AI for customer support: do it without making customers hate you
Most AI customer-support deployments feel like punishment. Here's how to do it so customers actually prefer it.
We've all hit the customer-support bot that loops you four times before letting you talk to a human. That's not AI customer support. That's customer punishment. Here's how to do this properly.
Three rules
- Always offer a clean exit to a human. The agent's first message should always include 'or reply HUMAN to talk to a person'. Don't hide it. Don't put it on screen 4. Don't use it as a stick.
- Resolve in one message where possible. If the agent has the answer, give the answer. Don't ask for an order number when the customer's email already includes it.
- Hand off with full context. When the agent escalates, the human gets the full conversation, the customer's history, and the agent's suggested resolution. Not a fresh ticket.
What 'good' looks like
For a DTC e-commerce brand, a well-built CX agent resolves 60–75% of tickets fully (refunds, exchanges, order status, sizing). It escalates the rest with full context. Average response time drops from 18 hours to under 5 minutes. Customer satisfaction goes up, not down.
What kills it
- Letting the agent guess. If it doesn't know, escalate. Don't hallucinate a refund policy.
- Hiding the human option. Customers can tell, and they hate it more than the original problem.
- Not measuring escalation rate. If escalations climb week-over-week, the agent's drifting. Maintenance, not magic.
AI customer support should make customers say 'that was easier than expected' — not 'I had to fight a robot for 20 minutes'.
— Aaron Manton, AI Operator Club