Skip to main content
A vendor record is small, and you need it before the first purchase order. Typing one out of a business card at the end of a showroom visit is the kind of job that gets postponed until the order is already late. Atelier can read one instead — from a photograph, a document, or the vendor’s own website — and hand you a filled-in form to check. Nothing is saved until you confirm it, so the worst case is a form you close. Each import costs one AI credit.

Three ways in

The first is the usual door, and despite the name it takes more than cards: a specification, a quote, an invoice, a screenshot of an email, or a website address. Inside an existing vendor the offer narrows to a file — there is no website field there — and it only fills what is blank. Anything you have already written stays as it is, so reading a quote into a vendor you have been keeping by hand cannot overwrite your own work.

What comes back

The name, what they supply, and any contact details that were actually there — an address, a phone number, an email, and a named person if the source names one. It arrives as an ordinary vendor form, filled in. Correct anything that is wrong, add what is missing, and save.
Blank fields are usually correct, not a failure. Most supplier home pages carry no address or phone at all, and a vendor read from one comes back with a name, a category and nothing else. That is the honest answer for that page.

Why it leaves fields empty

This is worth understanding, because it looks like the import underperforming when it is the opposite. A model asked to fill a form will fill it. Given a page with no contact details on it, earlier versions of this import returned a phone number, a street address and a sales manager — none of which existed anywhere on the site. A plausible invention is worse than a blank, because a blank is obviously unfinished and an invention is not: it gets saved, and six weeks later an order goes to an address nobody checked. So the details that must be copied rather than worked out — email, phone, address, website, and the whole contact person — are checked against the source before they reach you, and dropped when they are not in it. A contact whose name is not in the source is dropped entirely, since a name is the one part of a person that cannot be inferred. The name and category are different: those are read off the page as a whole, so “Lighting” is a fair answer for a site selling pendant lamps even though the word may not appear.

Websites and documents differ

A website import is checked against the page’s own text, so a detail either appears on that page or does not appear in your form. A document import — a card, a quote, a scan — has no such text to check against. Obvious dummies are still struck out, but a plausible-looking phone number on a blurry scan cannot be told from a real one.
Give a document import a glance before saving, particularly the phone and email. That is the one path where a wrong value can look right.

When it cannot read something

It says so, rather than producing something that looks like an answer. A file that is too blurry to read, a document that turns out not to be about a supplier, and one that names no company each get their own message, and nothing is created. A website that is slow, script-heavy or behind a sign-in may time out. Try the vendor’s contact or about page rather than a deep link into their catalogue, or add the vendor by hand — the record is only a few fields.

Afterwards

Everything the import fills in is an ordinary vendor record. It has no special status, and nothing about it is locked: add the logistics contact it could not know about, write down the terms you agreed, attach the quote.
Import at the moment you meet the vendor, not at order time. The card is in your hand and the website is open; six weeks later you will be reconstructing both under pressure.