When people talk about building an online store, they talk about design, conversions and speed. What actually eats the most time and nerves is rarely mentioned: getting the products into the store.

Especially when the products don't come from one place and there are thousands of them. Luckily, AI is a wonderful helper here.

The main problem: every supplier's files are different

The client had several suppliers whose products had to run into a single store. The problem wasn't the number of products, but that every supplier's files were put together in a completely different way. It's not that the products couldn't be added to the store, but rather that the quality with which they sat there wasn't good and didn't create a consistent impression.

The differences were on every level:

  • Structure. Columns, structure and logic varied from file to file.
  • Language. Texts were in different languages and of different quality, some machine-translated, some original.
  • Attributes. Some files had product attributes, some didn't. And when they did, each expressed them in its own way.
  • Format. The same thing (for example colour, size or material) was written differently in different files.

Done by hand, this would mean reviewing, translating, rewriting and reformatting every single product. For over 20,000 products that's several months of work.

What we built

We built an AI workflow that takes a supplier's source file as it is and brings it into a single format for the store. The workflow does several things:

  • Unifies the structure. All products end up in the same format, regardless of the shape they came in.
  • Translates and writes. Texts are brought into one language, with consistent quality and in the brand's tone of voice.
  • Extracts and normalises attributes. Product attributes are pulled out, brought into a single logic and converted into the right format for the store.

Result

  • Over 20,000 products unified, translated and ready for the store.
  • The client likely saved several months of manual work alongside their other core activities.
  • The AI cost stayed under 200 euros in tokens.

Where AI works well and where it doesn't

Honestly, this workflow (or any AI workflow) isn't completely foolproof or a fully automatic press of a button.

AI is good where the work is high-volume, repetitive and pattern-based, meaning unifying structure, translating and extracting attributes. Exactly the part that's tedious and error-prone for a person.

Where a human is still needed:

  • Setting the logic. What goes where, which attribute matters, from what data and how the data can be extracted, in what order to run the tests, how to keep token costs under control, the system architecture from files to store, and awareness of the real technical limitations. All of this needs conscious, well-thought-out setup by a developer. AI can follow it, but even at this stage it needs guidance, a technical overview and control to quickly catch possible errors and quality issues.
  • Control. Spot-checking and catching edge cases. The messier the source file, the more control is needed.
  • Decisions that aren't pattern-based. Questionable or contradictory data needs a human eye.

The value of the workflow isn't that it replaces the human. The value is that the human deals with decisions and control instead of typing 20,000 rows by hand.

When is it worth considering AI for cleaning up a product feed?

This isn't a solution for every store. But if any of these describe your situation, it's worth a thought:

  • Products come from several sources and in different formats. Among other things, this means you can't build a proper filtering system, and so on.
  • The product range is large and updates regularly.
  • Cleaning up product data currently eats human hours that could go elsewhere.

If you recognised your own store, or you're still not sure, get in touch and we'll take a look.