, calendar timeout), the AI doesn't fake an answer — it hands off.
Custom triggers
Define your own. Keyword groups, intent classes, customer attributes, channel rules, business-hours overlays.
What the agent sees
A handoff that respects the agent's time
When a person is paged, they get the context they need to reply in seconds — not a fresh conversation thread to read from scratch.
- Short summary. Two-line summary of intent and what the AI already tried.
- Customer profile. Name, channel, language, prior conversation, lifetime value.
- Suggested reply. Templated response that matches your tone, ready to send or edit.
- Transcript link. Full conversation history, timestamps, AI reasoning.
- Lock state. Only one agent owns a conversation at a time. Reclaim or release explicitly.
Handoff FAQ
Frequently asked questions
How does the AI know when to hand off?
Each conversation flows through a routing policy with explicit triggers (keyword, intent, customer attribute, channel, time of day) and a confidence threshold. If any trigger fires, the AI doesn't try to answer — it pages the right human.
What does the agent see when paged?
A short summary of the conversation, the intent classification, the customer profile (if known), the suggested response template, and the full transcript.
Can the customer go back to AI?
Yes. Agents can release a conversation back to AI when it makes sense (after answering, after a payment, after escalation resolves).
Does the customer know they were handed off?
Configurable. Some teams want a clean transition message ('A teammate is joining you'), others prefer the AI/agent identity boundary to feel natural and silent.
What if no human is available?
Out-of-hours and unavailable agent flows are part of the routing policy. The AI can collect contact, propose a callback time, or accept an inbound voice note depending on the channel.
Related pages
Keep exploring
See AI human handoff in your own routing policy.
Bring a customer message that should be human and one that AI should handle. We'll show what the routing decision and handoff look like in your tenant.