Nordic B2B just tipped over. Litium's 2026 report surveyed 915 decision-makers across Sweden, Norway, Denmark and Finland in January, and the headline is hard to miss: AI is now named the single most important driver of where digital commerce is heading, digital already accounts for 31% of total B2B revenue, and 81% of firms expect AI to land hard before the year is out.
Wholesale is where this moves quickest. But every week I run into the same awkward thing underneath all that ambition. The catalog is a mess.
What nobody says out loud about AI and product data
AI doesn't fix bad data. It scales it.
Point a model at a source catalog with half-empty attribute fields, three spellings of the same colour, and a "material" column that sometimes holds a fabric and sometimes holds a marketing slogan, and it will cheerfully produce five languages of fluent, confident, wrong. At volume. Across every channel you feed it. The output reads polished, and that is precisely the trap, because nobody re-reads 4,000 auto-generated Finnish descriptions. It is the same failure mode we describe in stopping bad data at the gate: the errors that hurt most are the ones that look fine.
So the real question was never "should we use AI on our catalog." Obviously you should, and your competitor in Oslo already does. The question is whether there's a clean layer sitting between your supplier feed and the model, and a person in the loop where it actually matters.
AI doesn't fix bad data. It scales it. The output reads polished, and that is precisely the trap.
Why the Nordic channels punish weak data in particular
This isn't theory. The channels Nordic wholesalers actually sell through are ranking and matching engines, and they read your data literally.
Take CDON, the largest Nordic marketplace, with more than 1,500 merchants across four countries. It runs a quality scoring model that ranks listings, so richer content sits higher, and it sorts merchants into four performance tiers. EAN or GTIN, MPN and brand have been required in your product data since October 2021. Descriptions have to be in the local language: Swedish for SE, Danish for DK, Norwegian for NO, Finnish for FI, with English only as the international fallback. Titles go up to 150 characters, descriptions up to 10,000. Miss the localisation and you don't just drop in visibility. You convert worse, in a market where buyers are genuinely visual and detail-driven.
The price comparison engines are stricter still, because they're pure feed machines. PriceSpy and Prisjakt want nine mandatory fields, and your category has to map to theirtaxonomy, not the one you dreamed up internally. PriceRunner, owned by Klarna since 2022, matches on EAN or GTIN and needs every variant as its own feed row. A shoe in seven sizes is seven rows, each with its own EAN and SKU. Kelkoo throws out any product URL that doesn't start with http:// or https://. And the single most common reason a feed gets rejected, across all of them, is a category that doesn't match the platform's taxonomy. Getting the same source data into each of these formats cleanly is exactly the job of product feed management.
None of those rules care how good your AI copy is. They care whether the underlying data is complete, structured and mapped correctly.
| Channel | What it reads literally | Where weak data hurts |
|---|---|---|
| CDON | Quality score ranks listings, four merchant tiers | Thin local-language description sits below a richer competitor |
| PriceSpy / Prisjakt | Nine mandatory fields, maps to their taxonomy | A category that does not match gets the feed rejected |
| PriceRunner | Matches on EAN/GTIN, one feed row per variant | Variants not split per size never match, never appear |
| Kelkoo | Rejects any URL not starting with http:// or https:// | A single malformed URL drops the product |
The channels are ranking and matching engines. They read the data, not the polish.
The multilingual wall is where wholesale really breaks
Here's the part that hits wholesalers hardest. You're not selling into one Nordic market. You're selling into as many as five languages, and Finnish is not a Scandinavian language. It's Uralic, a whole different family, and models trained mostly on Swedish, Norwegian and Danish routinely fumble it. Norwegian, on top of that, has two written standards, Bokmål and Nynorsk, and they have to be handled apart. This is why AI descriptions and translation need a human on the languages you can't proofread yourself.
The old way over this wall is brutal. SHOPLAB puts professional translation of a 3,000-SKU catalog across the five Nordic languages at EUR 50,000 to 100,000, plus months of calendar time. That's the manual tax. AI shrinks it to days, which is honestly why adoption is spiking. But do it on top of a dirty source and you've just paid to publish the same mistake, translated flawlessly, in five markets at once.
An example I walk new clients through. Say you wholesale an outdoor jacket, and the supplier feed lists the shell as "Fleece." One weak source field, multiplied across channels and languages, becomes the entire problem in a single product.
One weak source field, multiplied across channels and languages. That's the entire problem in a single product. It usually starts upstream, in the raw supplier feed, which is why merging supplier data into one golden record matters before any model touches it.
A clean layer underneath, a person in the loop
The way we've built SyncRefine, and the way I'd argue about it even if you never touched our product, is that AI should do the heavy lifting with experts working next to it, never on its own. A few things I'd treat as non-negotiable either way.
Do that, and AI really is the growth lever those Litium numbers promise. Skip the clean layer, and all you've automated is your backlog of errors, now shipped to four more countries.
Before you scale
If you're a Nordic wholesaler about to point AI at your catalog, I'd ask one blunt thing first. If you exported your source feed right now and read ten rows at random, would you trust them enough to publish as-is in Finnish? If the honest answer is no, AI won't save you. It'll just get you to wrong faster.
So what does your source feed actually look like on a Tuesday, before anyone tidies it up for a demo?
Tell us in half an hour how your product data travels today, from your suppliers to CDON, PriceRunner and the comparison engines, and we'll say honestly which fields derive cleanly and which need real editing. If your data still lives in a spreadsheet, start with Excel chaos to a catalogue. To keep those errors out before they scale, read product data quality at the gate. And if you're weighing whether all this needs a PIM, we lay out the trade-off in do I need a PIM system.
