Customer story / Digital health
A French digital health platform keeps patient data on HDS-certified infrastructure while adding AI to clinical workflows
A French digital health platform serving clinics and teleconsultation practices wanted AI-drafted consultation summaries. French health data hosting rules and GDPR Article 9 left little room for error. YTT Secunet enforces a France-only rule for identified health data and pseudonymizes the rest before inference.
- Customer
- French Digital Health Platform
- Industry
- Digital health
- Headquarters
- Paris, France
- Gateway
- Managed gateway
- Policy
- Custom Rego
- Zones
- European Union
Results
- 0
- prompts with an unmasked NIR served outside the HDS environment in the first 90 days
- 1,412
- requests refused at the boundary by policy during development, each with a logged reason
- No findings
- on the AI module in the platform's HDS surveillance audit
- 5 weeks
- from DPIA approval to general availability for participating clinics
The platform stores consultation notes, prescriptions and appointment histories for clinics across France, and it is hosted with an HDS-certified provider as required by Article L.1111-8 of the French Public Health Code. Clinicians asked for AI-drafted summaries of consultations and referral letters, which meant sending clinical text to a language model.
Every option the product team found had a problem. The strongest hosted models were available in EU regions but not always on HDS-certified infrastructure. Open-weight models could run on the platform's own HDS hosting, but their quality on French clinical text was weaker for some tasks. And the data protection team was clear that health data is special category data under Article 9, so a DPIA was mandatory and the technical controls had to be specific, not a promise in a vendor contract.
The team also wanted a rule that could not be bypassed by a future feature. In a previous incident, a new analytics integration had sent identifiers to a non-certified service for three days before anyone noticed. The DPO wanted the next mistake to be impossible, or at least refused at the boundary and logged.
- 01Integrate
The team wrote a custom Rego policy, reviewed by the DPO in the visual editor, with two branches. Prompts in which the PII guardrail detects identified health data, including the French NIR, patient names and dates of birth, may only be served by the self-hosted vLLM cluster on the platform's HDS-certified hosting in France. Prompts that have been fully pseudonymized may be served by a hosted model in an EU region, with the re-identification vault kept on the HDS hosting.
- 02Enforce
In practice most summary drafts use the self-hosted path, and referral letter drafting, which benefits from a larger model, uses the pseudonymized path. The front end re-identifies the response locally before the clinician sees it, so the hosted model never receives a name, NIR or address.
- 03Evidence
The DPIA referenced the policy file by version and attached the receipt format. When the HDS surveillance audit came round, the platform exported ninety days of receipts and showed the auditor how to verify a random sample against the published Merkle roots.
In their words
“We wrote the rule once: identified health data does not leave our HDS hosting. The engine enforces it on every request, and when a developer gets it wrong the call is refused and logged instead of becoming an incident.”
- vLLM on HDS-certified hosting (France)
- Hosted model in an EU region for pseudonymized prompts
- Custom Rego policy
- PII guardrails with NIR detection
- Local pseudonymization vault
Sovereign route ready
Stop risking 4% global turnover GDPR fines on AI inference.
Deploy sovereign, cryptographically proven AI routing in 15 minutes.