Designing the handoff: when AI should pull in a human
The most important design decision in a voice or chat deployment is when to stop. Get it wrong and you'll either annoy customers or pay for a human anyway.
Every AI conversation will eventually meet a question it shouldn't try to answer alone. The handoff — the moment AI says 'let me get someone who can help with that' — is the single most important design decision in any voice or chat system. Get it wrong in either direction and the deployment falls apart.
Get the handoff threshold too high, and the AI tries to answer questions it doesn't really understand. Customers get the wrong information. Trust evaporates. You end up paying a human to fix the AI's mistakes — which is worse than just having the human do it from the start.
Get the handoff threshold too low, and the AI hands off everything. Customers feel like they're calling a robot whose only job is to put them on hold. The AI becomes a $300/month phone tree. You've spent money to make the customer experience worse.
The right threshold is specific to your business. We tune it across three dimensions. First, topic — there are categories the AI handles fully (booking, hours, pricing on standard items) and categories it never tries (legal advice, medical advice, complex pricing). Second, confidence — every response has a calibrated confidence score, and below a tuned floor we hand off. Third, tone — if the customer is upset or confused, we hand off regardless of confidence; emotional escalation is a human's job.
Done right, the handoff is invisible. The customer says 'I'd like to speak with someone about a refund' and the AI says 'I'll get someone over to you in just a minute' and texts your team simultaneously. Customer experience: smooth. Operational cost: minimal. Trust: intact. That's what we're tuning toward, every deployment.