An AI agent connected to your ERP or CRMIt reads, it checks, it prepares. Your team validates.

An assistant that answers questions is useful. An agent that fetches the information from your management software, prepares the quote or the reminder and submits it to you is what saves hours. The difference comes down to one thing: the connection to your real tools, with precise rights and a trace of every action.

Book a free diagnostic

Your situations

What your team still does by hand

Retyping a purchase order received by e-mail into the ERP, line by line.

Checking three tabs to see whether the customer paid before answering their reminder.

Preparing a standard quote from a request e-mail, fifteen times a week.

Updating the CRM after every meeting, when someone remembers.

Answering internal questions about stock, lead times and prices.

Chasing unpaid invoices, late, with no history of the exchanges.

What we deliver

What an agent can take over

Reading and data entry

The agent reads an e-mail, a PDF or a form, extracts the data and prepares the record in the ERP or CRM. You validate with one click, it saves.

Documented answers

A question about stock, a lead time, an order: the agent queries the software and writes the answer with the exact figure and its source.

Quotes and reminders prepared

From a request and your pricing rules, it prepares the quote or the reminder in your tool, ready to be sent by the right person.

Watching your data

Overdue invoice, stock below threshold, file with no follow-up for two weeks: the agent flags it and suggests the action, it does not act alone.

How it works

How we connect it, in six weeks

01
Scope and rights
We choose one task, write down what the agent may read, prepare and do, and what it may never do.
02
Connection to your tools
Odoo, Sage, HubSpot, Pipedrive, Dolibarr or in-house software: we connect the agent through the software's API, via the MCP standard, with a dedicated account and limited rights.
03
Supervised first version
For two weeks the agent suggests and a person validates everything. We measure errors and adjust the rules.
04
Production and log
The agent runs within its scope, every action is logged, a weekly report tells you what it did and where it hesitated.

Assistant or agent: the difference that matters

An assistant waits for a question and answers with what it knows. An agent receives a goal, fetches the information where it lives, chains several actions and comes back with a result ready to validate. The second is more useful, and more demanding: it needs access, so access must be limited, and everything must be traceable.

Three rules we do not negotiate

A dedicated account, minimal rights. The agent has its own account in your software, with only the rights its task needs. Never an employee's account.
It prepares, a person validates. Sending a quote, changing a price, deleting a record: these actions stay human. The agent prepares and submits them.
Everything is logged. Every read, every suggestion, every validated action is traced, with the source data. You can review and undo everything.

Why the MCP standard

Every tool has its API, and every AI vendor would like to sell you its connector. The Model Context Protocol is the open standard that describes to a model the actions a piece of software exposes. We use it for two reasons: the connector we write for your ERP works with any model, and you can switch models without redoing it.

With which software

ERP: Odoo, Sage, Dolibarr, EBP, or software built for you. We have connected agents to our own ERP, which taught us what breaks.
CRM: HubSpot, Pipedrive, Brevo, Zoho, or your ERP's CRM module.
E-mail and documents: Microsoft 365, Google Workspace, a file server.

If your software has no API, we say so before starting, and we suggest another task.

What it costs, and what it returns

A first agent takes six weeks, between twelve and eighteen engineer days. Running costs depend on volume: a few tens of euros per month for an agent handling a hundred e-mails, a few hundred for thousands of documents. On the other side, count the entry and search time your team spends today: it is usually the simplest measurement to make, and we make it with you in week 1.

References

The Model Context Protocol: the open standard that connects a model to your software.
The EU AI Act: transparency and human-oversight duties.
The CNIL's guidance: personal data and AI.

Proof

What we have already connected

In our own ERP: an AI check of accounting entries that flags anomalies before closing, and serial-number reading from photos for inventory.
For an engineering office: a tool that reads tender bids and prepares a justified score, reviewed by the buyer before deciding.
For a health organisation: an assistant that queries a territorial knowledge base and prepares a shareable summary, without ever storing personal data.

Frequently asked questions

Questions we are asked about agents

Can the agent do something stupid in our ERP?

It can only do what its account allows, and the actions that commit you, sending, price changes, deletions, remain subject to human validation. For the first two weeks everything is validated; then you decide, task by task, what may go through without review.

Do we have to give access to all our data?

No. The agent accesses only the tables or objects its task needs, through a dedicated account. It does not read payroll to prepare a quote. The scope is written down in week 1 and verifiable in the software's rights.

Which AI model do you use, and where does our data run?

The one that fits the use case and your constraints: a model hosted in France or in Europe when required, a model installed on your own server for sensitive data. Thanks to the MCP standard, the model is replaceable without redoing the connection.

Our software is old and has no API, is it possible?

Sometimes, through the database or exports, but it is more fragile and we say so. In that case we suggest starting with a task that does not depend on that software, or looking at a migration if one is planned anyway.

How long until a first agent is in production?

Six weeks for a well-delimited task, with a supervised version by the third week. The following agents are faster: the connection to the software is already done.

Which task would your agent take first?

Thirty minutes to choose it together and check that your software is ready for it.