Where AI ends and empathy begins

Most "AI for customer support" pitches confuse two questions. The first is: what can an AI agent answer correctly today? The second is: when it can't, what does the next thirty seconds feel like for the customer?

We spend a lot of time on the second question. The handoff is where customer experience is actually won — and where most deployments quietly fail.

What a sensible Tier-0 actually does

Tier-0 deflection is real value when it's scoped honestly. The intents we automate look like:

• Order status, refunds inside policy, return-label generation.
• KYC re-uploads when the failure reason is mechanical.
• Stock-availability checks, simple address changes, password resets.
• Forwarding the right structured context to a human when none of the above apply.

That last bullet is the one most "AI agents" don't ship. Without it, the bot becomes a wall the customer has to argue with before they can be helped.

The handoff is the product

When the agent escalates, three things have to be true:

1. The customer doesn't repeat themselves. The specialist who picks up sees the conversation, the diagnosis attempted, and the customer's stated goal. No "can you start over for me?"
2. The agent has confidence scoring on its own answers. We don't let it commit to anything it isn't ~85%+ sure about. Below that threshold, it asks one clarifying question, then escalates.
3. The escalation is silent to the customer. No "I'll transfer you" lag. The human is already in the conversation by the time the customer notices the tone change.

What we don't automate

We don't automate refunds outside policy. We don't automate the first contact after a billing error. We don't automate anything where the next sentence might need to acknowledge that something went wrong on our side.

The reason is simple: those are exactly the conversations where retention is decided. Getting the routine intents right is the price of admission. Getting the consequential ones right is the product.

The metric we trust

CSAT broken out by escalation type. If the bot handled it end-to-end and CSAT is high, fine. If a human picked up after escalation and CSAT is high, even better — that's the signal that the handoff worked. If CSAT drops after escalation, that's a process bug we go fix the same week.

Bot-only CSAT, on its own, doesn't tell you much. It tells you whether the easy questions felt easy.