Deflection is a vanity metric

Open almost any customer-support dashboard and you will find a number near the top called deflection rate. It is presented as a success metric. It is the percentage of incoming contacts that were resolved without reaching a human agent.

Read that definition again, carefully. It does not say the problem was resolved. It says the human was not reached. Those are different events, and treating them as the same is the single most expensive measurement error in customer experience.

What deflection actually counts

A contact is "deflected" when a customer who started toward a human didn't finish the trip. There are two reasons that happens.

One: the customer's problem got solved on the way — by the help article, the chatbot, the automated flow. Genuine resolution. Good.

Two: the customer gave up. The chatbot looped. The article didn't match. The callback never came. They closed the tab and absorbed the loss, or they churned, or they're now writing the review. Also "deflected." Also counted as success.

The deflection rate adds these two outcomes together and reports the sum as a win. A team optimizing it cannot tell the difference between helping a customer and exhausting one — and worse, the cheapest way to move the number is almost always the second one. Make the chatbot a little stickier. Bury the contact link one click deeper. The metric goes up. The dashboard turns green. The customers are not okay.

Why this survived so long

Deflection persisted because, for two decades, it was a faithful proxy for the thing the business actually cared about: support cost. A deflected contact is a human-minute not spent. When the only lever was rationing access to expensive humans, "contacts that didn't reach a human" was a reasonable thing to watch.

That era is over. Automated resolution is now genuinely good at a large class of problems — which means the metric and the goal have come apart. You can now have a high deflection rate because you're resolving issues or because you're frustrating people, and the number reads identically either way. A proxy is only useful while it tracks the real thing. This one stopped.

What to measure instead

Three replacements, in order of how hard they are to game.

1. Resolution without escalation. Not "did they avoid a human" — did the problem end. This requires you to actually know the outcome: a follow-up signal, a confirmation, a downstream check that the thing the customer wanted is now true. Harder to measure. Impossible to fake by making your chatbot worse.
2. Repeat-contact rate. If the same customer is back about the same issue within a week, the first contact was deflected, not resolved. This number is brutally honest and most teams already have the data to compute it.
3. Effort-to-resolution. How much work did the customer do — steps, channels, re-explanations — before the issue ended? A "deflected" contact that cost the customer three channels and forty minutes is not a success. Effort is the texture deflection rate flattens away.

We made the broader version of this argument in The Asymmetric Service Economy: support spend rises every year while satisfaction falls, and metrics like deflection are part of why — they reward cost reduction that is invisible to the customer right up until it isn't.

The test

Here is a clean way to check whether your CX org is measuring the right thing. Ask: if we made our self-service worse, would any of our headline metrics get better?

If deflection rate is on the dashboard, the answer is yes — and that should end the conversation. A metric that rewards you for degrading the product is not a metric. It's a trap with a green checkmark on it.

Count resolutions. Count the customers who came back. Count what it cost them to get unstuck. Those numbers can't be gamed by hiding the contact button — and unlike deflection, when they improve, it means something good actually happened.