Every AI support vendor publishes a deflection rate. Intercom says Fin resolves 51%. Zendesk says their bots deflect 60%. We say message.ai deflects 73%. None of these numbers are the same number, and you should not buy a tool based on any of them without asking three follow-up questions.
Three deflection definitions that get conflated
There are at least three definitions of 'deflection' rattling around the industry, and vendors pick the one that flatters them. Definition 1: percent of conversations that ended without a human reply (any reason, including the customer giving up). Definition 2: percent of conversations the AI tagged as 'resolved' (vendor-internal, often optimistic). Definition 3: percent of conversations where the customer rated the AI's last reply as helpful.
All three numbers can be reported for the same workspace. They are not equally honest.
The one we publish, and what it actually counts
“If a vendor will not tell you which definition they are using, the answer is the one that flatters them.”
The 73% we publish is Definition 3 with one extra constraint. We count a conversation as deflected if the customer did not reopen it inside 48 hours AND did not rate the AI reply as unhelpful AND was not later assigned to a human for the same intent. The 'and' makes the number conservative. We have a Definition 1 number that is 81%, and a Definition 2 number that is 79%. We do not publish either of those because they make the AI look better than it actually is.
Three things to ask before you believe any deflection number
(1) Which definition are you using? (2) Are you counting conversations that the AI never engaged on, or only conversations the AI actually answered? (3) What happens if I look at the conversations the AI deflected and ask my agents to rate them? If a vendor cannot answer all three, the number is marketing.
What we monitor internally that we do not publish
Two internal numbers we care more about than the public deflection rate. First, false-confidence rate: how often the AI gave a confident answer that turned out to be wrong, measured by agent overrides or customer corrections. Ours sits at 3.1% as of last quarter. Second, escalation-quality rate: when the AI escalates, what share of escalations were 'good catches' vs the AI giving up early. Ours is 87% good catches.
We will publish both of these the moment a competitor publishes either of them. Until then, they stay in our internal dashboards, where they keep us honest.