Chatbot Deflection Rate Is a Misleading Metric, Here's a Better One

Deflection rate measures one thing: did the customer close the chat window without opening a support ticket. It says nothing about whether their problem got solved. A client proudly reported a 70% deflection rate on their new support bot in a quarterly review. Nobody in that room asked the obvious next question: what happened to the 70% who didn't escalate?

What we found when we actually looked

We pulled a sample of deflected sessions and manually reviewed them. A meaningful chunk weren't resolutions — they were abandonments. The customer asked a question, got a useless canned response, and left. Not because they were satisfied, but because escalating felt like more effort than it was worth for whatever they were trying to do. The bot wasn't solving problems; it was successfully making people give up, and deflection rate counted that as a win.

Deflection rate doesn't measure whether a customer's problem got solved. It measures whether they gave up quietly enough not to show up in your ticket count.

The metric that actually matters

We replaced the headline KPI with a resolution-confirmed rate, pulled from a simple one-tap 'did this solve your problem' prompt at the end of every bot session, paired with a 30-day repeat-contact rate for the same issue type. If someone comes back within a month asking essentially the same question, the bot didn't resolve anything the first time, no matter what the deflection number said.

• Deflection rate alone incentivizes making escalation harder, not making answers better.
• A one-tap resolution confirmation at session end is cheap to build and immediately more honest.
• Repeat-contact rate within 30 days catches the cases where a customer gave up rather than got helped.
• Report these alongside deflection rate, not instead of it, so the tradeoff stays visible to leadership.

Once we switched the headline metric, the bot's real resolution rate came in around 40%, not 70%. That's a less flattering number to put in a slide deck. It's also the only one that tells you whether the thing is actually working.