AI-Assisted Customer Support Without Losing the Customer

Deflection is the wrong target. It measures how many people gave up, not how many were helped. The useful question is which contacts a model can resolve completely, and which it should escalate untouched.

Sort by consequence, not by volume

The tempting move is to automate the most common ticket. The better move is to automate the ones where being wrong costs nothing — and to route anything involving money, access, or an unhappy customer straight to a person.

• Safe to automate: status lookups, documentation retrieval, routine account changes.
• Escalate always: billing disputes, security concerns, cancellations, anything already escalated once.

Cite, do not assert

An answer with a source can be checked in seconds. An answer without one has to be trusted or re-researched, and re-researching is the work you were trying to remove.

A support model should be judged on the contacts it closes correctly, not the contacts it prevents.