There is a number most support automation vendors will happily put in front of you, and it is usually the wrong one. Deflection rate measures how many people did not reach a human. It says nothing about whether their problem went away.
Two very different outcomes, one identical ticket count
Picture two customers with the same question about a failed payment.
The first gets an answer that identifies the failed charge, explains the reason, and retries it. She is finished. The second gets a link to a help article, reads half of it, decides it does not apply, and gives up. He is also finished, in the sense that he never contacted you again.
Both count as deflected. Only one counts as resolved. And the second customer is now measurably less likely to renew, which is a cost your support dashboard will never show you.
How to separate them
You need three signals, and none of them are difficult to collect:
- Reopen inside seven days. If the same customer comes back with the same problem, it was never resolved. This is the single strongest indicator.
- Abandonment mid conversation. Someone who stops replying after an automated answer did not necessarily get what they needed.
- Downstream contact on another channel. The customer who gives up on chat often turns up in email two days later, or on social media, which is worse.
Instrument all three and your resolution number drops. That is the point. A lower honest number is worth considerably more than a high one you cannot act on.
If a metric only ever moves in the direction you want, you are not measuring anything. You are decorating.
The compounding part
Deflection failures are not evenly distributed. They cluster on your hardest, highest value questions, because those are exactly the ones a shallow answer cannot handle. Which means an automation strategy optimised for deflection systematically fails your most important customers while reporting success.
Over a year, this produces a support function that looks efficient on paper and generates a steady, unexplained drag on retention.
What good looks like
An automated system worth having does three things a help article cannot:
- Reads the specific account, not the general case
- Takes the action the answer implies, rather than describing it
- Hands over cleanly when it genuinely cannot finish, with the context attached
That third one matters more than people expect. An agent that knows the limits of what it knows, and escalates early with a useful summary, will beat a confident one every time.
The number to put on the wall
Resolution rate, defined narrowly: conversations closed without a reopen inside seven days and without a duplicate contact on another channel. It is harder to move. It is the only one that means anything.