FR | EN

Your data & AI

An AI agent should never see more than its task requires. That is a matter of design, not blind trust — and it is exactly our trade, born of identity and access management (IAM/PAM).

Four data levels

Every piece of information entrusted to an agent is classified before deployment:

LevelExamplesRule for the agent
1. PublicHours, services offered, FAQs, information from your websiteOpen access: the agent's normal working base.
2. InternalInternal procedures, document templates, operating rulesReserved for the "employee" role; never served to the public.
3. PersonalName, contact details, a client's or employee's fileOut of scope by default: the agent escalates to a human instead of handling it.
4. Sensitive or regulatedFinancial or health data, trade secrets, negotiated ratesNever exposed to the agent without a specific assessment, a suitable model (often local), and explicit approval.

Five guardrails, always

1

Least-privilege role access

Each agent has its own identity and only opens documents within its role. A document not authorized for the role is never served — even if asked for directly.

2

Minimization

The agent asks for and processes only the information its specific task needs. Nothing extra is collected "just in case."

3

Logging

Every answer, refusal, and escalation is logged with its source. You can check afterwards what the agent said and why.

4

Out-of-scope refusal

If the information is not in the authorized documents, the agent refuses and directs the person to a human. It never invents a price, a deadline, or a policy.

5

Human validation

A draft is never a send. Every sensitive action — sending, committing, a personal file — goes through a human's decision.

Local or online?

Depending on your data's sensitivity, an AI model can run locally, on a computer at your business, without your information ever leaving your premises; or online, for uses where only public information is processed. The assessment settles this question with you, based on real figures and constraints — it is never a default choice.

Our founding principle: the agent never replaces the human. It serves the human: it lightens their work by automating certain tasks, and always leaves the human in charge of decisions.

See the AI & Sensitive Data Assessment