Goods receiving in a coffee shop chain
Operations lived in chats and spreadsheets: invoices, purchase prices, a monthly plan. We built a cabinet the agent writes into, and a person reads plan, forecast and prices in one place.
A chain of 10 coffee shops: every day each location receives deliveries — food, milk, disposables. The supplier brings a paper invoice, and it has to be entered into the accounting system (iiko): line items split across the right location, matched against the product catalogue, amounts totalled. This used to be done by hand — and it was a whole separate job that never went away.
What breaks with manual receiving
A barista receiving the goods took a photo of the invoice "so it doesn't get lost", and someone else entered it later — line by line, by hand. Dozens of rows per invoice, and supplier names don't match the names in the system: "Tomatoes red grade 1" against "Fresh tomato", "Cream 10% UHT 1l" against "Cream 10% pasteurized". Every row has to be found in the catalogue or created from scratch, the right location and warehouse must not be mixed up, totals have to be reconciled. Invoices piled up and were entered late and with errors — and at the end of the day it's just retyping a piece of paper into a system, which doesn't make the work more accurate or faster.
Why iiko and manual entry didn't help
iiko is an accounting system: it answers "what's in stock" and tracks balances, but it can't read a paper invoice or a photo of one. Someone still has to retype the document by hand: find or create each item, enter the quantity, the price, the location. The more locations and suppliers you have, the more time goes into mechanics instead of inventory management.
What we did: an agent receives the invoices itself
We built a receiving cabinet and put the agent right into the work chat where the baristas of all locations talk. Receiving now looks like this: the supplier has arrived — the barista photographs the invoice and sends it to the chat. That's it. From there the agent parses the document and takes it all the way to posting in iiko without a human in the loop.
- The agent reads the invoice in any format it arrives in — a photo, a file, an email — and extracts line items, quantities, prices and the total.
- It matches every row against the product catalogue and the iiko nomenclature — "Tomatoes red grade 1" is found as "Fresh tomato", no duplicates are created.
- It allocates deliveries to the correct location and warehouse of the chain.
- It posts the goods in iiko itself: the invoice gets the status "loaded to iiko", and nobody has to retype anything.
- It keeps a price history per supplier: week by week, you can see who raised prices, when and on which items.
- If it meets an item that isn't in the catalogue yet, it creates it and shows it to a human for approval — instead of silently skipping it.
Baristas don't need to learn anything
No new systems or interfaces: people are already in Telegram, where they photograph the invoice exactly as they used to photograph it "for memory". The only skill needed is taking a photo of the document when the goods arrive. Everything else — recognition, matching, allocation across locations, posting — happens on its own.
What stays with the human
Occasionally approving a new item the agent hasn't seen before and glancing at questionable rows. Management keeps watching supplier prices and deciding who to work with next — based on data that now collects itself. The machine allocates and posts, the human reviews and decides: manual entry goes to the agent.