Shopify is home to millions of merchants across every category imaginable: first-time sellers, established brands, dropshippers, retail businesses, social sellers, you name it.

Helping such a varied group get set up has always been a challenge, and for most of Shopify's history, the answer has been some version of the same thing: ask merchants a few questions during signup, and then use their answers to shape the guidance they receive when they land in the admin.

Dropshippers get help sourcing products, replatformers get tips for migrating their business. It's a reasonable system, but the guidance at the center of it has always been static, hard-coded copy. That guidance is what this project set out to replace.

The onboarding guidance for a merchant who identified as a dropshipper.

Hard-coded guidance has real limits. It goes stale the moment the product changes, every update has to be re-translated across supported languages, and if you want to show different guidance to different kinds of merchants, you have to write and maintain a separate version for each one.

And even then, a deeper problem remains: static guidance can't adapt to the individual.

Two merchants in the same segment can have completely different experience levels, different products, questions, needs. No single piece of guidance can work equally well for all of them. And the bigger and more varied your user base gets, the more this compounds. At Shopify's scale, it was already breaking.

Earlier in 2025, I was working on a project exploring whether conversational AI could replace Shopify's signup questionnaire. At some point during that work, I started thinking more broadly about how generative AI might be applied to onboarding guidance.

By that time, Shopify's AI assistant, Sidekick, was already inside every admin. It could understand a merchant's context, answer questions accurately, take actions on their behalf, and it was improving quickly. Using it to generate onboarding guidance, instead of serving the same fixed copy to everyone, started to feel obvious.

So I tested the idea. I built a lightweight prototype in Cursor that simulated how Sidekick might use a merchant's signup context to generate guidance for each task, recorded a short video making the case for it, and shared it with our Director of Design. She approved it, and I took it on as lead designer. We set a clear hypothesis going in, that personalized, capable guidance would lift onboarding completion, and agreed up front on how we'd measure it.

An early proof of concept showing AI-generated task guidance in place of hard-coded copy.

With the project approved, the real design challenge came into focus. We were taking the entire experience of onboarding guidance and rebuilding it inside a chat medium. Everything a merchant relied on to get set up, the instructions, the sense of where they were, the feeling of progress, now had to work inside a conversation with an AI. That was the core product design problem, and it was a large one, spanning prompting, conversation design, and a lot of engineering to orchestrate it all.

I started by asking what guidance should even feel like when it's no longer a static message but a live conversation. How should Sidekick behave when a merchant opens a task for the first time? How should that change as they move through different states of a task? And how does a multi-step process like onboarding stay legible inside a conversation, when conversations are linear and onboarding isn't?

Those questions sorted into two streams of work. The first was how Sidekick should communicate with a merchant, across onboarding as a whole and on any individual task. The second was how to give a multi-step process structure and cohesion inside a linear chat. They called for different kinds of thinking, so I'll take them one at a time.

Designing what Sidekick says

Our old, static guidance had to work for every merchant within a segment, so it was broad by necessity. All it could really do was describe the task in a general sense.

Generic guidance for the add your first product task.

Moving guidance into Sidekick meant we were no longer bound to that format. We weren't writing one broad block of copy for a whole segment, we could shape the guidance to the merchant in front of us. But the bigger shift was that Sidekick didn't have to describe the task at all. Rather than telling a merchant how to add their first product, it could help them actually do it.

Sidekick can generate a full product listing from a description and image.

That changed the question we were designing around. If the first message could do almost anything, what should it do?

It had to orient the merchant, first of all. They needed to understand the task, what finished looked like, and where to begin. Showing off what Sidekick could do was useless if the merchant didn't know what they were there to accomplish in the first place.

It also had to teach the merchant what Sidekick was. For most new merchants, opening that first task would be the first time they'd ever used the tool. So the message wasn't just about explaining a task, it was about setting expectations for Sidekick itself.

