AI applied to FHIR data has moved past pilots in 2026. Four production applications ship real value; understanding what shipped and what didn't shapes AI investment decisions.
1. Predictive risk scoring on bulk data. Population-level risk models (readmission, sepsis, deterioration) trained on bulk `$export` data feed clinician dashboards. This is now table-stakes in academic medical centers.
2. Ambient documentation with FHIR extraction. Voice-recognition tools capture clinician-patient conversations, extract structured Observation, Condition, and MedicationRequest resources. Nuance DAX, Suki, and Abridge lead the vendor space.
3. Prior authorization AI on Da Vinci CRD. ML models predict PA approval likelihood at order-select time via CDS Hooks with Da Vinci CRD payload. Reduces submission-review-approval friction.
4. LLM-assisted chart summarization. Long chart histories summarized into concise clinical narratives via LLMs. GPT-4 class models, prompted with US Core-shaped FHIR data.
What hasn't shipped
1. Autonomous diagnosis. Regulatory (FDA SaMD) and liability concerns keep this out of production. 2. Fully generative treatment plans. LLMs still hallucinate on clinical guidelines; used with heavy clinician review. 3. Continuous learning models. Model drift + regulatory concerns mean models are periodically retrained, not continuously updated.
Integration patterns
| Application | FHIR touch point | Delivery pattern |
|---|---|---|
| Risk scoring | Bulk $export |
Nightly batch |
| Ambient documentation | Subscription/API | Real-time, in-workflow |
| PA prediction | CDS Hooks | Real-time, in-workflow |
| Chart summarization | Patient/Encounter reads | Real-time, on-demand |
Data quality gates
AI on FHIR is only as good as underlying data. Track `$validate` pass rates, reference integrity, and terminology conformance. Sites <95% on these metrics don't get good model performance.
AI + FHIR is now production reality for the four use cases above. The pattern is consistent: FHIR provides the data, purpose-built models provide the intelligence, clinicians retain decision authority.
