Where the idea came from
Earlier this year I sold my shares in HeliumDeploy. The habit I kept was interviewing customers. I'd always done it: who bought, who almost did, what they were trying to do, what got in the way. Something happened almost every time. A customer would say one sentence and we'd have an edge we didn't have that morning: a better way to position a product, an angle for an ad, a product we should carry and didn't. After I left, I built an AI that does the calls, and with HeliumDeploy as the first store to try it, it ran 137 fifteen-minute conversations with their customers.
Reading them back, most of what people wanted was pretty ordinary. They bought accessories on Amazon and would have preferred one box from the store. They left a product page to find out one fact that could have been printed on it. They waited on an order and nobody told them the date had moved. Good information, and it had been sitting in the store's own conversations the whole time. That's the thing I kept noticing, there and at other stores: Shopify stores don't know their customers. Not because they don't care, because nobody has the hours.
So, Mia
Mia sits on a store's product pages. The owner clicks one install link, and about forty seconds later she has read the catalog and is ready to talk to shoppers: help them pick, compare two products, hand over the link, and where the store allows it, look up an order. Every conversation shows up on a page the owner can read whenever they like.
Training her is meant to be casual. The owner writes how they sell in plain words, and she goes with it from the next chat. If they use Claude Code or Cursor, one line connects their own coding agent to her, and it can read her conversations and teach her directly.
I built it with Claude Code holding the keys: Shopify, the hosting, the Anthropic platform, email. It wrote the service, the widget, the install flow and the tests that run against a real store page. I say what I want, look at what comes back, and we go again. The first version of the install page asked owners to paste code into their theme, which no store owner is going to do, so that went.
customer conversations run by the AI interviewer, read back for what people actually asked for
to install Mia; she sets herself up from the catalog in about forty seconds
stores trying it before launch: an audio retailer and a luggage brand, both in Canada
What I'd say I learned
The store usually knows what to fix. It's in the questions shoppers ask and the reasons they leave. Building the software has gotten cheap. Picking what to build, and for whom, is still the job.