Ticket deflection is the share of customer conversations that end without a human being involved. Take the conversations your bot or help centre handled, divide by the total that arrived, and the percentage is your deflection rate. Every support tool on the market reports some version of it, and it is the number most vendors put in the headline of a case study.
It is also a metric that cannot tell the difference between a customer helped and a customer abandoned. That is worth understanding before you buy anything on the strength of it, or judge your own team by it.
What ticket deflection actually counts
The arithmetic only knows one thing: did a human touch this conversation? A visitor who got a correct answer at 11 pm and went to bed happy counts as deflected. So does a visitor who got a confident wrong answer, believed it, and never came back. So does one who gave up halfway through and bought elsewhere. All three look identical in the report, and two of them cost you money.
The inverse is worse. A bot that recognises the limits of your documentation and offers a human counts against the score, every time it does the right thing. Under this metric, honesty is a defect. A vendor optimising for the number has an incentive to make their bot answer more often and admit less, which is precisely the behaviour that produces the wrong answers you will later clean up.
Why the number is so popular anyway
Because it is easy, and because it flatters everyone involved. It requires no judgement, no reading, and no awkward conversations about answer quality. It moves in the direction you want as soon as a bot is switched on, which makes it perfect for a renewal deck and useless as a signal.
We have written before about measuring an AI support agent without vanity metrics, and this is the metric that piece is mostly about. The argument is not that deflection is meaningless. It is that deflection without an answer-quality check beside it measures traffic diverted, not problems solved.
What we report, and what we do not pretend it means
Serochat’s dashboard shows the same measurement under a plainer name: answered without help, the share of conversations that finished without an escalation to your team. We show it because the trend is genuinely useful — a sharp drop usually means a document went stale or a new question started arriving.
The dashboard also shows an estimate of hours saved. That figure applies a flat assumption of four minutes per conversation your bot resolved on its own. It is a heuristic, we say so in the product, and you should treat it as arithmetic rather than measurement. If your questions are more involved than four minutes, the number is low; if half your bot’s answers were wrong, it is meaningless. We would rather show a rough number and label it honestly than dress a guess as data.
The metric we watch harder is the escalation rate, and specifically what is in the escalated conversations. Those are the questions your content does not answer yet, and they are the only report that tells you what to do next.
How to audit a bot’s deflection during a trial
A trial is the one moment when you can check a vendor’s number instead of accepting it. It takes an afternoon.
- Read twenty conversations end to end, including the ones marked resolved. Count how many contain an answer you would be happy to have sent yourself.
- Ask ten questions your documents genuinely do not cover, and see how many produce a confident answer instead of a handoff.
- Check whether an unanswered question is recorded anywhere you can act on it, or simply disappears from the report.
- Look at what happens after a handoff: does your team get the conversation with its context attached, or a bare notification?
The first of those four is the real deflection rate. It is always lower than the dashboard version, and it is the one that predicts whether your inbox actually gets quieter.
Common questions
What is a good ticket deflection rate?
We will not give you a number, and we would be sceptical of anyone who does. A shop with excellent documentation and repetitive questions will deflect a large share honestly; a business with complex, high-stakes enquiries should deflect far less, and forcing that figure up would be the wrong outcome. A rate is only interpretable next to the quality of the answers behind it.
Does an honest handoff count as a deflection?
Not under the standard formula, no. It counts as a failure, which is the flaw in the metric rather than the behaviour. If you want a fairer version, count a handoff as a success when it reached your team with context and the customer got a good answer quickly.
Should we ignore deflection entirely?
No. Watch it as a trend, alongside satisfaction ratings and the volume of escalated questions. Deflection moving up while satisfaction moves down is the pattern worth catching early, and no single number will show it to you.
Deflection is a real measurement of a fairly narrow thing: how much traffic never reached a person. Whether that traffic was helped is a separate question, and it is the one your customers are answering.