Method & demo

No big-bang technology rollout: a four-step progression, each step validated before moving to the next.

The four steps

1. Assessment (1 to 2 weeks). Together we pin down the problem, the success metric, the authorized documents, and the escalation rules. This is also where your data classification is decided.
2. Prototype. A working agent is built and tested — on public or fictitious documents only. You see its answers, its refusals, and its escalations before it touches anything real.
3. Measured pilot (4 to 6 weeks). One task, a narrow scope, one before/after metric. Systematic human validation. At the end: an honest report, backed by numbers.
4. Framed deployment. If the pilot justifies it, gradual expansion to other tasks or teams, with the same guardrails.

A recognized framework: ANSSI's seven principles

In October 2026, ANSSI — France's national cybersecurity agency — published its recommendations for securing agentic AI systems (Securing an Agentic AI System). Our method follows them point by point:

• 1. Assess risks from the design stage — every project starts with an inventory of the data the AI will be able to see and use.
• 2. Limit autonomy — the agent works within a documented scope, expanded gradually and never by default.
• 3. Classify actions — what the agent may do on its own, what requires your approval, and what is prohibited.
• 4. Isolate environments — development, testing and production kept separate, with distinct access.
• 5. Enforce authorization outside the model — access is controlled by code and by roles, never by the agent itself; out of scope means a refusal.
• 6. Contain the impact — no secrets entrusted to the agent, limited interactions, shutdown and recovery planned.
• 7. Maintain demonstrable control — activity logs and security tests you can verify.

These recommendations are not a certification: we apply them as a design framework, and we state honestly what gets put in place over the course of a pilot.

The 15-minute demo

We present our demonstration agent on entirely fictitious documents: you will see it answer while citing its sources, refuse what falls outside its scope — including an attempt to make it "ignore its rules" — and escalate a sensitive case to a human with a draft ready for approval. It is the best way to judge our approach before any commitment.

Case shown in the demonstration: a welcome agent for newly arrived families' general questions, developed through exploratory exchanges with the francophone school community of southern New Brunswick, from public information and fictitious documents. No real family or student data is used.

Request the demo

The tests every agent passes before being shown

• Role leakage: the public role never receives internal content.
• Outside the documents: a price or policy absent from the sources gets a clear refusal.
• Citation: every factual answer names its source document.
• Injection: "ignore your rules" gets a refusal; the role's limits hold.
• Sensitive data: no disclosure, escalation to a human.
• Escalation: urgent or sensitive case = draft only, human decision.