Expertise · AI agents
We do not build demos: our agents run in production, inside our own products, every day. We know what it takes for them to be useful, bounded and safe.
Figures as of 30 September 2026, taken from our own operations.
Three agents, three jobs, one method.
AI host and assistant
A voice for the LeBisou community, and a day-to-day assistant.
Anna welcomes new members and hosts the chat of LeBisou, a community of over 135,000 members. The same agent also assists management by SMS and by phone: taking messages, reminders and outreach.
Operations agent in Slack
An AI colleague for the tech team, with deliberately limited rights.
Alfred lives in Slack. You mention him to query the cloud infrastructure and databases: he answers briefly, with evidence. He was cloned from a general-purpose agent whose skills we kept, but not its authority.
General-purpose engineering agent
An agent that builds, ships and monitors our products.
Hermes works from Discord: it writes and tests code, drives CI, deploys, watches servers, handles emails and publishes content. This very website was designed, translated and put online by it in one morning, under the supervision of Eddy Fayet.
What separates an agent that helps from an agent that worries you.
We clone an agent's skills, never its authority. Each agent gets its own credentials, limited to its scope.
Free to read, approval to write. Irreversible or financial decisions are validated by a person.
One agent, one container, one state. A mistake in one agent does not spread to the others.
A log of every interaction and action: you can always explain what an agent did, and why.
The right model for the right task: small models for triage, top-tier models for anything the customer sees.
Beyond agents, AI is built into each of our products, wherever it serves a use case.
35 tools expose bookings, flights, members and maintenance to AI agents. Authentication follows each conversation, so the agent inherits the member's rights, nothing more.
A club's emails become proposed documents or news items, approved by a manager before publication.
Photos, selfie certifications and chat are filtered automatically; doubtful cases go to a human team.
Each has its own style. They compose trios of proposals the customer refines, for 116 identities delivered since March 2026.
Classification, extraction or writing: each task goes to the model with the best quality-to-cost ratio.
From first idea to production, with the same standards as our own products.
Spot the tasks where an agent creates value, and those where it must never act alone.
Slack, Discord, SMS, voice or web: an agent installed where your teams work, with its own personality and scope.
Expose your tools and data to agents safely, with authentication and per-user rights.
Cloud rights, isolation, logging, human approvals: the part that lets you sleep at night.
An AI agent is an assistant that does more than answer: it uses tools (databases, email, browser, APIs) to complete a task, with defined rights and scope.
Yes, provided it is framed properly: minimal rights, free to read and approval to write, isolation and logging. This is the method we apply to our own agents.
We pick the model per task (Claude, OpenAI, Gemini…) and use routing that keeps cost proportional to value.
With a narrow, measurable, low-risk use case. A first useful agent goes live on a simple scope, then grows once trust is established.