Case study · in production, 2026
RFSU AB, Sweden
RFSU's order mail, read, understood, and decided before anyone opens it. An agent workflow reads each incoming order, matches the customer, checks every line against the catalogue, and scores its own confidence. The system's job isn't to handle every order; it's to know which orders it can handle.
- Client
- RFSU AB, Sweden
- Role
- AI agents for operations, solo
- Stack
- Business Central · OCR · RAG
The shape of the problem
Wholesale clients order by PDF invoice. Placing one meant mapping every line to the right SKU, converting the units of measure and keying it into Business Central by hand. Customer support's days went to order entry.
What I learned about the field
Everything a human does can be automated. It just takes time. The support team had run these orders for ten years, and every nuance they carried took months of iterating to teach. The automation came out beautiful; the real work was understanding them.
What runs autonomously
The system handles
- Reads each incoming order document
- Resolves every line to the right item in Business Central
- Converts the units of measure and picks the most efficient packaging
- Places the finished order
People keep
- Anything below the confidence bar waits for a human decision; nothing uncertain ships
- 5
- orders processed a day
- ~20
- items on an average order
Production volume, August 2026.


The confidence gate
The idea the whole system stands on: every order gets a confidence score, and the score decides. High goes straight to a booked order. Uncertain goes to a person, with the reasoning attached. Unknown customer? The system asks instead of guessing.


Outcomes
- Order processing
- End to end, without a person in the loop
- Customer support
- Out of order entry entirely
- Volume
- Five orders a day, ~20 items each
Written by Arsenii Demydov · Published 28 August 2026 · Updated 29 August 2026
