The AI that runs the business with its own data
We built a data platform where AI analyzes the business, recommends the purchases and even sends the orders to the supplier. Over the client's real data, on their own server.
They had the data, but couldn't talk to it
Several pharmacies, each with its information on its own. To know what was selling, what was piling up or what to buy, someone built Excel sheets by hand. And worse: the sales figure their own software gave was miscalculated, it didn't count the part Social Security pays on prescriptions, so they'd been making decisions with a number that fell short by almost half.
A data platform with the AI inside
We gathered all the pharmacies' information in one place, clean and reliable, and connected the AI directly to it. Now management can ask in plain language ("which product dropped this month?", "what should I buy?") and the AI answers with the real business figures. And it doesn't stop at answering: it recommends the purchases and even sends the order to the distributor on its own.
How it works, step by step
- 1
We gather the data
Every 10 minutes, sales, purchases and stock from all the pharmacies come in on their own and are organized into a single model.
- 2
The AI reasons over it
We connect the AI to that data with our own servers. Ask anything about the business and it answers with your exact figure.
- 3
It recommends and acts
It analyzes stock, margins and expiry dates, proposes what to buy and sends the real order to the distributor.
- 4
All in your house
It runs 24/7 on the client's own server. Their data doesn't leave, and the software is theirs.
Under the hood
For the curious: the real engineering behind it.
- Python ETL that ingests over SFTP every 10 minutes, idempotent, over a DuckDB analytical database in a star schema
- Over 1 million sales lines, 21,255 purchase orders (330,829 lines) and 138,000 phone-system calls, unified
- Two of our own MCP servers that give the AI access to the data and the accounting
- Real orders sent to the distributor via Fedicom v3 (REST + JWT)
- Invoice pipeline OCR (Mistral) + extraction (Gemini) that posts into the accounting, verifying totals
- Pseudonymized e-prescription (GDPR, fail-closed). It all runs on the client's own cloud server, no lock-in
The result
| What we measured | Before | Now |
|---|---|---|
| Sales figure | miscalculated (~50% short) | reconciled to the cent (Δ €1.31) |
| Unified sales lines | scattered per pharmacy | +1,000,000 |
| Purchase orders integrated | by hand | 21,255 |
| Questions to the business | manual Excel | in natural language, instantly |
The wow
The AI doesn't just look at the data: it corrects it, decides and executes. It reconstructed a sales figure that the client's own software miscalculated by almost half and reconciled it to the cent (eleven millionths of deviation), and over more than a million real lines it recommends purchases and sends actual orders to the distributor. All on the client's server, running on its own, day and night.
Want something like this in your business?
If you have data everywhere and decisions made by gut feeling, this same thing can be built over your business.
Tell us your case and we'll tell you what can be built. The first analysis is free, with no commitment.