Service 13 · Useful AI, properly scoped
A custom AI agent, for one specific task, approved by a person.
An agent does more than answer: it reads, sorts, pre-fills, suggests. Well scoped, it frees your teams from a repetitive task. Badly scoped, it sends out wrong e-mails one after another. We build agents with guardrails, human approval on anything that commits you, and a capped budget.
First conversation free.
What exactly is an AI agent
A language model, tools it can call (your e-mail, your CRM, a document base), a sequence of steps and a control system. An assistant talks; an agent acts: it reads your e-mails, sorts them, writes a draft, feeds your ERP, suggests a follow-up. A good first use case follows three rules: volume, a tolerable error, a reversible action.
- 01
Sorting and qualifying incoming e-mails. The agent reads, sorts, prioritises and suggests a standard reply.
- 02
First drafts of quotes. From a short brief and your templates; a person approves and adjusts before sending.
- 03
Extraction from your documents. Invoices, contracts, PDFs: the agent reads them and feeds your ERP. The recognition rate is measured during acceptance testing, on your own documents, before the agent goes into service.
- 04
Summaries of calls and meetings. Transcription, summary, action points: the follow-up documents itself.
- 05
Searching internal documentation. The agent finds the right document and cites its sources.
Who it is not for
A task where a mistake cannot be put right, or an agent that would send e-mails to your clients on its own. That does not go into production with us without human approval.
The agent prepares, the person decides
We do not build agents that replace your teams, but workflows in which AI does the repetitive work and people keep the decision. The agent prepares, suggests, pre-fills; the person approves, adjusts, commits. That is the model that makes it into production and that teams accept, not the full automation that frightens people and gets things wrong in silence.
Tools
Claude, GPT or Mistral; connectors with a controlled scope to your CRM, your ERP and your databases; a .NET or Next.js application, hosted on our servers in France.
The guardrails, in every agent
- A cap on steps: the agent never loops indefinitely
- A capped budget for model calls: a bug cannot blow up the bill overnight; alert at the threshold
- Human approval on any action that commits you: sending an e-mail, writing to a database, a commitment to a client
- Read-only by default: write access only covers a scope approved with you
- A log of every action and costs tracked continuously: an agent you can audit at any time
Translate, compose, check, correct
An agent is data processing like any other: it runs on our infrastructure, measured in the same way.
- OVHcloud data centres housing our hypervisors
- 3 in FranceOVHcloud public API (dedicated server records)
- Active monitoring probes across our machines
- 58Internal Uptime Kuma monitoring
Measurements taken on 5 September 2026, values recorded by our own instruments.
Before handing a task to an agent
How is it different from a chat assistant?
An assistant answers questions; an agent decides on actions, carries them out through tools, observes the result and starts again until it reaches the objective. It does the work instead of talking about it, which is why it needs guardrails.
Can an agent do whatever it likes in my systems?
No, and that is deliberate: a cap on steps, a capped budget, read-only by default, human approval on any action that commits you. An agent that suggests a draft for approval is safe; an agent that sends things on its own does not go into production with us.
Do we need a technical team to run it?
No. Your teams use it every day with no technical skills: it presents what needs approving and raises an alert if something looks wrong. On the technical side, light maintenance is enough for cost tracking, adjusting instructions and model updates.
How much does it cost?
The prototype is priced during scoping, release to production is billed by the day, and the model's usage costs are tracked and capped. Price ranges are published on the pricing page.
Timelines, quotes, code ownership, end of contract: the answers are in the FAQ.
Quoted per project. Price ranges are published on the pricing page, along with what makes a budget vary.
Your first agent
A repetitive task to hand to an agent?
Describe the task, its volume and what happens when it is done badly. We will price a prototype, or tell you that now is not the right time.
Reply within 24 working hours, from one of the two founders.