Deflection rate is the share of incoming customer messages that get resolved without a human ever having to step in. If your team gets a thousand tickets in a month and four hundred are answered fully by an AI agent, your deflection rate is forty percent.
It is the single most useful number for measuring whether automated support actually reduces workload. Other metrics tell you how the AI performs in isolation. Deflection tells you how much human time the AI freed up.
How it differs from automation rate
Some vendors quote "automation rate", which can mean almost anything: messages the bot touched, messages the bot started, messages the bot eventually closed even if a human had to fix things. These numbers tend to flatter the tool.
A clean deflection rate counts only the conversations that ended without a human reply and where the customer did not come back with the same issue. If the AI answered, the customer accepted the answer, and the ticket stayed closed, that is a deflection. Anything else, including the conversation the AI answered and the customer immediately re-asked of a human, is not.
Why it varies so much
Deflection depends on what your team handles. A SaaS team with rich documentation, predictable product questions, and good account data can hit fifty to seventy percent. A high-touch consultancy with custom contracts and one-of-a-kind requests will be lower. A retail brand with clean order data and a tight policy lands in the middle.
It also depends on how the AI is configured. A very conservative threshold yields high accuracy but low deflection. A very loose threshold deflects more but makes more mistakes. The right setting is the one your CSAT and your re-open rate both like.
Why it matters
A point of deflection is a point of staff time you can spend on the conversations that need it. For most teams, the value of AI support is not in handling the hard cases. It is in clearing the routine ones so your humans can focus on the hard ones.
In Keloa
In Keloa, the dashboard reports deflection on conversations where the AI agent replied, the customer did not come back, and no human took over. We separate this from automation rate to keep the number honest. See pricing for how deflection maps to plan tier.