How to run it
Each week, pick five tickets at random, from different agents, and include the AI's answers if you use one. Score each against the card below, write one sentence on what worked, and one on what to change. Share the notes with the agent the same day.
The five questions
Score each yes, partly or no:
- Accurate: is everything in the reply true, and does it follow policy?
- Complete: did it answer every question the customer asked, so they will not need to write again?
- Owned: does it say what the agent did or will do, with a time, rather than pointing the customer elsewhere?
- Clear: could the customer act on it, with short sentences and steps in order?
- Human: does it sound like a person who read this customer's message, not a template?
Weigh accuracy first
A kind, clear, wrong answer is the worst result. If accuracy fails, the ticket fails, whatever else is good. Everything else is style.
Review the AI the same way
Score its answers with the same card. Patterns show up quickly: it may be accurate but cold, or helpful but missing a policy. Fix the source (a macro, an article, a fact) rather than the single reply.
Praise specifically
Review is not only for fixing. When a reply is excellent, say exactly why, and consider turning it into a macro so the team can reuse it.
Track a trend, not a grade
Count how many of the five pass each week. Do not rank agents: the goal is a shared standard, and a leaderboard turns it into a contest. If a pattern appears across several agents, treat it as a training or knowledge gap.