AI for ecommerce: start with the catalog, not checkout

Summary: AI tools are moving into everyday ecommerce work: catalogs, images, campaign drafts, inventory text, and basic analysis. Start there. Do not let AI near checkout, payments, tax, or fulfillment until the workflow is tested and someone owns the boundary.

The big demo is tempting: describe a store, get a finished shop, hit publish.

But online stores usually break in boring ways. The wrong price. The wrong image on the right SKU. A sale that promises too much. Product copy that sounds nice but misses measurements, material, or allergy information. A campaign that sells out something the stockroom does not have.

So the useful starting point is not "AI should run the whole store." It is simpler: let AI clean, suggest, and prepare the catalog while checkout stays locked.

What actually changed

Manus has launched Shopify on Manus. According to Manus, users can create a Shopify development store from a prompt, draft branding, products, pages, and a working cart, or connect an existing Shopify store for catalog work, sales analysis, and campaign drafts. Manus also says Shopify still handles checkout, payments, taxes, shipping, orders, and fulfillment.

Source: Introducing Shopify on Manus: Build, Run, and Grow Your Store From One Chat

Perplexity also launched its Main Street AI Accelerator with a clear small-business angle: $25 million in Perplexity Computer credits, $250 per company for up to 100,000 eligible companies, according to Perplexity's own program page. Applications were not open when checked, and the SBA relationship should be described as Perplexity-stated until a matching SBA page exists.

Source: Main Street AI Accelerator

On the infrastructure side, Cloudflare is making it easier to route Grok/xAI through AI Gateway and consolidate AI costs through Cloudflare billing. That is not an ecommerce feature by itself. But it shows where the market is going: AI is getting easier to enable, easier to buy, and easier to spread into more workflows than anyone first planned.

Source: Cloudflare AI Gateway: xAI / Grok provider

Source: Cloudflare AI Gateway: Unified Billing

Start with the catalog, because the mistakes are visible

The catalog is a good first surface for AI because much of it is already half-structured. There are SKUs, images, supplier sheets, old product descriptions, collections, sizes, inventory status, and customer questions. AI can help find gaps and write the first draft.

That does not mean AI gets to decide what is true. It can suggest that a jacket lacks measurements, that an image seems attached to the wrong product, or that a category should be split. A person still needs to know what is actually on the shelf.

A good first AI workflow for an online shop could look like this:

• Upload a product sheet and images.
• Ask AI to match image, SKU, title, and description.
• Ask AI to flag uncertainty instead of guessing.
• Review anything that changes price, stock, legal information, or a customer promise.
• Publish only after a named person has approved the batch.

It is less spectacular than an autonomous store. That is also why it works.

Let AI draft campaigns, not promise delivery

Campaigns are the next reasonable step. Manus describes using Shopify data to create campaigns around slow movers, repeat buyers, hero collections, landing pages, Instagram posts, email, and ads.

Source: Introducing Shopify on Manus: Build, Run, and Grow Your Store From One Chat

The practical setup should be narrower than the demo. Let AI suggest:

• Which products need better descriptions.
• Which older products could become a campaign.
• Three newsletter subject lines.
• A landing page structure.
• Questions customers are likely to ask.

Do not let it decide discount levels, inventory truth, delivery promises, return exceptions, or which customers should receive a specific offer. A bad guess costs more there.

Checkout is not the first test surface

Checkout carries a different kind of risk than product copy. When something goes wrong in the catalog, you often get a correction path: change the text, swap the image, contact the customer. When something goes wrong in checkout, you may have the wrong payment, the wrong tax, the wrong shipping promise, or an order that needs cleanup afterward.

So checkout should have a plain rule: AI may prepare the material, but it does not own the transaction.

For a smaller online shop, the boundary can look like this:

• AI may suggest product copy, collections, and campaign ideas.
• Inventory or price mismatches should go to review.
• Customer replies about delivery and returns should come from approved answers.
• Payment settings stay out of scope.
• Tax, shipping zones, and discount rules need manual approval.
• Any campaign that affects price or delivery needs an accountable person before publication.

This is not process for process's sake. It is the difference between useful AI support and a shop that has to apologize later.

A simple starting plan for ecommerce AI

If you want to test AI in an online store, start with a workflow small enough to check.

1. Pick one catalog surface. Take one collection, not the whole store. For example: "all lamps without alt text" or "products from supplier X".

2. Decide what AI may change. In week one, the answer may be: nothing directly. AI only creates suggestions and an error report.

3. Write the review rule. Who approves product copy? Which role may change price? Where does an inventory mismatch go? If nobody owns the point, AI should not touch it.

4. Run on sample data. Use a development store, export, or copy before connecting live flows.

5. Measure something boring. Time saved on product copy. Missing images found. Fewer manual campaign drafts. If the only metric is "AI feels smooth," you do not know whether the workflow is worth keeping.

Where Hammer fits

This is a typical Tool Forge job: not picking the most impressive AI demo, but drawing a workflow people can trust.

For an online shop, it usually starts with three questions:

• Which catalog areas are safe enough for AI drafts?
• Which decisions require human approval every time?
• Where is the stop button if the tool, credits, or provider behaves differently tomorrow?

If your shop already has products, images, campaigns, and order flows, AI is not a magic new layer above all of it. It is a fast intern. Give it the catalog first. Keep checkout behind the glass until the routine holds.