AI for healthcare
Healthcare teams spend a growing share of their time on documentation. AI can give clinicians hours back every week, provided it is deployed with the safeguards patient data demands.
The biggest near-term value of AI in healthcare is administrative rather than diagnostic. Discharge summaries, referral letters, coding, scheduling correspondence and protocol lookups consume a large share of clinical time, and these are recoverable-error, human-reviewed workflows where language models perform well today. Organisations that start there build the trust, governance and data foundations that any future clinical application will require, under the strictest data-protection regime there is.
Patient data is special-category data
Health data enjoys reinforced GDPR protection, and in France its hosting is itself regulated: providers must use HDS-certified hosting for personal health data. Architecture and vendor choices are compliance decisions before they are technical ones.
Clinical safety boundaries
Where AI becomes a medical device or a safety component of one, it falls under medical-device regulation and the EU AI Act’s high-risk regime. Drawing the line between assistive drafting and clinical decision-making, and keeping deployments on the right side of it, is a core design task.
Clinician trust and adoption
Clinicians will not accept opaque suggestions, and they are right not to. Tools earn adoption when every output cites its source, fits the existing workflow, and visibly saves time on work nobody wanted to do.
Fragmented information systems
Hospital information systems and EHRs are heterogeneous and integration-resistant. Successful projects scope tightly around real document flows instead of waiting for a perfect data layer.
Clinical documentation support
First drafts of discharge summaries, referral letters and consultation notes for clinician review. In most institutions this is where the largest time savings sit.
Medical knowledge access
Literature synthesis and internal protocol Q&A with citations, so staff query institutional knowledge instead of hunting through PDFs.
Patient communication
Clear, multilingual drafts explaining procedures, instructions and administrative steps, reviewed before sending and readable by the patient.
Administrative operations
Coding and billing support, correspondence drafting and back-office document processing. Relief for teams under the same staffing pressure as clinical ones.
A healthcare deployment starts from data-protection law and works backwards: a lawful basis under GDPR for any processing of health data, HDS-certified hosting where French personal health data is stored, minimisation and pseudonymisation wherever the workflow allows, and human review on anything that approaches care. Use cases are assessed against the EU AI Act. Administrative assistance generally sits outside the high-risk tier, while anything qualifying as a medical device is engineered to that standard or deliberately left out of scope. The goal is a system your DPO can defend.
Can we process patient data with an LLM at all?
Yes, lawfully and carefully: an appropriate legal basis, a DPIA, suitable hosting (HDS-certified in France for personal health data), minimisation or pseudonymisation where possible, and enterprise agreements that exclude training on your data. Many valuable workflows can also run entirely on de-identified or non-clinical data.
Is healthcare AI automatically high-risk under the EU AI Act?
No. AI that is a medical device or its safety component falls in the high-risk regime; administrative and documentation assistance generally does not. Each use case needs its own classification, which also means you can start delivering value in the lower-risk tiers now.
Where should a hospital or clinic start?
Administrative documentation with mandatory human review: discharge summaries, letters, protocol Q&A. The time savings are significant, the risk is contained, and the governance you establish becomes the foundation for more ambitious projects.