The distributor was already digital. Every product was in the billing system, priced, in stock, ready. The marketplace was live. And customers were not buying.
The reason was visible the moment you looked at a product page. The listing said SOFT 100ML. No proper name, no description, no photo. That is a perfectly good entry in a ledger, where the person reading it already knows what they sold. It is useless to somebody deciding whether to buy.
Their data was not missing. It was written for the wrong reader.
Why nobody had fixed it
Because it is hundreds of items per distributor, and each one needs a real name, a description, the pack size pulled out of the text, a category, a brand and a photo. It is a week of somebody's life for one distributor, it has to happen again for every new one, and it is the least interesting work in the business. So it stays undone, and the marketplace stays quiet.
What we ran
One pass over 703 listings. It read each entry, proposed the real product name, generated a description, pulled the pack quantity out of the shorthand, and assigned a category and brand. 656 names were cleaned. Local-language entries needed four separate rounds and 870 corrections, because the raw text arrived with prices, punctuation and quantities collapsed into the product name.
Then per-distributor passes as new ones came on: 42 items, then 102, then 111 for a new brand onboarding.
The mess underneath
Cleaning the listings exposed the brand list. It had 242 entries and should have had about 80. One household brand appeared eight separate times; another four times. A hundred and twenty-three brands had no products attached at all — leftovers from imports nobody had cleared.
It went down to 81. Everything was backed up before deletion, and afterwards all 1,280 products were checked to confirm none had been left pointing at a brand that no longer existed. That verification step is not optional. Merging a brand list is exactly the kind of tidy-up that silently orphans a hundred products.
Separately, 1,317 product photos were moved to where they belonged across nine distributors and checked byte for byte afterwards. On one brand's range, official pack photos replaced phone snaps: storage dropped by around 40% while 115 of 116 images got physically larger on screen. Better pictures, less storage, because the originals were oversized phone captures of a small printed pack.
What actually changed for them
Products that had been sitting at the pricing stage, unlistable, became sellable. A new distributor's range now goes live in a session rather than waiting for somebody to find a week. And the catalogue stopped being the reason to delay onboarding the next one.
That last one is the real result. The bottleneck on a marketplace like this is never the software. It is that every new supplier arrives with a spreadsheet written for their own bookkeeper, and somebody has to translate it before a customer can buy anything.
The transferable part
Almost every established business has this. Data that is complete, accurate, and formatted for the person who entered it rather than the person who now needs it. Inventory written for stock control being asked to serve a website. Job records written for invoicing being asked to answer a customer. Notes written for a colleague being asked to train a new starter.
This is the most reliable place to start, because the information already exists and you can check the output against it. Nothing is being invented. It is being rewritten for a different reader — and that is a job where being wrong is immediately obvious, which is exactly the kind of job to give a machine first.