Deflection Rate vs Resolution Rate: The Support Metric That Flatters, and the One That Doesn’t

The short answer#

Deflection rate is the share of support conversations that never reached a human. Resolution rate is the share of customer problems that were actually solved. They are different numbers because a conversation can end without a person and without a solution — the customer gave up, got a plausible wrong answer, or came back tomorrow through a different door. Deflection is easy to measure and flattering; resolution is hard to measure and honest. A vendor quoting one as if it were the other is not lying, exactly, but they are letting you hear the better word.

The practical rule: any percentage you are shown should come with its numerator, its denominator, its time window, and a statement of who observed the outcome. If it comes with none of those, it is a story wearing the costume of a measurement.

What is deflection rate?#

The fraction of conversations an automated layer handled without a human agent being involved.

deflection rate = conversations with no human involvement ÷ all conversations, over a window

What it can see: whether a person was pulled in. What it cannot see: whether the customer's problem went away. A customer who asked, got an answer, and closed the tab is "deflected" whether the answer was right, wrong, or a polite refusal they walked away from. The metric measures the absence of an escalation and nothing else.

That makes it a real number about cost — it tells you how much human time the layer is absorbing — and a poor number about quality.

What is containment rate?#

The same idea measured inside one channel. A conversation that starts in the chat widget and ends in the chat widget is "contained", even if the same customer emails twenty minutes later. Containment is the channel's own view; deflection is usually the portfolio's view across channels. The names are used interchangeably by many vendors, so ask which denominator they mean.

What is resolution rate?#

The fraction of problems that were solved, end to end: the answer was correct, any action needed was taken, and the customer did not need to come back.

resolution rate = problems confirmed solved ÷ problems raised, over a window

The word confirmed is where the difficulty lives. Something has to observe the outcome. In practice there are three ways to observe it, in descending order of honesty:

1. The customer says so. A "did this solve it?" that is answered yes. Reliable when answered; most customers do not answer.
2. No return within a window. The customer did not reopen, re-ask, or contact through another channel within, say, seven days. A proxy, and a decent one if you can actually see the other channels.
3. The absence of an escalation. The customer did not ask for a person. This is deflection wearing resolution's name, and it is the one to watch for.

The third is where most "resolution" claims come from, because it is the only one a chat layer can count without help from anywhere else.

A worked example: 80% deflection, 50% resolution#

One hundred conversations in a month. Here is what happened to them and what each metric calls it.

Deflection rate: 80 of 100. Resolution rate: 70 of 100 if you count the handoffs the humans solved, 50 of 100 if you are measuring the automated layer alone. Same month, same customers, and a number that can be quoted as anything from 50% to 80% depending on the question you decide it answers.

Notice the fifteen wrong answers. They are the reason deflection flatters: the layer's worst outcome — a confident, wrong reply about your own product — is counted as a success, because nobody escalated. Improving the honesty of the agent, so that those fifteen become refusals and handoffs, would lower the deflection rate while making customers better off.

What a support agent can honestly count on its own#

A chat layer sitting inside your product can see three things with certainty: that a question was asked, whether it admitted it could not answer, and whether a handoff was recorded against that admission. It cannot see whether the problem went away.

This is the view hiy takes, and it is worth stating plainly because it is the opposite of the industry's habit. The two rates at the top of a hiy agent's Insights are 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 — and deflection rate, its complement: conversations where the agent never said it couldn't answer. Deflection there is defined as the absence of an admission, and the docs say in as many words that it is not the presence of a solution. Nothing in the product is called a resolution rate, because nothing in the product can observe one.

Three rules make those two numbers quotable rather than decorative:

• Every percentage carries the two numbers it is made of. 20 of 25 conversations · last 30 days, inside the same tile, so you can check the arithmetic and see how big the sample was.
• Below twenty conversations, no percentage at all. A rate off a handful of people swings five points when one of them changes their mind. Until twenty, the tile shows a dash and how far along you are.
• A busy window says so rather than guessing. If there are more conversations than one read can total exactly, the tile says that, and no rate is stated. A rate counted from part of a window is not slightly off; it is unknown.

The full definitions, including why clearing the queue of unanswered questions never moves a past rate, are in deflection and escalation. If you want resolution — customer confirmation, reopen tracking, a helpdesk that sees email as well as chat — that is analytics hiy does not ship today; it is on the roadmap, and until it ships the honest thing is to say so rather than relabel deflection.

Five questions to ask about any vendor's number#

1. What is the numerator? "Resolved" meaning the customer confirmed it, or meaning nobody escalated?
2. What is the denominator? All conversations, or only the ones the bot chose to engage with? Excluding "out of scope" from the denominator is the oldest trick in the category.
3. What is the window? A rate over a launch week is not a rate over a quarter.
4. Is there reopen data? If the same customer came back within a week, was the first conversation still counted as resolved?
5. Who observed the outcome? The customer, another system, or the absence of a complaint?

A vendor who answers all five clearly is one you can compare with another. A vendor who answers with a bigger number is giving you a different metric.

The honest version#

Resolution rate is the number everyone wants and nobody inside a chat widget can measure without help. The honest options are to instrument the confirmation yourself — a one-tap "did this help?" that most customers will ignore, plus a reopen check across every channel you run — or to accept deflection and escalation as what they are: numbers about where the work went, not about whether the work was done.

And a high deflection rate is not a goal. The agent whose deflection rate is highest is the one that never admits a gap. That agent is inventing your product.

Where to start#

Ask your current tool the five questions and write the answers down; you will learn what its headline number means. If you are setting up a hiy Support Agent, the two rates appear in Insights on their own once twenty conversations have happened — with their counts and window beside them — and the questions it could not answer wait in the queue below, which is the part of the number you can actually act on.