When a chat closes in Chatixy it receives exactly one of five outcomes, and only two of them, confirmed and assumed, count toward the AI's resolution rate. The other three (human assisted, unresolved and abandoned) are recorded and kept out of the wins. The outcome is decided once, at close, in a fixed order of precedence, so a chat that fits two descriptions still lands in one place.
We built it this way because a resolution rate is usually one figure that mixes "the visitor said this helped" with "the visitor stopped typing". Those are different kinds of evidence, and a person reading your report cannot tell them apart unless you keep them apart at the source.
If you have to put that number in front of someone, run the checks below against your own closed chats first. Each one names a transcript to look for in the shared inbox and the outcome it should carry. If the metric itself is new to you, What is ticket deflection, and how do you improve it? covers the ground underneath it.
- `human_assisted`: What puts a chat there: A person replied or was assigned; Counts toward the AI rate: No
- `unresolved`: What puts a chat there: Thumbs-down or negative sentiment; Counts toward the AI rate: No
- `resolved_confirmed`: What puts a chat there: Thumbs-up or closing gratitude; Counts toward the AI rate: Yes
- `abandoned`: What puts a chat there: AI never answered, or visitor left mid-question (open loop); Counts toward the AI rate: No
- `resolved_assumed`: What puts a chat there: AI answered, nothing negative, went silent; Counts toward the AI rate: Yes, shown separately
1. Silence is labelled assumed and never merged into confirmed
How to check: find a closed chat where the AI gave an answer and the visitor never wrote again. No thumb, no thanks, no complaint.
What good looks like: the outcome is resolved_assumed. Three things have to be true for it: the AI answered, nothing negative happened, and the visitor went silent.
A transcript that ends after an answer is consistent with two stories. In one the visitor read it, got what they needed and closed the tab. In the other they read it, decided it was useless and closed the tab. Nothing in the text separates them. So the chat counts toward the rate, but under its own label, and the assumed total is shown separately from the confirmed one.
2. A visitor who left mid-question is abandoned, not resolved
How to check: look for two kinds of chat. In the first, the last message is the visitor's and the AI never answered it. In the second, the conversation was still in motion when the visitor left: a question half asked, a loop still open.
What good looks like: both are abandoned, and neither counts as a win.
This is the check that separates an honest rate from a flattering one, because abandoned and assumed look alike from a distance. Both end in silence. The difference is whether the loop was closed before the silence started. If the AI had answered, silence is at least compatible with success. If the question was still open, silence is only a walk-out.
3. Confirmed requires a signal from the visitor, after the answer
How to check: open a few chats marked resolved_confirmed and find the evidence. It should be one of two things: a thumbs-up, or a closing thank-you.
What good looks like: in every one, that signal comes after the AI answered. Order matters here. Gratitude that closes a conversation is a response to the answer. A "thanks in advance" attached to the opening question is politeness, and the rule is not looking for it.
Because confirmed has this narrow definition, the confirmed share is a figure you can quote by itself without a footnote. Each chat in it has a visitor who said the answer helped, and none is there because of what a visitor did not do.
4. A thumbs-down outranks the fact that an answer was given
How to check: find a chat where the AI did answer, with sources, and the visitor pressed thumbs-down or the conversation turned negative.
What good looks like: unresolved. Producing an answer is not the same as resolving the question, and a visitor telling you so is the strongest evidence you will get in either direction.
These chats are also your content backlog. An unresolved outcome on a chat where the AI answered usually means the source page was thin, outdated or written for a different question. How to Write FAQ Answers an AI Support Agent Can Actually Use covers how to fix that.
5. One human touch moves the chat out of the AI's column
How to check: find a chat where a teammate replied, and another where someone was assigned without replying.
What good looks like: both are human_assisted. That includes the awkward case where the AI had already given a correct answer and a teammate added one line to be helpful. A chat a person touched is never counted as an AI win.
This costs the AI some credit it arguably earned. We accept that, because the alternative is arguing chat by chat over who really solved it, and a number built from arguments is not one you can report. Human-assisted chats still matter, but the place to judge them is the handoff itself, which AI-to-human handoff: 6 best practices for support covers.
6. A customer who comes back undoes the resolution
How to check: find a customer who returned after their chat had closed as resolved.
What good looks like: the earlier resolution is counted as reverted. It does not stand as a clean win.
A close is a claim about the future, namely that the customer will not need to come back about this. When they do, the claim was wrong. Without a reverted count, the cheapest way to raise a resolution rate is to close chats quickly. With one, a quick close that did not hold shows up as exactly that.
7. The reported figure has two parts and a narrow denominator
The definition, as written in our ticket lifecycle doc:
honestairesolution_rate = (confirmed + assumed) / (AI-only closed)
How to check: work out what sits in the bottom of that fraction. "AI-only closed" means closed chats that no person replied to or was assigned to. Since every closed chat has exactly one outcome, that is the four outcomes other than human assisted.
What good looks like:
- Human-assisted chats are absent from both the top and the bottom. They neither help nor hurt the AI's rate.
- Abandoned and unresolved chats are in the bottom and not the top. Walk-outs and thumbs-downs pull the rate down, as they should.
- Chats that are still open are not in it at all.
Then report it in parts. Give the rate, then the confirmed share and the assumed share that add up to it, then the reverted count. A reader who sees a rate made mostly of assumed chats knows it rests on silence and can weigh it accordingly. A rate made mostly of confirmed chats needs no defence.
The single blended figure is the one to avoid. It may well be accurate, but nobody reading it can tell which of those two situations they are looking at.
Run it on your own chats
To see these outcomes on your own traffic, add your website address, customize the support agent and watch it answer in minutes. Pricing lists plans, limits and yearly discounts. For demos, Business plans and volume pricing, email info@chatixy.com.