Product data, product descriptions, customer questions and stock – with the facts from you and the responsibility staying with the shop
I help online shops use AI where the time disappears: product data that needs completing, product descriptions that need writing, customer questions about orders and returns, and material for purchasing and stock – connected to the e-commerce platform you already have.
The information about the products comes from you and your suppliers. AI phrases, sorts and calculates suggestions. A person decides what is published and what is ordered.
An online shop is a catalogue that is never finished. New items arrive with the supplier's description in English and half the fields empty, customers write in to ask whether a part fits their model, and on top of that someone has to decide how many of each to buy in for next season. All of that is work with text and figures that already exist somewhere – and that is where AI can take some of the load.
But in e-commerce the text is part of the purchase. What it says on the product page is a promise to the customer, and a language model that writes “waterproof” about a jacket that is only water-repellent has created a problem, not saved time. So the work is as much about where the information comes from as about what the tool should do.
Supplier files, PDF data sheets and price lists in different formats can be read by a language model that picks out the attributes – dimensions, weight, material, colour, compatibility – and puts them in the right fields. The same size written four different ways becomes one. Items where a required field is missing can be put on a list instead of being published with a guess. A value that is not in the source material should not be in the shop.
Product descriptions for the whole range
Once the attributes are there, the texts can be built on them: a description for each item, texts for category pages, meta descriptions for search engines and translations into more languages. The shop's tone is set once in an instruction, and every text is checked against the product data before it goes out. The more similar items, the bigger the time saving.
Customer questions before the purchase
“Does this fit my model?”, “How big is it in real life?”, “Can it go in the washing machine?” The answers are often already in the product data, but the customer cannot find them. A chat or a draft reply that only draws on the shop's own information, and hands over when the answer is not there, is set up to reduce both the emails and the wrong purchases.
Order and returns questions
Questions about delivery status have an answer in the order and in the carrier's tracking. Returns and complaints can be sorted, summarised and given a draft reply with the order history attached. The decision on a complaint is made by the shop – that is where the consumer's rights are settled, and where a mistake costs the most.
Forecasts for purchasing and stock
Sales history, season and planned campaigns can become a suggestion for how much to order, and a list of items that are about to run out or are sitting still on the shelf. The calculation itself is done with ordinary statistics. The AI describes the deviations in plain language. A forecast is an estimate, and for new items with no history it is uncertain – that should be visible in the material, not hidden.
The platform
Connected to the shop, not next to it
An AI tool that requires someone to copy products into another window and paste the texts back quickly becomes a tool nobody uses. So the work goes through the e-commerce platform's API: products, orders, customers and stock levels are read from there, and whatever needs to go back – a new text, a completed attribute – is written there after approval.
Many platforms now have their own AI features for product descriptions – check what yours has before building anything of your own. If they are enough for your range, you should use them, and I will say so in the first conversation. A flow of your own pays off when the material comes from many suppliers in different formats, when the texts need to be based on your own product data rather than a general template, or when the same information also needs to go on to the accounting system, price comparison sites and marketplaces.
I build it so that every change written to the shop is logged, so you can see what was changed, when and on the basis of what material. The starting point is that prices and stock levels are not changed by a language model.
Honestly
What it says in the shop is the shop's responsibility
Information that becomes part of the purchase. Under the Swedish Consumer Sales Act (konsumentköplagen, 2022:260), goods must match the information about their characteristics that the trader gave in the marketing or otherwise before the purchase (chapter 4, section 2). If they do not, the goods are defective. A feature that a language model added on its own is therefore not a turn of phrase – it can become a complaint.
Misleading marketing. The Swedish Marketing Practices Act (marknadsföringslagen, 2008:486) prohibits incorrect claims and other representations that are misleading. That applies whoever wrote the text. Claims about the environment, health and safety are the hardest to stand behind if they do not come from material you can show yourselves.
Product images. An AI-generated lifestyle image can be fine for the mood. An image meant to show what the product looks like should be a photo of the product.
Customer data. Orders contain names, addresses and purchase history. Whatever is sent to an AI service needs a data processing agreement and clear information on where the data is processed. How to arrange that is on the page about AI and GDPR, and what applies to a chat that talks to customers is covered under labelling on the page about AI for marketing.
This describes the situation in November 2026. I am an AI consultant, not a lawyer, and in borderline cases you should ask one.
How it works
From a first meeting to a flow in operation
A first meeting. You tell me where the time goes – new items, customer emails, returns or purchasing – and which platform the shop runs on. The meeting costs nothing.
A look at the material. An export of the product data and a sample of customer questions say more than a description. They show which fields are missing and which questions can actually be answered from what is there. After that you receive a fixed quote.
One flow at a time, tested on part of the range. The first flow is tested on one category, and the result is compared with what you would have written or replied yourselves. You see where it goes wrong before it touches the whole shop.
Approval and handover. You decide who approves what, and receive documentation, instructions and access in your own name. If staff are to work with the tools themselves, there is AI training for businesses.
Honestly
When an online shop should hold back
When the range is small. Fifty items that change once a year are better written by you. The time saving is in the volume.
When the material is missing. If the supplier has not stated the material, no model can know it. Then the first job is getting the information in, not phrasing it.
When the platform's own feature is enough. A build of your own next to something you already pay for is then not always worth the money.
When the replies require judgement. An unhappy customer needs a person who can make decisions, not a faster standard reply.
Background
Who builds it
My name is Mikael Glännström and I run the one-person business Glänne & Söner in Lindesberg. I build AI solutions and connections between systems, and I would rather say in the first meeting that something will not pay off than after a build.
The work can be done remotely, so I take on assignments from online shops all over Sweden. All the AI services are gathered on the page about AI consulting, and if AI is to go into more parts of the business than the shop, it starts with an AI strategy.
Can AI write product descriptions for our online shop?
Yes, if it gets the facts from you. A language model writing from a product's own attributes – dimensions, material, compatibility, care instructions – can produce a draft for each item in a few seconds, in the shop's tone of voice.
What it must not do is fill the gaps. The set-up is therefore that a text is based only on fields that exist in the product data, that items with missing information are set aside instead of guessed, and that someone reads the texts before they are published.
Who is responsible if an AI-written product description is wrong?
The shop. For the consumer it makes no difference who or what wrote the text. Under the Swedish Consumer Sales Act, goods must match the information given about them in the marketing, and the Swedish Marketing Practices Act prohibits misleading claims.
A feature that a language model has invented can therefore become both a defect in the goods and misleading marketing. This describes the situation in November 2026 and is not legal advice.
Can an AI chat answer questions about orders and returns?
Some of them. Questions like “where is my parcel?” have an answer in the order status, and a chat can look it up if it is connected to the platform and the carrier.
Complaints and disputes are something else. There the shop decides, not a model, and the chat is set up to hand over to a person with the case summarised.
Does it work with the platform we already have?
It is the platform's API that determines what is possible. Where product data, orders and stock levels can be read and written through an API, the AI steps can sit between the platform and the systems you already use. What your platform and your plan allow is something I check before the quote.
How much does it cost?
It depends on how large the range is, how orderly the product data is and which systems need to be connected. You receive a fixed quote after a first meeting, free of charge.
Where does the time go in your online shop?
Tell me which platform you use and what takes the most time – new items, customer emails, returns or purchasing. After a first meeting, free of charge, you will know what can be taken off your hands, and receive a fixed quote.
