At the patisserie, the orders aren’t taken by a person
Client
Artyom Kirilovskiy is a pastry chef and the founder of an artisan patisserie in Moscow. Five years in, his own workshop, a team, supply to cafés and restaurants. His signature piece is sourdough panettone: three days from starter to box, and those three days can’t be rushed.
Everything about his business already works: a website with a catalogue, delivery across Moscow and the country, production running smoothly. This isn’t someone who "needs automation" — it’s a craftsman with a business that already runs.
Problem
The website handles a simple purchase: pick a box, pay, receive it. But an artisan patisserie doesn’t live off a box on a shelf — it lives off conversation. "Will this feed ten people?", "Can you leave out the nuts?", "Do you deliver to Kazan?", "Can you have it ready by Saturday?" A catalogue doesn’t answer questions like these.
So part of every order was closed by a person — by hand, in chat. And it all came down to that one person: a question would arrive at midnight, the answer came in the morning, and by then the customer had already ordered from whoever answered first.
The specialist
We built Pudra for incoming messages — an AI specialist that carries that same conversation through to a paid order. That’s what it looks like — the recording above.
Cards expand on hover
Collects the items, the occasion, the date and the address, and writes the order into the patisserie’s database — from there it goes straight into 1C. Not into a chat thread where someone later has to go hunting for it.
Takes the customer’s address and works out the real cost: courier within Moscow or CDEK nationwide. States the time and the price right in the conversation, before the order is placed.
Sends a QR code for instant bank transfer straight into the chat and checks the payment status itself. No one asks the customer "did you pay?" or checks a bank statement by hand.
The owner types "Napoleon cake is now 3,200" — and in the very next conversation it quotes the new price. No control panel, no "email the developer."
Keeps a history of orders and preferences: what they bought, for what occasion, what didn’t work. A regular customer doesn’t have to explain themselves from scratch every time.
It doesn’t make up an answer. It hands the conversation to a person and leaves them the full thread — no need to work out what was said. Same with failures: a payment doesn’t go through, delivery falls through — the bot writes to the owner itself and asks for help.
The marketing bot
The second specialist faces the other direction — toward people who’ve already bought something. The owner doesn’t build the mailing list by hand or export spreadsheets: they say what they want to say, and the bot helps work out who to send it to and in what words.
There are four kinds of broadcast, and they form a ladder from "the same message to everyone" to "something different for each person."
The owner gives the text — it goes out to the whole list as is, a photo included if there’s one. This is a plain announcement: we’ve opened, we’ve moved, we’re working through the holidays.
One occasion — "seasonal strawberries" — but each message is written separately: with that person’s tastes, their past orders, their own way of being addressed. One topic, a hundred different letters.
The bot picks out who this is worth offering to — "people who order for children’s parties" — and writes only to them. The owner sets how strict the selection is; in a test of five people, three received the letter.
The top step: the bot picks a product from the catalogue for that specific person and builds an offer around them. Not "20% off everything," but "here’s this, for you, and here’s why."
Before sending, it shows what the message will look like and how many people will get it. It sends only after approval — never on its own.
It learns from its own broadcasts
After sending, the bot comes back and checks what came of it — for every recipient separately. It doesn’t take a purchase on faith: it checks the orders that came in during the three days after the message and ties them back to it.
Here’s what it worked out on its own, word for word — nobody wrote it these rules:
- on priceShowing "was / now" heads off the "too expensive" objection before it’s even raised
- on the productState the weight, the filling and the shelf life right in the message, without waiting to be asked
- on peopleLeave out of the next broadcast anyone who’s complained before about getting too many messages
- on discountsTie a discount to the person’s own occasion date, not to the end of the week
The lessons build up separately for each kind of broadcast: the first one is already on its twenty-third revision.
How this could be useful to you
If your business lives on conversation — questions in chat, clarifications, "could you do it like this" — you probably already know what happens next: you need to answer fast, and there’s only one of you. The specialist takes that conversation on and carries it through to a paid order, bringing you in only where it genuinely needs you. It doesn’t have to be a patisserie: a custom-order craftsperson, a salon booking appointments, a tailor fitting a dress — all of them have to talk a customer into a purchase too.