AI for an online store, and the eleven jobs behind it

What an AI team for an online store actually is

An AI team for an online store is several AI workers, each holding one job in the shop, working from the same catalogue and the same order records. One keeps product listings correct. One writes the descriptions. One follows up the basket somebody filled and abandoned. One answers where is my order, at midnight, in whatever language the customer used.

A shop is a strange business to run because the work is both constant and invisible. Nobody sees the hours that go into checking a variant is in stock, fixing a product feed that stopped, replying to the same delivery question forty times, or noticing that a good customer has not ordered since March. That work does not stop when you are busy. It just does not get done.

So the roster is arranged around the moments where money is lost. The listing that was wrong, the basket that was left, the question that went unanswered until the customer bought elsewhere, and the buyer who was never asked to come back.

The team, and what each one does

Eleven jobs. Most stores start with the order desk and the cart chaser.

• The catalogue keeper. Adds products, keeps variants, images and tags consistent, and flags the listing that says in stock when the stock record says otherwise.
• The page writer. Produces titles, descriptions and the details that decide a purchase: materials, sizing, what is in the box, what it does not do.
• The feed manager. Keeps the product feeds that shopping adverts run on clean, because a feed with errors quietly stops showing your best sellers and nobody notices for a fortnight.
• The basket chaser. Follows up an abandoned basket on whichever channel that customer actually uses, once, usefully, rather than five times until they block you.
• The browse follower. Handles people who looked repeatedly at one product and never added it, which is a different customer with a different reason.
• The shop assistant. Suggests the right product when somebody describes what they need rather than knowing what it is called, and says when you do not sell it.
• The messaging assistant. Answers on WhatsApp, sends the product link, confirms availability and takes the conversation to checkout, which in this market is where a large share of orders actually begin.
• The order desk. Answers where is my order, when will it arrive, can I change the address, has my payment gone through, by looking it up rather than guessing.
• The returns guide. Walks a customer through a return or exchange under your written policy and escalates the ones that are not routine.
• The repeat nudger. Works out roughly when a customer is due to reorder and gets in touch then, which is the cheapest revenue in any shop.
• The store reporter. Reports what sold, what stalled, which products get returned most and where baskets are being abandoned.

A Tuesday inside an AI store team

12:20am. A customer messages asking whether a jacket comes in a larger size. The messaging assistant checks the catalogue, answers, sends the link and confirms delivery to their city. That order is placed before anyone wakes up.

7:30am. The catalogue keeper finds two products marked available that the stock record says are gone. It does not change the shop by itself, because unpublishing a product is a decision with money attached. It flags both.

9:00am. A customer sends a voice note in two languages with a screenshot of a bank transfer. The order desk reads both, matches the amount to an unpaid order and marks it for confirmation. This is a normal message here, not an unusual one.

11:00am. The basket chaser works through last night. Nine abandoned baskets, each contacted once on the channel that customer used, with the item named rather than a generic reminder that they left something behind.

1:40pm. The feed manager reports that eleven products dropped out of the shopping feed overnight because a required field went missing during an import. Left alone, that would have been a fortnight of adverts not running.

3:00pm. The returns guide handles four returns under your written policy and escalates a fifth, where the customer says the item arrived damaged. Damage means a photograph, a decision and probably an apology, and all three belong to a person.

5:30pm. The repeat nudger contacts customers whose usual reorder time has passed. Not a discount, just a reminder. The discount is a decision you make, not one it makes.

6:00pm. The store reporter posts four lines. Orders, revenue, baskets abandoned, and the one product whose return rate has climbed enough to be worth a look.

What stays human in a shop

Money going back stays human. Refunds, goodwill, replacing an item at your cost. These are relationship decisions and they should be made by somebody who can weigh what this customer is worth against what the item cost.

Price and discount stay human. A worker offering a discount to close a sale will close many sales and quietly train your customers to wait for one. Where discounting is allowed at all, it runs to rules a person wrote.

Promises about stock and delivery stay human when they are not certain. A system reads what your records say. If your records are optimistic, it will repeat that optimism to a customer, and the customer will hold you to it.

Damage, loss and anything that went badly stay human. Somebody who received a broken item wants to hear from a person. The fastest possible automated apology is still an automated apology.

How the shop connects to other departments

A store is where the departments overlap most, because the same product data feeds the adverts, the pages, the answers and the reports. When each of those is maintained separately, they drift, and the customer is the one who finds the difference.

