We do not sell a chatbot. This is the story of how an AI engine learned to read the WhatsApp orders of a real restaurant — with its quirks, its pre-bookings and its lifelong customers — and gave the owner hours back every day.
A lunch delivery restaurant — a thriving family business, ~140 orders a day — ran on its owner's shoulders: hours every day reading WhatsApp chats, writing tickets by hand, matching payments from memory. The business had a unique competitive edge (it sells from the afternoon before and produces against confirmed orders, with almost no waste), but growth was capped by manual administration.
The project's golden rule came from reality itself: no method could add work for the owner. The AI had to feed on what already existed — the chats exactly as they arrive — or it was useless.
An engine that reads real conversations and produces what the kitchen needs. The hard part was not understanding the easy order — it was surviving the cases that break bots.
"Send me two of the fish one but one without soup and with blackberry juice" — the engine turns it into a structured ticket against the menu in force that day (it rotates daily), with every customization recorded.
It spots orders booked for a future date ("20 for Thursday") and does not mix them with tomorrow's kitchen. It separates several customers writing from the same chat, each with their own ticket. It dates ambiguous orders using the business rules — even the dish on the menu gives away which day it is for.
A dashboard shows the tickets grouped by delivery date, the consolidated production list and the pre-bookings separately. The owner reviews and corrects before dispatch: the judgment is still human — the typing is not.
The muscle that reads orders also analyzes complete industrial operations, every day, without anyone asking.
Four daily analyses of 13 refrigeration units across 2 platforms, with an engineer's judgment: it tells a normal defrost from a real fault and respects each business's operating windows.
Every night, the filtration analysis of more than 40 machines: compliance, anomalies and ranking, straight to management.
Routines that verify the reports themselves went out — if an analysis fails, someone finds out within minutes.
Every engine is built with the owner, around how their operation actually works — it is not a generic model left to guess.
Days watching how the real work flows: where the time goes, where the data gets lost.
WhatsApp, sensors, spreadsheets, invoicing — the engine feeds on what already exists, without adding work for anyone.
The AI drafts and the human approves. Whatever proves itself gets fully automated.
Tell us in a three-line email. We will tell you frankly whether AI can solve it — and how.
Tell us your case