I ran customer support for 60 Shopify stores. Then I built the tool I wanted.

For seven years my job was other people's customer inboxes. Replo, a Shopify GDPR app, and an agency where I was on support across roughly 60 Shopify stores at once.

Sixty stores, different brands, same five questions every single morning. Where is my order. I typed the wrong address. Can I return this. Do you ship here. Is this in stock in medium.

Support does not break because the questions are hard. It breaks because the volume never stops, and because every real fix means a human opening the Shopify admin and doing something. Answering "yes I can change that address" does not change the address. Someone still has to go change it.

The part that bothered me most was the pricing. Nearly every tool in the category charges per ticket, per AI resolution, or per agent seat. The bill grows every month the store grows. Store owners get charged more for succeeding, and more the better the automation works. That is a strange thing to sign up for when the problem you are solving is volume.

So I built Arbyn. It is live on the Shopify App Store now.

It answers on support email and live chat in the store's own voice, pulling live Shopify order data. Then it performs the action instead of stopping at the reply. Shipping address changes it handles on its own. Cancels, refunds, discounts, gift cards, reships and returns wait for the store owner to tap approve, and then Arbyn executes the real Shopify mutation and writes it to the order timeline. Anything that touches money waits by design, because "the AI refunded someone at 2am" is not a message I ever want to wake up to.

It sells in the same conversation too. In-chat upsells, bundles, guided quizzes that hand the shopper a pre-filled cart, and revenue attribution on every order it sources.

Pricing is flat. Free forever at 150 AI conversations a month, $59 for 500, $99 for unlimited. Every plan includes every feature. Nothing per ticket, nothing per resolution.

Two things I got wrong worth passing on.

I underestimated how much of the work is guardrails rather than the agent itself. Shadow mode, confidence thresholds, refund ceilings, escalation triggers, kill switch. The model was the easy part. Making it safe enough to trust with a stranger's money is where the months went.

And I let a channel appear in my marketing copy that was not actually built yet. It got picked up and read as fact before I caught it and pulled it. Now nothing goes into copy unless I can point at the code that does it.

Where it stands: shipped, on the App Store, early, and I am the whole go-to-market side of it.

Happy to answer anything about building on Shopify, the App Store review process, or what running support at real volume actually looks like.