What is deflection and how do you measure it without fooling yourself?

Deflection is the share of questions that never reached an agent. It's the easiest number to overstate and the most useful to measure honestly.

Deflection is the share of customer questions answered without an agent involved, for example by a help centre, an order status page or an AI agent. The number is only meaningful if you measure whether the customer contacted you afterwards anyway, because a conversation that ends isn't the same as a question that's answered.

Deflection = (completed self-service conversations with no follow-up contact / total questions) x 100

The no-follow-up condition is the whole difference. Without it you also count customers who gave up and emailed anyway, and then you're measuring how often your self-service ends a conversation rather than how often it helps. Take a window of, say, 72 hours and count contact about the same topic.

Why the default number is nearly always too high

Most systems report deflection as the number of conversations the AI or help centre closed. That includes three groups that don't belong: customers who gave up, customers who got an answer that was wrong, and customers who contacted you through another channel afterwards.

That last group is the nastiest, because it moves your work instead of removing it. A chat closed by the AI that produces a phone call an hour later counts in most reporting as successful deflection and as a new phone conversation.

What you can measure

Tie contact to the customer rather than the conversation. Per customer, check whether a conversation about the same topic followed within a few days of a self-service moment. What remains is deflection you can defend.

Also measure the other side: the number of conversations where a customer opens by saying they already looked on the site. That's a sentence often literally in the opening line, and it's the most direct indication your self-service didn't cover the question.

Why deflection shouldn't be a target

The moment a team is judged on deflection, the cheapest way to score is making contact harder. The contact button disappears under a menu, the phone number moves to a page you have to find, and the number rises without a single question fewer.

So steer on questions per order rather than deflection. That number only falls when a question is genuinely removed, and it rises when you force customers to come the long way round.

Where it goes wrong

Counting closed conversations as answered questions

A customer who clicks away also closes the conversation. Without a window in which you measure follow-up contact, you count giving up as success.

Measuring per channel instead of per customer

Chat deflects and the phone rings. Measure per channel and that looks like a win on both sides.

Setting deflection as a team target

The number is too easy to hit by making contact harder. Use it as an observation, not a target.

Adopting your vendor's number

Always ask which definition sits underneath and whether follow-up contact is counted. If that differs from your own definition the two numbers aren't comparable, and the higher one is usually the optimistic one.

What to do this week

Take fifty conversations your AI or help centre closed last month and check per customer whether another conversation about the same topic followed within three days. The percentage that remains is your real deflection. If that's far below what your reporting shows, the gap sits in the conversations where the customer gave up, and those can be read back.

What this looks like in a system

In Cuego a conversation hangs on the customer rather than the channel, so a chat closed by the AI and a phone call an hour later sit on the same customer card. That makes it visible whether a closed conversation was actually the end. Conversations the agent handed over on uncertainty land in a separate list, and that list is also your overview of missing knowledge base articles.

Knowledge base & AI

Frequently asked questions

Without the definition attached the number is meaningless, and definitions vary so widely that comparing with a benchmark yields little. More useful is whether your own figure rises under an unchanged method, alongside questions per order. If the first rises and the second doesn't fall, you've moved work rather than removed it.

Your Cue to Go.

Everything around your customer. Together.

The Customer Contact Platform where conversations, customer data, knowledge, workflows, people and AI come together. Book a 30-minute demo and see it against your own situation.

30-minute demo · then we set it up together

Rather look for yourself first? Take the free website scan