Running your online store with AI: creating products, building pages and maintaining content from one workspace
Nick van der Blom5 min read
Most online stores use AI for product copy and the occasional image. That is step zero. The question owners ask me is: "How do you actually use it effectively?"
The answer is not a better tool. It is a different way of working: running your store from one AI workspace that knows your business, does the work on a schedule, and only calls on you where it matters. This article describes that route in five steps, with what is realistic at each step and what is not.
What "running it with AI" does and does not mean
Not: a chatbot that answers your questions. Also not: a button that generates a hundred pages.
Yes: a workspace in which AI agents have access to your product data, your house style and your goals, and use them to do recurring management work. Creating products from supplier data. Writing category pages that match what you actually sell. Designing pages within your design system. Reviewing and improving existing content. And not once, but every week.
The basis for that is Claude Code, an AI agent that works on your computer or server and is connected to your systems. What that is exactly is in our explainer on Claude Code; here it is about what you do with it.
The staircase in five steps
Start with an agent that is only allowed to look. Weekly: which products have no description, which pages lost traffic, which searches on your own site returned zero results. A message in Slack or email, three to five sentences.
This sounds small. It is the step where you learn whether the agent understands your store, without anything being able to go wrong. At one client, the zero-results log surfaced a whole gap in the range that nobody on the floor had seen.
Step 2: content with business context
Now the agent may write, but only from context you have set down: what you sell, to whom, how you sound, which claims may never be made, and the product data as the only source of facts.
- Creating products. Supplier data in, complete product page out: title, description in your tone, specifications, internal links to the right category. The agent does not invent properties; what is not in the data does not go in the copy.
- Category copy. Written from the range that is actually in it, with the search intent that comes from Search Console.
- Improving existing pages. The agent analyses a page, proposes a better version, and you approve.
Generic AI copy is not a model problem. The system did not know your store. This is the step where that changes.
Step 3: building and designing pages
This is where it gets interesting for anyone who wants more than copy. With a design system (colours, typography, components, set down as instructions) the agent can deliver complete pages that look like the rest of your site: a landing page for a promotion, a guide page for a product group, a comparison page.
How to brief AI for pages that convert is a discipline of its own; the core is that the design system sets the limits, so that "make a page about X" always lands within your house style. Our own sites run this way: one design language, hundreds of generated pages, no loose ends.
Step 4: on a schedule, not on attention
Everything so far you can start by hand. The gain comes when it runs on a schedule: the content check on Monday, new products processed on Wednesday, improvement proposals on Friday. Nothing depends on whether someone remembers.
This is also the step where guardrails become mandatory. What is never allowed without a human? For us: publishing without a feedback round, claims outside the product data, changes to prices. It lives in one file that every agent reads.
Step 5: everything from one workspace
The last step is not a new task but coherence. SEO, content, advertising and reporting share the same business context and the same limits. Decide that a product group gets priority, and every agent works with that from that moment on. That is what we call a command center, and it is the point where "using AI" turns into "running your store with AI".
Our travel site Guide2Japan works this way: agents research, write, edit, translate and publish; there are exactly two approval moments per article. The story is in the article about that site.
What the human keeps doing
| Step | The agent | You |
|---|---|---|
| 1 Read | Flags | Judge whether it is right |
| 2 Content | Writes from context | Set the context, approve |
| 3 Pages | Builds within the design system | Define the system, approve |
| 4 Schedule | Runs | Guard the limits, read the reports |
| 5 One workspace | Collaborates | Make the decisions that steer everything |
The expert stays. The click work moves into the system. That is the whole thesis.
What it realistically delivers
- Time. Creating product pages goes from an afternoon per batch to fifteen minutes of approving.
- Volume. Work that used to require an extra hire becomes feasible without extra people, provided the context is right.
- Consistency. Every page in the same tone, from the same source of facts.
- Overview. Every week you know what was done and what was proposed.
What it does not deliver: strategy. Which product groups matter, where the margin is, where the store is heading. That remains the source the system draws from.
Where to start tomorrow
- Step 1, one task. A weekly notification of products without a description is set up in an afternoon.
- Write your business context. One page. This is the work everyone skips.
- Choose your two approval moments before you let anything be written.
- Four weeks per step. Faster is possible, but then you miss the lesson of the previous step.
Frequently asked questions
Does this work with my store platform? Shopify and WooCommerce have ready-made connectors; Lightspeed, Magento and custom builds through their API. Anything with an API or export is reachable.
Do I need a developer on staff? No. Setting up connectors and limits is specialist work (we do that); day-to-day management is giving instructions and approving.
What if the agent publishes something wrong? That is why it does not publish without a feedback round, and only works from product data. Errors stay in the draft and are logged.
How does this compare to an agency? The agency sets it up, guards the limits and reads along. The agent does the work that used to be billed by the hour.
Want to know which step your store is on and where to start? Request a free audit.