Six tasks that really take hours in a shop
This article is part of the series "AI for business by industry" and answers the same five questions: what, with what, how much, what not to do and where to start. It covers only tasks that make sense in an online shop and not in a café or a clinic.
- Product titles and descriptions from specifications. The supplier sent a table with SKU, size and composition; the shop needs text for the product page and for the product feed. Titles too, which at the supplier look like "T-sh. men wht XL art. 4471".
- Merging supplier price lists. Three suppliers name the same attribute differently: "колір", "цвет", "Color"; sizes are in centimetres in one list and S, M, L in another. Before the goods go on the website, someone reconciles this by hand.
- Replies to the same questions in messengers. "Is it in stock?", "Which size for 176 cm?", "When will you ship?" Most answers are already in the product description or the delivery terms.
- Reasons for refusals and returns. A manager writes "wrong size", "colour not as in the photo", "changed their mind" in the comment field. These notes pile up in the hundreds, and nobody knows which reason is the main one.
- Customer reviews per product. What exactly people praise and complain about in a given product: "runs small", "seam splits after washing". Material for fixing the description and for a conversation with the supplier.
- Product photos. Removing the background, evening out the light, staging a product in an interior. Google has firm rules here; they get their own section below.
What to use
| Task | Tool | How to check the result |
|---|---|---|
| Descriptions and titles | An assistant for dozens of products; a model API or a built-in platform feature for hundreds | Every fact checked against the specifications; in the feed, a separate field marked as AI |
| Supplier price lists | An assistant with the table: "map the columns to this list of attributes" | 20 rows checked by hand; product counts must match |
| Messenger questions | Reply templates from an assistant, or a bot with a knowledge base | 50 real questions before launch; hand over to a person for anything about money |
| Refusal and return reasons | An assistant with exported comments, without customer data | Classify 20 random records by hand and compare |
| Reviews | An assistant: "group the reviews of this product by topic with counts" | Back each topic with two or three quotes from the reviews |
| Photos | Built-in editor tools and image generators | Colour, pattern and material match the product; AI metadata not stripped |
How to choose between a button bot and an AI assistant for messengers is in the article on chatbots for business. How to put AI descriptions into the feed so that Google does not ignore them is in the article on AI product descriptions. We will not repeat either here.
What it costs
Two lines of spending: an assistant subscription for people, and paying for a model through the API if descriptions are generated in bulk.
Subscription. According to ChatGPT's pricing page on 11 October 2026, Plus costs $20 a month, and Business $25 per user billed monthly or $20 billed annually, from two users. At the National Bank of Ukraine rate for 11 October 2026 (UAH 44.8526 per dollar), Plus is about UAH 897, and Business for two people billed monthly about UAH 2,243. The difference is not only price: on Business, content is not used to train models by default; on Plus you have to switch that off yourself. Other assistants' prices are in the series hub.
Descriptions through the API. An illustrative calculation for an 800-product catalogue: one description takes about 700 input tokens (specifications and instructions) and 450 output tokens (the description). Prices from OpenAI's API pricing page on 11 October 2026, standard mode.
Prices per 1M tokens, standard mode, short context: Luna $0.10 input / $0.50 output, 6.1 Sol $2 / $10, Astra $10 / $50. Batch mode halves the prices. NBU rate on 11.10.2026: UAH 44.8526 per dollar. One description: 700 input + 450 output tokens.
Even the most expensive model costs about a thousand hryvnias for the whole catalogue, and 30 new products a month on the mid-range model cost under UAH 10. These sums are not worth economising on.
Another number is. If a person checks each description for three minutes, comparing sizes, composition and contents with the specifications, 800 descriptions take 40 hours. That is the real price of AI descriptions. It pays off only if descriptions used to be written from scratch or did not exist at all. If you already have decent descriptions from the supplier, rewriting them with AI for the sake of "uniqueness" is 40 hours with no obvious result.
Product photos: what Merchant Center requires
For products a shop sends to Google Merchant Center for ads and free listings in Shopping, firm rules apply to AI photos. The help page on AI-generated content says every image created with generative AI must carry metadata marking it as such under the IPTC standard, in particular DigitalSourceType with the value TrainedAlgorithmicMedia. Embedded tags added by AI tools must not be removed. AI images are allowed as the main, additional and lifestyle images.
The second set of rules is on the image_link attribute page. It applies to any photo, but AI editing is what most often breaks it.
| What you do to the photo | What the rules say | Conclusion |
|---|---|---|
| Generate a scene or background around the real product | AI images are allowed with IPTC metadata; the main image must clearly show the product itself | Yes for lifestyle images; use a product shot as the main one |
| Generate a product photo from a description, without a shoot | The image must accurately match the product variant: colour, pattern, finish, material | Risky: an image generated from a description does not guarantee that |
| Recolour one photo for other colour variants | Each variant needs its own image accurately showing the selected colour | Only if the colour in the photo matches the real product |
| Add a "Sale" label, the shop logo, a frame | Promotional text, overlays, added logos and borders are grounds for image disapproval | No, with or without AI |
The practical consequence: if a photo passes through several programs (a generator, an editor, compression for the website), check after the last one that the IPTC metadata is still there. A separate article will cover AI product photos and when generation is cheaper than a shoot.
What not to do
- Do not paste customer data into an assistant. Analysing refusals and returns needs the product, the reason and the comment, not the surname, phone number and branch address. Name, phone and address are personal data under the Law of Ukraine "On Personal Data Protection". On personal assistant plans, content may be used to train models by default. Where to switch that off is in the article on data for an AI assistant.
- Do not let a bot make promises. A discount, a delivery date, agreement to a return: the bot takes these only from the shop's rules; a disputed return goes straight to a person.
- Do not ask AI to "issue a receipt". A fiscal receipt comes only from a registered software or hardware cash register; when an online shop needs one is in the article on online acquiring.
- Do not publish descriptions without checking. An invented property in a description is a promise to the customer and a future return.
First step for one week: refusal and return reasons
Of the six tasks, this one gives a result in hryvnias, not impressions. A refused parcel costs the shop at least the delivery; a return costs delivery both ways. How much these losses take from every order is in the article on online shop unit economics.
- Export three months of refusals and returns: product, variant, the manager's or customer's comment. No names, phone numbers or addresses.
- Ask the assistant to sort the records into 6–8 reasons with a count for each, overall and per product. Ask for three examples per reason.
- Check 20 random records by hand. If the assistant is wrong in more than one case in five, refine the list of reasons and run it again.
- Take the main reason that can be fixed on the product page: size chart, colour in the photo, what is in the box. Fix the 10–20 products where it appears most often.
- A month later, compare the share of refusals and returns for these products with the previous three months.
The work takes a day; the subscription is either already paid or costs UAH 897. If the refusal share for the fixed products has not changed, the cause is not in the description but in the product, or in advertising that brings the wrong buyers. That is a result too.
Our position
Do not start by generating descriptions for the whole catalogue. The model costs next to nothing, but checking takes dozens of hours, and the gain from new descriptions is hard to measure. Start with the refusal and return analysis: it is cheap, it shows what exactly to fix, and the result shows up in figures a month later.
The position does not hold for a shop that is just launching or adding a large new range with no descriptions. There descriptions are not an optimisation but a necessity, and AI drafts with a human check are the fastest way to get them.
ChatGPT prices were checked on 11 October 2026 on ChatGPT's pricing page, API prices on OpenAI's API pricing page, the rules for AI images and product photos in Google Merchant Center Help, and the exchange rate at the National Bank of Ukraine for 11 October 2026. The description cost calculation is illustrative: the number of tokens depends on the length of the specifications and the text.