The link worth building first is between the order desk and everything else. What customers ask about most is the clearest instruction any other department will get all year.

• The marketing team runs shopping and retargeting campaigns off the same product data the catalogue keeper maintains.
• The content team produces the product images and descriptions, which is what the page writer works from rather than inventing.
• The support team handles anything past a routine order question, with the whole conversation handed over rather than restarted.

What an ecommerce system costs and what changes the number

An audit, then a build, then a monthly fee while it runs. The catalogue is what moves the figure most. Forty products with two variants each is a short build. Four thousand listings with inconsistent naming, missing attributes and images of different quality is a data project first and an AI project second.

The other variable is how you take money. Cash on delivery changes the shape of everything: the confirmation step matters more, refusals cost real money, and the messaging around delivery becomes part of the sale rather than an afterthought.

• How many products and variants, and how consistent the existing data is.
• Which platform the shop runs on and what it will let a system read and write.
• Whether cash on delivery is used, and how confirmation is handled today.
• How many channels customers order and ask through.
• Whether product photography and descriptions are included or already exist.
• Whether Wobble operates it monthly or hands it to your own team once the handover is done.

Stores too small for this to be worth it

If you take a few orders a week, a person handles all of it in less time than it takes to explain the system. Everything on this roster earns its place through repetition. Without volume there is only setup, and setup is the expensive part.

If your stock numbers are wrong, automation sells things you do not have, faster and more confidently than a person would. Fixing stock accuracy is dull, internal and not something a supplier can do for you, and it comes first.

If your real problem is delivery, this fixes nothing. A shop with a courier that loses parcels does not have a support problem, it has a supplier problem, and better messaging around a late parcel is a more articulate version of the same disappointment.

If your catalogue is small and never changes, several of these workers have nothing to do. A single product business is usually better served by the messaging assistant and the repeat nudger alone, and can safely ignore the rest.

Where the shop team stops, how fast it starts, and what stays yours

Three things this will not do. It cannot fix a product nobody wants, and a store team applied to a bad catalogue produces faster confirmation of the same result. It will not set a price, because pricing is a margin decision with somebody's name on it. And it does not touch a refund or a cancellation alone, since a wrongly cancelled order costs you the sale and the customer at once.

Anything ambiguous goes the same way. A payment screenshot that does not match the invoice, an address failing validation twice, a complaint mentioning a previous complaint. The system attaches what it has and hands it to a human rather than trying a third phrasing.

The first week is the catalogue audit, which is where the real problems usually turn out to be. One worker runs inside a fortnight, normally order confirmation, and the rest follow across the second month. Ibrahim owns build and workflows and takes this kind of engagement. The storefront work on the work page shows the finished end: a Shopify build for Moni Honolulu and a theme migration for Oahu Golf Apparel.

The store, the payment provider, the messaging numbers and the data sit in your own accounts throughout. If you leave, the flows carry on and the product data is yours, which matters because the product data took the longest. Wobble stays accountable for the workers it installed.

Common questions

What does AI do for an online store?

It covers the repeated jobs: keeping the catalogue correct, writing product pages, maintaining shopping feeds, chasing abandoned baskets, answering order and delivery questions, guiding returns and prompting repeat purchases. Each is a separate worker rather than one assistant told to help with the shop.

Can AI answer order questions on WhatsApp?

Yes, and it should look the order up rather than guess. Under the business platform's rules customers opt in, messages you start use templates approved in advance, and free-form replies are limited to a window after the customer's last message. Those rules shape what is possible.

Does it work with cash on delivery?

It has to in this market. The useful part is the confirmation conversation before dispatch, where an unclear or unanswered order becomes a refused parcel and a real loss. That step is worth more than most of the rest of the roster combined.

Will AI written product descriptions hurt search visibility?

Thin descriptions do badly whether a person or a machine wrote them. What ranks and what sells is the same thing: real detail. Materials, measurements, what is included, what it does not do. If those facts exist, the writing tool is not the deciding factor.

How many orders make this worthwhile?

There is no universal number, but the test is repetition rather than revenue. If the same question is answered more than a few times a day, or if baskets are abandoned faster than anyone can follow them up, the volume is there. If not, a person is still cheaper.

Which store worker should be installed first?

The order desk, because where is my order is the largest single category of messages in almost every shop and the easiest to answer correctly from real data. The basket chaser is usually second, since it produces visible revenue within weeks.