A product description is, in most SME shops, the last item on the "we'll get to it someday" list. The photo is there, the price is there, the category is there. The description is a sentence and a half copied from the supplier's invoice — or an empty frame. And it's been like that for five years. AI sounds like the answer, and sometimes genuinely is. You just need to know the point where it stops helping and starts quietly cluttering the catalog.
At one of our clients — a 15-person wholesaler of fasteners near Warsaw — the B2B portal held 3,800 items. About 60% had no description at all, the other 40% had something like "M8x30 bolt DIN 933 galvanized". Zero chance for a customer to find the product by its application ("steel bolt, galvanized, for outdoor construction"). An AI pipeline (an LLM API plus a simple script) generated descriptions for 2,400 items in two weeks. After technical review, 1,800 went live unchanged, 420 needed minor edits, and 180 were rejected — and those 180 are the most important part of this story.
Why descriptions are always last on the list
Writing a decent product description takes 15–30 minutes. Multiplied by 3,800 items that's a thousand hours — half a full-time job for a year. Nobody at a 15-person company has that. Hiring a freelance copywriter means 30–50 PLN per description, which is 150,000 PLN total. The owner looks at that number and closes the spreadsheet.
The result: 90% of SME catalogs in Poland look like that one near Warsaw — raw supplier data, no context, no keywords, nothing that would help a customer choose between two similar products. The competition on big marketplaces wins not because they have better products, but because their products can actually be found.
This is where AI actually changes something. Not because it writes "better" than a human — it doesn't. Because the cost of generating 2,000 descriptions drops from 60,000 PLN to 150 PLN (token cost) plus 40 hours of human review.
What AI really can do — and what it won't do
A language model will cleanly turn dry technical data into readable copy in natural Polish or English. Feed it "M8x30 bolt DIN 933 galvanized A2L steel 8.8" and it spits out three sentences of description plus five SEO sub-keywords. Give it three examples of "this is how we write descriptions in our company" and it adapts to the style.
What it won't do:
- Verify that your input data makes sense. If the file says "M8x30" and reality is M10x30, the model writes a beautiful description of a product that doesn't exist.
- Figure out how customers actually use the product. If you don't tell it the bolt goes into outdoor balcony frames, the model will invent an application — sometimes right, sometimes a lie.
- Spot that two items in your catalog are practically identical and only one should stay. AI generates descriptions, it doesn't clean up catalogs.
- Replace technical review. More on this in a moment.
That isn't a limitation of a specific model. It's the limit of what you can do at all without a human in the loop.
How to build a simple pipeline
You don't need a 20,000-euro-a-year platform. You need three things: a good input, a good prompt, and a good output control.
Input. Export from your ERP or supplier file to CSV, with columns: name, code, technical parameters, category, price group. The more structured data the better. If a product group has 10 typical applications — attach them as shared context. 20 minutes of work for someone who knows the products.
Prompt. One, polished, with three examples of an "ideal description" from your industry. Don't write thirty versions of the prompt — one per product category (bolts, nuts, washers etc.) and that's it. Every prompt change means re-reviewing everything that went through it.
Output. Nothing goes live until a human has read it. The next section covers what the reviewer does. But the ironclad rule: zero direct publication from AI into the production system. Always an intermediate layer.
Token cost for 2,000 descriptions on a GPT-class model is tens to a few hundred PLN. The biggest cost is the reviewer's human time. Realistically: 2–3 minutes per description = 70–100 hours of technical work. One person for a week, not a year.
Quality control — where AI gets it wrong
Those 180 rejected descriptions at the client near Warsaw had three typical error patterns. Worth knowing before you start the pipeline.
First — invented parameters. The model gets "M8x30 bolt" and sometimes adds "with cylindrical hex socket head" — because most M8x30s in its training set had one. Except DIN 933 has a hex head. The description reads well, but misleads the customer. A reviewer who knows the products catches it in 20 seconds.
Second — invented applications. "Ideal for food industry applications" on a bolt without a food-contact certificate. Or "for marine construction" on standard galvanized steel (which corrodes in seawater in three months). These are errors that generate real complaints.
Third — overly confident tone on generic products. "Our premium M8x30 bolt stands out for its exceptional quality" — on a 15-grosz-per-piece commodity that three competitors within 50 km stock in exactly the same spec. Technical description yes, marketing laurel no.
The reviewer doesn't have to be a copywriter. They have to know the products. Best done by someone from the technical department or a senior sales rep. 15 hours a week for a month and the catalog is sorted.
Where to start in your shop
Not with the pipeline. With a list of 20 items that most often land in the cart — or, conversely, 20 items that have traffic but a desperately low conversion rate. Generate descriptions for those 20 by hand, through a chat interface (not an API). Review them. See how many are "ready to publish", how many "need edits", how many "throw out".
If the ratio is around 70/20/10 — building a full pipeline makes sense and pays back in the first month. If it's 30/30/40 — you have a problem with input data, not with the model. Clean up the product master first.
If you want to talk through whether this makes economic sense in your industry and how to organise the review — the first call is free and takes half an hour. We'll show you how to work out the cost per description for your business, and the pitfalls specific to your category.
AI doesn't write descriptions better than a human. It writes them a hundred times cheaper — and that changes everything, provided someone actually reads them afterwards.