Security and control
Security. On the client’s infrastructure, under its control.
The code and the data stay with the company.
/01 · Commitments
Three commitments.
Client infrastructure
- In the company’s cloud (AWS, Azure, Google Cloud) or on its servers.
- Behind its single sign-on and access rules.
- Logs kept in its environment.
Code and data ownership
- Code delivered to the company’s repository, documented, as the project goes.
- No lock-in: an internal team or another provider can take over.
- No model trained on the company’s data.
Model choice
- Claude, GPT, Mistral or an open-source model hosted in-house.
- Chosen task by task: quality, cost, data sensitivity.
- An architecture that allows switching without rebuilding everything.
/02 · Build practices
Build principles.
/01
A person decides
Every outward action (sending, publishing, paying) follows permissions set by the company, one-off or standing. At first, the AI prepares and a person approves.
/02
Every action logged
What the system read, what it did, who authorized it: everything is traced and available to the teams.
/03
Permissions inherited from the systems
A solution only accesses what it needs, and a person sees through it only what they could already see.
/04
Measured quality
Evaluation sets approved by the business teams, replayed before every update; monitoring in production.
/05
AI that says it is AI
An agent or assistant that talks to customers or suppliers introduces itself as AI, and hands over to a person when asked.
/06
Sourced answers
When an answer relies on a document, it cites it. Without a reliable source, the AI says so rather than making things up.
/03 · Security review
Security questionnaire, answered before the start.
The architecture adapts to the company’s requirements.