← Back to Field Notes
Every support bot we’ve ever used has the same problem. You tell it something wrong, and it agrees with you. Politely, enthusiastically, and completely unhelpfully.
“I’d like to cancel because feature X doesn’t work.” “I’m so sorry to hear that. Let me help you cancel right away.”
Except feature X does work. The customer is just using it wrong. And now instead of a two-minute redirect, you’ve lost a customer — because the bot was too agreeable to say “actually, here’s how that works.”
We started wondering what would happen if we built a support agent that could respectfully disagree. Not argue. Not be rude. Just… honest. “I understand the frustration. But I think what’s happening is this — let me show you.”
The results from the experiment genuinely surprised us. CSAT scores went up 12% when the agent pushed back compared to when it blindly agreed. Resolution time dropped 35%.
People don’t want to be agreed with. They want their problem solved. And sometimes solving the problem means telling them they’re looking at it wrong.
The catch — and there’s always a catch — is tone. Too direct and satisfaction cratered. The agent needed to lead with empathy, then evidence, then the redirect. Empathy sandwich. It sounds simple but the calibration is everything. One degree too far in either direction and you’ve either got a pushover or an argument.
Full experiment: AI Support Argues Back. We’re now building this as a reusable pattern for Relay’s workflow templates. The question we haven’t cracked: how does this translate to higher-stakes contexts — billing disputes, contract issues, compliance conversations — where “respectful disagreement” has legal weight?