# Why Most AI Customer Support Tools Are Optimized to Lie to You

> Deflection rate is the wrong number for AI customer support — here's what we track instead and why it changes how the agent behaves.

_Arslan Nasir · 2026-07-23 · Support_

Canonical: https://peeve.ai/blog/why-most-ai-customer-support-tools-are-optimized-to-lie-to-you/

Every AI customer support vendor pitch deck has the same slide: a deflection rate going up and to the right. 40%. 60%. "We resolved 8 out of 10 tickets without a human." It's the number that gets a deal signed, and it's the number that quietly rots your support quality from the inside.

Here's the problem: deflection rate measures whether the AI *answered*, not whether the customer's problem *went away*. Those are different things, and the gap between them is where trust dies.

## The deflection rate trap

A chatbot can "resolve" a billing dispute by explaining your refund policy in four confident paragraphs. Ticket closed. Deflection: successful. Customer: still charged twice, now also annoyed that a robot lectured them about terms of service.

Deflection rate rewards this behavior because it's cheap to measure — did the conversation end without a human being pulled in? It says nothing about:

- Whether the customer came back within 48 hours with the same issue
- Whether the "resolution" required the customer to go do something manual themselves
- Whether the AI quietly gave up and buried an escalation path three clicks deep
- Whether the answer was even correct, versus just plausible

We've watched teams chase deflection rate for two quarters and then get blindsided by a churn spike they can't explain. The support dashboard says everything's fine. The NPS comments say otherwise. That gap is deflection rate doing its job — hiding the actual outcome.

## What we track instead at Peeve

Because Peeve [executes real backend actions](/solution-actions/) (refunds, plan changes, password resets, whatever the product allows) instead of just generating text, we can measure something deflection rate can't touch: **did the underlying state actually change in the way the customer needed?**

That gives us a different scoreboard:

- **Action completion rate** — of the cases where the fix required a backend action, how many actually got the action executed versus just explained
- **Recontact rate on the same issue** — not "did they open a new ticket," but did the same root cause resurface within 7 days
- **Escalation quality** — when Peeve hands off to a human, does the agent get full context (order history, what Peeve already tried, why it stopped) or does the customer have to re-explain everything from scratch
- **Time-to-actual-resolution**, measured from first contact to state change, not from first contact to "conversation closed"

None of these numbers look as good on a slide. Action completion rate especially — it's humbling, because it forces you to admit how often the AI *should* have acted but instead just talked. That's a design failure, and deflection rate would've hidden it completely.

## Why this changes how the agent is built, not just how it's measured

If you're optimizing for "conversation ends without escalation," you build an agent that's good at sounding done. Hedged language, polite non-answers, gentle redirects to a help article. It's the AI customer support equivalent of a store employee pointing vaguely toward aisle 12 instead of walking you there.

If you're optimizing for actual resolution, you build something different: an agent that's willing to say "I can't fix this, here's a human who can, and here's everything they need to know already" — and treats that as a *good* outcome, not a failure metric. Escalation isn't the enemy of good AI customer support. Confident wrongness is.

This is also why we think agents that only live in a chat window are structurally limited. If Peeve can [see the actual screen state](/solution-deflect/) and execute the action in place — cancel the subscription, apply the credit, update the shipping address — there's no ambiguity about whether the problem got solved. The state either changed or it didn't. Chat-only tools have to take the model's word for it, which means the whole system is built on a foundation of "trust me."

## The uncomfortable tradeoff

Being honest about action completion rate means our numbers are, on paper, sometimes worse than a competitor's deflection rate. We're fine with that. A vendor reporting 75% deflection and a vendor reporting 55% action completion aren't comparable — they're measuring different things, and only one of them tells you what actually happened to the customer.

If you're [evaluating AI customer support tools](/compare/) right now, ask vendors this instead of the deflection number: *of the tickets your AI closed last month, how many customers contacted support again about the same issue within a week?* Most won't have that number ready. That's the tell.

Support leads don't need an AI that looks resolved on a dashboard. They need one that makes the recontact rate go down, quietly, over months, because the thing actually got fixed the first time. That's a slower story to tell in a demo. It's the only one worth measuring.