BelOrta is one of Europe's largest fruit and vegetable auctions. Every day, orders, delivery questions, changes and complaints arrive by email from supermarkets, wholesalers and caterers — and every one of those emails is a link in the operation. Cloudway built the AI mailbox assistant that now sorts them, without BelOrta changing a single thing about how its people work.
BelOrta works with three specialised teams — Vegetables, Fruit and Organic — and each runs its own mailbox. On paper that is a clean split. In practice, customers do not know which team handles what, and they have no reason to care. A question about conventional fruit arrives in the organic mailbox. An order goes to all three at once, on the reasonable assumption that at least one of them is the right one.
The volumes turn that into real work. The vegetables team receives around 110 orders a day, 25 of which are not theirs. Fruit receives 70, of which 50 are meant for someone else. Organic receives about 100 a day — and roughly 90 of those turn out not to be relevant.
So before anyone can act on an email, somebody has to read it, work out whose it is, and move it. Every day, in three mailboxes. BelOrta wanted that to stop, without adding a new system for their teams to learn.
The design decision that mattered most was made early: no new platform. The assistant sits behind the existing Outlook environment. Nobody was asked to adopt a tool, log into a portal or change a habit. The intelligence went in behind the scenes.
Every incoming email is analysed automatically — and not just the message. Attachments are read too: PDFs, Excel files and images, which is where the actual order information usually lives.
From that analysis the assistant answers two questions in order. First, does this email belong in the mailbox it landed in? A fruit order in the organic mailbox is flagged as not relevant for organic; a vegetable order in the fruit mailbox is recognised and labelled as such. Second, what kind of message is it — a new order, a question about an existing one, or a complaint? Each relevant email is filed into the right folder on that basis. The mailbox arrives sorted.
Because customers write in more than one language, the classification works on the meaning of the message rather than on keywords, so the same rules hold whichever language an order comes in.
The content stays where it belongs. Emails and attachments are processed inside BelOrta's own environment, are not shared with other organisations, and are not used to train external AI models.
The most interesting problem on this project was not technical. It was working out what counts as fruit.
Botanically, a tomato is a fruit. At BelOrta a tomato is a vegetable, because that is how its customers order it and how its teams work. Several exceptions like that surfaced during the project, and each one mattered: an assistant that classifies produce correctly in theory and incorrectly in practice is worse than no assistant at all, because now somebody has to check it.
So the solution was tuned to BelOrta's own product logic rather than to a general-purpose one. The system categorises produce the way BelOrta categorises produce. That is the difference between a generic AI model and something that actually fits an operation — and it is usually where the real work of an AI project sits.
This is what a lot of useful AI looks like. Not a transformation program but an everyday process organised more intelligently, with no new tools and no change management, giving back time, focus and oversight. Sometimes innovation is not doing something new, but doing something you already do every day rather more cleverly.
our experts are happy to look at it with you.