An MCP server makes the ERP available to the AI assistant the company already uses. The question is written in plain language, and the answer comes from the real data, with the permissions of whoever asks.
All the data is in the ERP, but getting it out means knowing where to look. People who do not work in it all day ask those who do, wait for Monday's report or export to Excel and redo the sums by hand.
Meanwhile some people in the company already use an AI assistant to write and summarise, but that assistant knows nothing about the ERP. The most useful part is exactly what is left out.
We expose the functions you need (find a customer, order status, stock level, revenue by period) through the Model Context Protocol, the standard assistants use to talk to external software.
The assistant sees what its user sees, because it goes through the same authentication. A salesperson who cannot see margins in the ERP cannot see them here either.
Questions first, then, if needed, actions, such as a draft quote or a note on a customer. Every write asks for confirmation and stays in the logs.
The same server works with Claude, with ChatGPT and with the agents you build later. If the tool changes, the work does not need redoing.
These are the numbers we set in discovery, before writing any code. The starting value is measured on your data, and the prototype is judged against it.
An assistant nobody uses is useless, however accurate it is. It is the first number to look at, per person and per department.
How many still reach whoever manages the ERP. It is the time that goes back to that person.
A sample checked against the ERP. One wrong answer on the numbers costs more than ten right ones.
If the ERP has no API and no readable database, that access has to be built first, and sometimes it is the longest part of the project. If instead the questions are always the same five, a dashboard is enough.
In Kleio, our own ERP, questions about the data were the AI feature users adopted first, and by a wide margin.
How we put the models inEveryone is pitching something with AI and nobody in the company can evaluate it. A technical counterpart who lines up the use cases and decides with you what to buy and what to build.
Orders arriving by email, as PDFs, on WhatsApp. An agent reads them, checks prices and availability in the ERP and prepares the order for confirmation.
A free 30-minute call, with no commercial follow-up unless you want it.