In development
Online reputation agent
Plans the editorial calendar, writes for each network and creates visuals in your brand. Nothing is published without your approval.
Task automation
Custom AI agents
An AI agent is not an assistant you question. It is software that takes responsibility for a repetitive task from end to end: it prepares, proposes, waits for your approval, then executes.
We design these agents around your tools and data, then operate and maintain them for you.
Contact usWhat it does
What an AI agent actually does
The difference comes down to one word. An assistant answers; an agent acts. These are tasks we can entrust to an agent. They are not demonstrations: they are jobs already occupying someone in your organisation.
Under study
Video editing agent
Turns your footage into short videos for each platform, using your titles, colours and expected formats.
Under study
Email assistant
Sorts incoming messages, summarises what matters and prepares recurring replies. You review and send.
Under study
Administrative assistant
Drafts quotations and invoices from your templates and terms, including the follow-ups you have defined.
Under study
UGC video agent
Produces testimonial or demonstration videos from a script for your social channels.
Yours to define
Your agent, still to invent
Which task returns every week, follows the same rules and teaches its owner nothing new? If you have an answer, there may be an agent to build.
A working example
Our first client is ourselves
We are building a communication agent for social media, and 3N is its first user. We will not sell you a tool that we do not use ourselves.
It proposes
A monthly calendar: topics, dates, formats and networks.
You clarify
Open a publication, explain what you want to say and choose your images.
It produces
Copy for each network, a visual in your brand and animation where needed.
You decide
Nothing leaves without approval. Correct it, regenerate it or validate it.
Our guarantees
An agent does not decide in your place
It is the first justified question every leader asks. Our answer rests on three rules enforced in code rather than left as a commercial promise.
Nothing leaves without you
For anything visible outside your organisation—a post, email or quotation—the agent prepares and stops. A person validates. That is the default.
The agent does not invent
A model left alone fills gaps with convincing inventions. We strictly define the subjects and information it is allowed to use.
Everything is checked again
Dates, formats and amounts are checked by code before storage. Incorrect data silently recorded is worse than a visible error.
Already in production
AI in our projects is not new
Before building agents, we integrated artificial intelligence into systems delivered to clients and running in production.
LDLC
An AI assistant inside a wind-farm maintenance application, with mobile and back-end algorithms working together.
Association BVH
An AI chatbot inside an association CRM covering members, donors, projects and accounting.
ATMO Grand Est
Algorithmic processing of noise and pollution sensor data for the Grand Est region.
None of these three is an agent in the sense used on this page. They assist, analyse and answer, but do not carry a task from end to end. We make the distinction because “agent” is now applied to many systems that are not agents.
The two costs
What it costs, and how we know
An AI agent has two costs. First comes design and connection to your tools. Then comes usage: each time the agent works, it consumes computing capacity from the model provider. This second cost is often left unmeasured.
- The number of tasks entrusted to the agent.
- The number and condition of the tools it must connect to.
- The volume processed every month.
- The level of control required before each action.
- Your confidentiality and hosting constraints.
We begin with a measured trial on one task, not with a full construction project. It reveals the cost per operation and the time genuinely saved. You decide whether to continue using your figures, not our estimate.
Where to begin
Identify
Start with a frequent task governed by stable rules and where an error can be recovered. Choose the most profitable, not the most spectacular.
Measured trial
A reduced version runs on real data. We measure output quality, time saved and cost per operation.
Build
The agent connects to your tools with the controls and approvals agreed together.
Launch
Your teams are trained, errors are monitored and the agent is adjusted through its first weeks.
Questions
- Is an AI agent a chatbot?
- No. A chatbot answers questions. An agent takes responsibility for a complete task: it gathers what it needs, produces an output, then executes it or submits it to you. You do not have to prompt it for the work to happen.
- Can the agent publish or send something without my approval?
- Not by default. For anything leaving your organisation, the agent prepares the work and waits for approval. Automatic validation is possible, but it is an explicit decision made task by task.
- How much does usage cost?
- It depends on the task and volume. We measure the cost of a real operation during the trial so that you decide using an observed figure rather than an estimate.
- Do I own the agent?
- No, by design. 3N builds and operates the agent. You use it while we run, maintain and improve it. What belongs to you remains yours: your data, your content and everything the agent produces for you.
- Will my teams need training?
- Yes, but only lightly. The essential skill is reviewing and approving. We train the relevant people at launch, and operating guidance remains available from their workspace.
Which task would you like to stop doing by hand?
Describe the task, its frequency, the tools involved and what must remain under human control. We will tell you whether an agent is the right answer—and say so clearly if it is not.