---
title: "Deflection and escalation"
description: "The two rates at the top of Insights — what each one counts, and what neither of them claims."
source: https://hiy.ai/docs/deflection
---

# Deflection and escalation

Two rates sit at the top of **Insights**, above the activity chart. They are
the numbers a support agent is usually judged on, and they are counted over the
same conversations, from what the server itself recorded — never from anything
a visitor's browser said happened.

## What each one counts

- **Escalation rate** — conversations where the agent said it couldn't answer,
  or where a handoff was recorded against an answer it admitted it couldn't
  give.
- **Deflection rate** — conversations where the agent never said it couldn't
  answer, as the server saw it.

They are complements over the same set of conversations: the {rateWindow} of
conversations in which somebody asked your twin something. The window is
printed beside each rate, so the number and the period it covers can never come
apart.

The sentence under each rate on screen is the same one this page is written
from. If a caption here and a caption there ever disagreed, one of them would
be claiming something hiy doesn't measure — so there is one wording, kept in
one place.

## Every percentage carries the two numbers it's made of

A rate is never shown on its own. Under it sits the count it came from —
*20 of 25 conversations* — so you can check the arithmetic and see how big the
sample was. A percentage whose numerator and denominator are hidden is
something you have to take on trust, and these are meant to be numbers you can
quote to somebody else.

## Below twenty conversations, no percentage at all

Until twenty conversations have happened in the window, neither rate is stated.
The panel says **Not enough conversations yet**, and how far along you are —
*7 of 20 conversations so far*.

That is the state most twins are in, and it isn't an error. A rate off a
handful of people is a story wearing the costume of a measurement. Twenty is
where one conversation stops being able to move the headline by more than five
points; below it, a single person swings the number. Rather than print
something that will swing and then settle, hiy says what it has so far and
waits — the count is real, and it is going up.

## A busy window says so, rather than guessing

At the other end, a window can hold more conversations than one read can total
exactly. When that happens the panel says so — *too many conversations in this
window to count exactly* — and states no rate.

It is deliberately not an approximation. A rate counted from part of a window
isn't a slightly-off rate, it's an unknown one, and there would be nothing on
the screen to tell it apart from the truth. So the honest answers are a
percentage or a reason there isn't one, and never a number with a caveat.

## What these rates do not tell you

**They do not say whether anyone's problem was solved.** hiy has no way of
observing that and will not invent one. Deflection counts conversations where
the agent never said it couldn't answer — that is the absence of an admission,
not the presence of a solution. Nothing here is a resolution rate, and it is
never described as one.

**Deflection is not a claim that nobody was handed off.** A handoff is
something the visitor's browser reports, and these counts never read it. A
conversation where somebody asked for a person, but where the agent never
admitted it couldn't answer, is counted as deflected. Your **People** list and
your Inbox are where handoffs are recorded — see [Where a lead came
from](/docs/conversations). When your agent hands over, who that reaches and
what the visitor reads are yours to set: [Handing over to a
person](/docs/escalation). Changing those settings never moves these rates.

**They aren't a score you can improve by tidying up.** The escalation count
includes the questions your teach queue recorded, counted at the moment they
were raised. Clearing your queue doesn't move a past rate, because a number the
person being measured can change by pressing *dismiss* isn't a number worth
printing.

## The questions behind them

Under the two rates, Founding shows the questions your agent recorded a gap
for, grouped the same way your teach queue groups them — so you can see what
your twin keeps getting asked and can't answer yet, and go
[teach it](/docs/sources).

That list is **part of what drove the rates, not all of it.** When your agent
says it can't answer, hiy records that it happened; it does not store the
wording of every question that led to it. So a conversation can count toward
the rates above without appearing in the list below them, and the list never
claims to be the whole story.

## Where the counting happens

Everything on this panel is counted server-side from your own twin's records:
the questions people asked it, the times it said it couldn't answer, the gaps
it raised for you, and handoffs recorded against an answer it had already
admitted it couldn't give. Nothing is modelled, nothing is estimated, and no
number here comes from a visitor's browser asserting something happened.