Deflection rate is one of the most popular metrics in AI-powered customer service. It’s easy to see why: when the number goes up, fewer tickets reach your agents, and costs go down.
So what’s the catch? There are a few.
What deflection rate is (and isn't) measuring
Deflection rate measures how many customer queries did not pass to a human agent. The presumption is that the tickets were resolved by the AI system. But that's not always the case.
In practice, there are three kinds of deflected conversation, and only one is the outcome you actually want:
Good deflection: the customer asked a question, the AI gave the right answer, the customer is happy, and no agent time was needed. This is the dream.
Bad deflection: the customer got a frustrating non-answer, hit a dead end, and gave up. The ticket never reached an agent — but that's because the customer decided the experience wasn't worth their time, not because their problem was solved. They may well be on your competitor's website right now.
As one CX leader put it to us: "A deflection is good for something like a common FAQ, a general question — but a full resolution means this was a real problem and we were able to fix it. If it was deflected, my whole thing is: we don't know [on the surface] what happened to the customer."
That’s the core issue. A deflected conversation is closed, but it is not necessarily resolved.
False-positive deflection: the customer thinks they got the right answer. Your dashboard agrees. But the AI hallucinated, and somewhere out there a customer is acting on incorrect information — a wrong returns policy, a misquoted delivery date, an ingredient that isn’t supposed to be present.
All three show up the same way in a deflection rate dashboard.
The perverse incentive problem
When you optimise for deflection rate, you create a quiet but significant problem. You incentivise the AI to avoid saying “I don’t know” and to avoid handing off to a human.
An AI that freely admits uncertainty, asks clarifying questions, or escalates to a human will have a lower deflection rate than one that confidently answers every query — even when it shouldn't. If deflection rate is the headline number in the dashboard, it becomes the thing people (and the system) optimise for.
The result can be an AI that is less likely to escalate genuinely complex queries to agents, less likely to express appropriate uncertainty, and more likely to give a plausible answer even when it is wrong. It is optimised for the appearance of resolution, not resolution itself.
Why deflection rate is so dominant
The short answer is that it's easy to measure. You don't need post-conversation surveys, sentiment analysis, or quality review processes. You just count whether a human stepped in or not. In a world where customer service teams are under constant pressure to do more with less, a metric this clean and this legible is understandably appealing.
That does not make it wrong. It just makes it incomplete.
It’s similar to automation rate. If you reply with an autoreply to every single message you receive, even if it's just the same line, you could say that that ticket has been automated. But there is a big difference between that sort of automation and fully resolving a ticket through automated means.
What to measure instead (or alongside)
Deflection rate is still worth tracking, but you need a few other measures alongside it. It is not a North Star metric on its own. Combined with other measures, it can tell you how well your AI is actually performing.
Accuracy rate / false confidence rate: Is the AI actually right when it answers? How often is it confidently wrong?
CSAT: Are customers who were "deflected" satisfied with the outcome?
Completed conversations: Did the conversation reach a natural conclusion, or did the customer abandon mid-flow?
Escalation quality: When the AI does hand off to a human, is it doing so at the right moment, with the right context?
If deflection rate is rising and these measures are improving, you’re genuinely doing well. If deflection rate is rising while CSAT is falling and accuracy is drifting, you are likely rewarding the wrong behaviour.
Should you even be maximising deflection?
One CX leader put it bluntly: "If your goal is to deflect customers, that's not good customer experience."
It’s worth pausing on that. Is the goal to minimise human involvement as much as possible?
Not necessarily. There are ticket types where a human agent adds real value — complex complaints, high-value customers, emotionally sensitive situations. An AI that aggressively deflects everything, including those, isn't doing a good job for you or your customers. And depending on your brand positioning, "we'll always connect you to a person when it matters" might be a competitive advantage, not a cost centre.
Automation also does not have to be all-or-nothing. Many tickets can be partially automated: the AI handles the first response, gathers relevant information, and speeds up resolution time — but a human closes the loop. Average Handle Time comes down, customer gets a fast initial response, and your agents are focusing on the parts of the job that actually need them. Deflection rate goes up modestly whereas quality goes up meaningfully.
The bottom line
Deflection rate is not a useless metric. It's a useful signal. It was designed to be easy to measure, not to be the whole story. Treating it as the whole story creates real risks: AI systems that avoid admitting uncertainty, dashboards that look healthy while customers quietly churn, and optimisation pressure pointed at the wrong target.
Measure deflection. But measure it alongside accuracy, CSAT, and completion rates. And ask yourself not just whether your deflection rate is going up, but whether the right conversations are being deflected — and whether the ones that aren't are landing with the right people.
DigitalGenius helps ecommerce brands automate customer service with AI that's built to be accurate, not just fast. Speak to us today.
