Answering a customer well is only half the job. The other half is noticing when an answer didn't land, and not making the same mistake again. Alongside learning facts from your team (covered in "Arbyn learns from your team"), Arbyn quietly keeps an honest record of how its own conversations actually turn out, and uses the recurring failures it finds to change its own behavior on future conversations, without you having to do anything.
What it does
For every reply Arbyn sends, it later checks what happened next: did the customer's next message read like the question was answered, did the conversation just go quiet, did your team have to step in, did the customer rate it well or badly, did your team thumbs-down the reply, or did they edit it before it went out. None of this changes anything about the conversation itself. It's purely a record, kept privately, of what actually happened.
Where this becomes useful is when the same kind of failure shows up again and again. Arbyn groups its own recurring failure patterns (a particular kind of question it keeps escalating, a reply it keeps having to repeat because the first one didn't land, a topic where customers keep rating it badly) and turns each pattern into a specific instruction for itself: something concrete to do differently the next time that exact kind of question comes up.
Where to find it
Open Knowledge → Learned facts, then the Behaviour tab. This shows the failure patterns Arbyn has found, how often each one has happened, and when it last happened. It's shown as the actual instruction Arbyn gives itself, not a raw number, so you can read exactly what it's trying to change rather than decode a label.
How it works
Every agent reply is tracked to one outcome: the customer replied normally, the conversation resolved and stayed resolved, a teammate stepped in, a customer rating came in (good or bad), your team gave the reply a thumbs-down, your team edited the draft before sending it, or the conversation simply went quiet with no further reply from the customer. This classification is plain rules and record-keeping, not guesswork, and it happens automatically in the background after every reply.
From there, Arbyn looks across many conversations at once to find the ones that keep recurring: the same kind of question ending in an escalation over and over, the same kind of reply your team keeps having to rewrite, a topic where customers keep going quiet instead of confirming they were helped. A single bad outcome is just a bad day and isn't treated as meaningful on its own; it takes a genuine, repeated pattern (several occurrences of the same shape of failure) before Arbyn treats it as something worth changing.
Once a pattern is established, Arbyn turns it into a concrete instruction that gets added to how it thinks through future conversations of that same kind, for as long as the pattern keeps happening. Concretely, that means things like: "you've ended conversations like this without the customer confirming they were helped, so before treating this as resolved, ask whether that answered their question," or "you've repeated yourself instead of moving the conversation forward on this topic, so try something different this time: look something up, ask one specific question, or hand off, rather than saying the same thing again." These are Arbyn's own measured patterns on your specific store, not general advice, and they're scoped to the kind of question they came from, so a mistake on shipping questions doesn't change how Arbyn handles refund questions.
If a pattern stops recurring for a while, it's retired rather than kept forever, since re-warning Arbyn about something it's already stopped doing would waste effort and could even nudge it towards the wrong lesson.
Good to know
This never changes what a customer sees directly. It only ever changes how carefully Arbyn approaches the next conversation of a similar kind, not the one it already answered.
A single bad outcome isn't treated as a pattern. It takes several occurrences of the same shape of failure before Arbyn adjusts its own behavior because of it. One rough conversation is not enough to change anything.
It stays private to your store. Outcomes are derived only from your own conversations, orders, and ratings, and nothing about a specific conversation is shared anywhere outside your Arbyn account.
Your store still comes first. As always, Arbyn answers from your real orders, inventory, and policies. What it learns from outcomes changes how carefully or differently it approaches a topic, never what the actual facts are.
It's deterministic, not a black box. Every classification is built from concrete, observable signals (a reply, a rating, a handoff, a thumbs-down), not from a model's opinion of how a conversation went.
FAQ
Does this change the replies Arbyn sends to customers? Yes, indirectly. Arbyn doesn't rewrite a conversation that already happened, but a genuinely recurring failure on a specific kind of question does change how carefully or differently Arbyn approaches the next conversation like it.
Do I need to turn anything on for this to work? No. It runs automatically for both chat and email conversations, with nothing to configure.
Does it use my customers' personal data? It uses the shape of what happened in a conversation (a reply was followed by silence, a rating came in negative) to spot patterns. It doesn't build advertising profiles of shoppers or share anything about a specific customer outside your Arbyn account.
How is this different from the Learned facts tabs? Learned facts are specific answers Arbyn should give, usually taught by a human action on your team. This is about patterns in how Arbyn's own replies are landing, including mistakes it's noticed itself and is actively working to stop repeating, whether or not anyone on your team was directly involved in that particular conversation.
Can I see exactly what it's changed? Yes. The Behaviour tab shows every active pattern as the actual instruction Arbyn is following, so you can read precisely what it's trying to do differently and why.