The 30-second answer

Deflection rate is the share of customer contacts that get resolved by an automated system without escalating to a human agent. If 1,000 customers ask a chatbot a question this week and 600 of them don't continue to a human, the deflection rate is 60%.

That's the whole formula. The interesting part is what counts as "didn't continue."

How it's measured

There are two common definitions, and most teams use the first while talking as if they're using the second.

  • Contained. The bot handled the conversation end-to-end and the customer didn't escalate within the session. This is what most chatbot vendors report by default. It's a single-channel, single-session number.
  • Deflected. The customer didn't return through any channel for the same issue within a defined window — usually 7, 14, or 30 days. This is the more honest version, because it tries to detect customers who gave up on the bot and called, emailed, or filled in a form.

The second is harder to measure, because it requires identity stitching across channels. Most companies don't have that. So most companies report containment and call it deflection.

Why it's a tempting metric

Deflection rate is easy. It comes straight out of one channel's logs. It rolls up to a single percentage that fits on a slide. It moves when you tune the bot, which makes it feel like a real lever. And it maps neatly to a cost story: each "deflected" contact is a human ticket that didn't happen, multiplied by a cost-per-ticket figure.

All of that is true on the surface. The problem is what's underneath.

What it actually misses

Deflection rate doesn't measure whether the customer got the problem solved. A "deflected" customer might call the next day. They might have given up. They might have left for a competitor. They might have written a complaint somewhere you don't track.

Deflection counts tickets your AI closed. It doesn't count problems your customers actually solved.

This is what we call the AI Volume Paradox: deflection rate goes up, support tickets don't go down. The bot closes more sessions, the human queue stays the same size, and nobody can explain where the contacts went. They didn't go anywhere — they came back through a different door.

How to read it well

Deflection rate is fine as a directional signal if you pair it with the things it doesn't see:

  • Cross-channel resolution. Did the customer come back through another channel within 7, 14, or 30 days?
  • CSAT on deflected contacts. Specifically the ones the bot closed, not the overall average.
  • Repeat-contact rate. How often does the same customer come back about the same issue?
  • Human-loop quality. When the bot does escalate, is the handover clean, or does the human have to start from zero?

A deflection rate reported on its own, without these, is a vanity number. It tells you the bot is busy. It doesn't tell you the bot is useful.

We've written more on the deflection metric in our field note: a high AI deflection rate doesn't mean your customers got help.