You don't have to train Arbyn. You don't have to write anything down. Every time someone on your team corrects a reply, edits a draft before it goes out, or steps in to answer a customer themselves, Arbyn quietly remembers what happened, and the next customer who asks something similar gets the better answer. This is on from your very first conversation.
What it does
Arbyn watches what your team actually does, not what they say they would do, and turns the genuine, human moments into knowledge it can use later. It never watches a rating a customer left without you being involved, and it never learns from its own drafts. It only learns from a real person: your reply, your edit, your note, or a customer's own rating of something a human already touched or that closed the loop.
Where to find it
Open Knowledge, then Learned facts. This is the Learning hub: a Pending approval tab for anything Arbyn picked up that hasn't been reviewed yet, an Approved Q&A tab for facts your team has signed off on, a Customer facts tab for things Arbyn has noticed about individual shoppers, and a Behaviour tab covering the voice Arbyn has picked up and the mistakes it's actively avoiding (more on that in a moment). There's also a Teach Arbyn button at the top of the page, the one deliberate way to teach it something directly.
How it works
Most of what Arbyn learns comes from things your team is already doing as part of their normal job, not a separate task:
You reply to a customer yourself, with no draft involved. Your reply becomes the answer.
You edit one of Arbyn's drafts before sending it. The edited (sent) version becomes the answer, never the original draft. The edit is the correction; that's the part worth keeping.
You leave a handoff note when handing a conversation back or closing it out. Your note becomes an explicit instruction for next time.
You thumbs-up or thumbs-down a reply, whether it's already been sent to a customer or it's still sitting as a draft waiting for your review. A thumbs-down with a short note of what was wrong is one of the clearest signals Arbyn gets. A bare thumbs-up still counts, just at lower confidence than something you actually wrote, since a click is a weaker signal than a sentence.
You resolve an escalated conversation with an actual decision (granting an exception, explaining a ruling). The decision becomes reusable knowledge for the next similar case, not just something that happened once and was forgotten.
A customer rates a conversation 5 stars. The reply that earned that rating is treated as proven to work, and kept.
You use "Teach Arbyn." This is the only one you have to do on purpose: type a question and the answer you want given, and it's saved immediately at the highest possible confidence, since you're stating it outright rather than Arbyn inferring it from something else.
Everything else on that list is just your team doing their job. Arbyn is simply paying attention.
It learns by meaning, not by matching words
This is the part that actually makes it useful. Teach Arbyn the answer to "do you ship to Canada?" and a later customer who asks, in completely different words, "can I get something sent up north?" will still get your answer. Not one word matches between the two questions, but Arbyn understood they mean the same thing. You never have to anticipate every way a customer might phrase something. Answer once, in your own words, and Arbyn carries it forward to every phrasing that means the same thing.
It also learns how you write, not just what you know
When your team replies or edits a draft, Arbyn doesn't only keep the facts in it. It also picks up on how your team actually writes: reply length, tone, whether you open with an apology, how you sign off. The next time it drafts a reply on a similar topic, it draws on real examples of how your team has actually answered that kind of question before, rather than a generic description like "friendly and professional." Ask about a late parcel and you'll tend to get your late-parcel voice; ask about a return and you'll tend to get your returns voice.
What it will not learn from
This is deliberate, and it's the guardrail that keeps the whole system honest: Arbyn never learns from its own drafts. A reply that was never reviewed, approved, edited, or acted on by a real person is never treated as a fact. An assistant that learned from its own unreviewed output would drift a little further from the truth every time, and nobody would notice until it was badly wrong. So the loop only closes when a human is genuinely in it, whether that's your team or a customer reacting to something a human already touched.
Good to know
You don't have to review or approve anything for it to start working. Facts Arbyn extracts on its own land in Pending approval so you can check them, but Arbyn is learning from day one regardless.
You can see everything it has learned. The Learning hub shows every fact, where it came from, and how confident Arbyn is in it. You can edit, approve, unapprove, or delete anything.
Your correction always wins. If Arbyn learned something wrong, delete it or just answer that question yourself once; your answer replaces it.
What Arbyn learns at your store stays at your store. It never shares what it learns with any other business using Arbyn.
FAQ
Do I have to train Arbyn before it's useful? No. It answers from your real store data from day one, and it gets better every time your team touches a conversation, without any setup on your part.
Can I see what it has learned? Yes. Open Knowledge → Learned facts to see every fact, where it came from, and how confident Arbyn is. Anything you don't want kept can be edited or removed.
Does a thumbs-down actually change anything? Yes. A thumbs-down with a short reason is one of the strongest correction signals Arbyn gets, whether it's on a reply that already went out or a draft still waiting for your review.
Does it learn across other stores? No. Everything Arbyn learns from your team and your customers stays private to your store.
What's the difference between this and the Behaviour tab? This page is about facts, the actual answers Arbyn should give. The Behaviour tab is about the patterns Arbyn has noticed in how its own replies land, including the mistakes it's been making and actively working to stop repeating.