And it had to be efficient. Every turn in a conversation costs something, in compute and in the merchant's patience. Guidance that took five exchanges to get someone moving was worse than guidance that delivered in the first one.

These pressures pointed in the same direction. The most useful thing the first message could do wasn't to walk a merchant through a task step by step. It was to give the task shape and show the specific ways Sidekick could shorten it, by offering to write the description, suggest titles, or generate images.

Testing early variations of how Sidekick opens on onboarding tasks.

Sidekick's guidance also had to adapt to where the merchant was in a task. A task isn't a single moment, it's a series of states. A merchant might be opening one for the first time, returning to one they'd started, or coming back to one they'd already finished. Each state called for a different message, so Sidekick's guidance had to be stateful in the same way the tasks were.

With these principles in place, I mapped every state a merchant could be in when they opened a task, wrote prompts and examples for each, and worked with engineering to refine them until they held up in real situations.

Giving the conversation structure

The second area was making the onboarding journey legible inside a conversation. Merchants already had a persistent view of their tasks on Home, where they could see everything at once and track progress. Inside Sidekick, nothing like that existed. And onboarding can take anywhere from a few hours to a few weeks to complete, so a merchant needed to be able to leave and come back without losing their place.

The problem is that chat alone isn't built for this. A conversation is linear, with no fixed structure. If we handled onboarding entirely through messages, a merchant would have no way to see the full set of tasks at once. To check what they'd finished or what was left, they'd have to scroll back through the conversation.

So we needed to add that structure ourselves. Something that could show the full set of tasks at once, track which ones a merchant had finished, and stay in the conversation without getting in the way.

That's what we built. Working closely with the Sidekick team, I helped design a new to-do list component that renders when a merchant opens their first onboarding task.

The full to-do list renders in chat after clicking on an onboarding task.

When it first renders, the card shows the full set of onboarding tasks, so a merchant can see everything they need to do in one place. Once they start a task, the card collapses to keep the conversation clean, but it stays pinned at the top, so it's always one tap away.

Hovering over the collapsed component expands it.

As the merchant completes tasks, the to-do list marks them off in real time, so the card always reflects where they actually are and what's next. And because it lives in the conversation, a merchant can leave and come back days later to find it right where they left it, still showing what's done and what's next.

When a merchant saves their store name, the task is marked complete.

Rather than building a new component, the simpler option would have been to let each task spin up its own conversation. It would have taken a fraction of the effort. The problem is that navigating Sidekick's conversation history isn't intuitive today. A merchant who worked on a task, closed the conversation, and came back later would struggle to find their way back to it. It was a small thing, but it was the kind of paper cut that adds up across an onboarding flow.

The new component also scaled beyond onboarding. A way to give multi-step workflows structure inside a conversation serves Shopify and its merchants well past setup, which made it worth building as reusable infrastructure rather than a one-off.

That broader value is also why the Sidekick team was the right partner. We needed it for structured onboarding, they needed it for multi-step workflows in general, and we built and tested it against both.

When a merchant onboards to Shopify today, the experience looks nothing like it used to. Sidekick opens when they start a task, takes them to the right place in the admin, and gives them guidance built for their situation. They can ask questions, go deeper on anything, leave, and come back a week later to pick up where they left off. The to-do list keeps the whole journey visible inside one conversation, so nothing gets lost.

The thing the old system could never do, adapt to the individual merchant, is now the thing it does best. It's the difference between being handed a manual and having someone walk you through it.

The bet paid off. The hypothesis was that personalized, capable guidance would lift onboarding completion, and it did, by 5%. The shift from static to generative guidance has been one of the most impactful things our team has shipped, driving $25M in gross profit alongside that lift.

But the number that changed how we think about our work was Sidekick adoption. Before this project, most new merchants never sent it a single message. After launch, new merchant engagement went up 64%.

It also changed how our team works. We don't write and publish guides anymore. We write prompts, ship them, and measure what happens. The iteration speed is a different order of magnitude, and we've already run experiments that would have taken months under the old system.