LLMs in the EHR: Real-World Deployment, Adoption, and What Actually Drives Value
EHRs have improved access to clinical data — but often at the cost of a massive increase in administrative time. In some settings, physicians spend more than 50% of their time on non-clinical tasks. LLMs are frequently framed as the solution, yet most evidence remains theoretical or benchmark-driven. This summary highlights a real-world deployment of an LLM chatbot integrated into Epic EHR in a European university hospital, focusing on feasibility, adoption, and how clinicians actually use it (not on clinical “performance”).
Key takeaways
- LLMs can significantly reduce EHR-related administrative burden, especially through summarisation and information retrieval.
- Real clinician use focuses on summarisation and retrieval — not diagnosis.
- Native EHR integration, with no workflow disruption, is critical for adoption.
- The main barriers are not technical, but governance, security, compliance, and trust.
- On-premise, GDPR-compliant deployments are feasible and accepted in European hospital settings.
Context & why it matters
While EHRs improved access to clinical data, they also increased administrative overhead dramatically. LLMs are often presented as a remedy, but most work remains theoretical or based on synthetic benchmarks. This study reports a real deployment of an LLM chatbot integrated into the Epic EHR in a European university hospital, focusing on feasibility, adoption, and real physician usage rather than clinical decision-making accuracy.
Where AI already works today
In the pilot (28 physicians, 9 specialties, 1 month), usage clustered around:
- Patient chart summarisation (≈ 30%)
- Targeted clinical information retrieval (≈ 29%)
- Support for clinical reasoning (≈ 25%)
- Note drafting (≈ 15%)
Clinicians primarily use AI as a cognitive augmentation tool: quickly aggregating dozens of documents, locating a precise data point, and saving time when preparing for consultations.
Why adoption remains limited
The obstacles identified were mainly organisational:
- Complex data flows (internal documents, regional eHealth sources, literature)
- High requirements for security, privacy, and regulatory compliance (GDPR, MDR, upcoming AI Act)
- Risk of hallucinations or omissions, requiring traceability and explainability
- Low trust in automated diagnostic use cases
Diagnosis remains marginal: physicians do not expect AI to decide — they expect it to structure information and make it usable.
What enables trust and scale
The study suggests adoption depends on:
- On-premise hosting (no patient data leaves the hospital)
- Controlled RAG (internal documents, eHealth sources, locally indexed literature)
- Access control via the EHR (SMART on FHIR)
- Systematic explainability (sources are visible and clickable)
- Strong governance: DPIA, ethics committee oversight, internal validation, model versioning
Result: 64% of participants used the tool daily.
What healthcare leaders should do
- Invest first in low-risk use cases (summarisation, retrieval, documentation)
- Embed AI inside the EHR — not alongside it
- Establish clear governance (security, validation, continuous monitoring)
- Measure real workflow impact before targeting advanced clinical use cases
- Design AI as clinician support, not a replacement
👉 Bottom line
This study shows generative AI in healthcare already creates value today — provided it is restrained, integrated, secure, and well-governed. The future of hospital AI will depend less on new models and more on strong, real-world implementation.