Trust and security
Your data, under your control: hosted where it is allowed to be, anonymized before it leaves, validated by people.
Privacy, sovereignty, governance and interoperability decide whether an AI system is acceptable at all. We treat them as design inputs, from the diagnostic to the run, whether the system lives in the public cloud or on your own servers.
- +Hosting matched to the data: cloud, Azure France, private or on-site
- +Anonymization before any external model
- +Human validation on sensitive decisions
- +GDPR and AI Act handled per use case
01 — Principles
Trust starts with explicit design choices.
Anonymization first
Personal data is detected and masked in the pipeline before any model call, with our own detection models running on-site or in France.
Data minimization
Move only what the workflow needs, and make data movement visible in the architecture.
Identity-aware access
User context, roles and permissions stay attached to the workflow from the start, including for agents and MCP tools.
Traceable behaviour
Logging, approval checkpoints and evaluation show how the system behaves over time.
Interoperable stack
Architectures that can change models, frameworks, runtimes and hosting as requirements evolve, independent of your ERP vendor.
Isolation by client
Our multi-tenant platform isolates each client's data, with dedicated isolation for regulated customers.
02 — Deployment patterns
Hosting matched to the sensitivity of the data.
Pattern A
Public cloud
Azure or another cloud, with the region, isolation and key management your policy requires. Fits most product work and non-sensitive data.
Pattern B
Hosted in France
Azure France region or sovereign cloud, operated by us under the run subscription, with isolation per client. For confidential business data.
Pattern C
Private or on-site
Open-weight models on your servers when data sensitivity, connectivity or policy require it. Pilots can run on our GPU infrastructure in France before you invest.
Pattern D
Hybrid
The sensitive path stays in a controlled environment; an anonymized path uses external models where quality or cost justifies it.
03 — Governance controls
Four questions to answer before wider rollout.
What data enters the model path?
Source systems, retrieval scope, retention, what is anonymized and any excluded content or records.
Which actions can the system take?
Tool access, write permissions, approval requirements and escalation for higher-risk operations.
Who reviews behaviour and changes?
Owners for logging, evaluation, model updates, prompt changes and sign-off, on your side and ours.
How easily can the stack evolve?
Options preserved across models, frameworks, inference infrastructure, hosting and connectivity layers.
04 — Compliance
GDPR and AI Act, handled as part of the work.
Compliance is part of every engagement, not an afterthought. Each use case is classified and documented, and the obligations follow from that classification.
GDPR
Processing records for each use case, data protection impact assessment when required, and a data processing agreement (Article 28) with every client.
AI Act
Classification of each use case; the commercial automation cases are minimal-risk in principle. Transparency obligations for AI-generated content are handled in the design, and evaluation results serve as documentation.
Security practices
Client isolation, encryption in transit and at rest, access logging, least-privilege tool permissions and reviewed changes to models and prompts.
Training data
Our anonymization models are trained only on commercially licensed data. Your data is never used to train models for other clients.
05 — Buyer checklist
Questions you should expect to answer with us.
- +Where will the model run, and where is the data hosted?
- +Is personal data anonymized before any external model?
- +Which systems can be queried, and with whose permissions?
- +Who validates an output before it leaves?
- +What logs are retained, and for how long?
- +How is quality measured, and what gates a change?
- +What changes require review, and by whom?
- +How portable is the architecture if a model or host changes?
Use governance as a design input, not a late-stage objection.
A good discovery workshop surfaces the constraints early enough to shape the right deployment model.