Modernizing healthcare referrals with agentic AI depends less on any single model and more on the data foundation beneath it. Health systems stand to benefit from AI across patient outcomes, physician workload, and revenue-cycle efficiency, yet many struggle to realize that value because their data and AI estates are highly fragmented.
The core problem: fragmented healthcare data
Patient records sit in electronic health record (EHR) systems, medical images live in PACS, bedside monitors stream telemetry, and clinical notes remain unstructured. These silos lead to inconsistent security policies, operational and clinical data that is effectively locked in, and AI initiatives that sprawl without shared enterprise semantics. Building and operating HIPAA-compliant AI agents adds further difficulty, particularly when balancing computational cost against clinical quality and safety.
Maximizing AI potential starts with the data
A common architectural response is to consolidate hospital data on a lakehouse foundation. In the Databricks ecosystem this combines large-scale storage with Lakeflow for ingestion and streaming, which suits real-time HL7/FHIR feeds and IoT bedside monitors, and Lakebase, a serverless Postgres option. An integrated foundation is positioned to lower total cost of ownership and reduce the vendor lock-in common in legacy healthcare IT, though actual savings depend on each organization’s existing systems and migration scope.
Governance is the second pillar. Unity Catalog provides unified controls across assets, which matters for protected health information (PHI) and HIPAA compliance. Its scope extends beyond traditional tables to predictive models, AI agents, PDFs such as scanned referral documents, and raw files.
Optimizing referral management
Once a governed data foundation is in place, health systems can supply their own clinical and operational context to AI rather than relying on generic models. For referrals specifically, that context can support matching patients to appropriate specialists based on factors such as availability and proximity, and surfacing the information staff need to route referrals efficiently.
Preparing data for these uses still requires modeling and transformation. Fivetran offers transformation tooling that includes prebuilt quickstart models for common SaaS platforms and integration with dbt and Coalesce, alongside newer capabilities such as an AI Column Builder that enriches data with additional fields using an external AI provider, and an audience-builder feature for custom segments. The transformed data can then be activated into the applications where teams already work, so insights are not trapped in dashboards, static reports, or warehouses.
Natural-language analytics and agents
Interfaces such as Databricks AI/BI Genie let hospital administrators, nursing managers, and medical directors query data in natural language, exploring questions about patient flow, readmission rates, or emergency-department wait times and receiving tables and visualizations in response.
For document-heavy work, production agents can extract structured data from unstructured records. As a verified example, AstraZeneca used Databricks Agent Bricks to parse more than 400,000 complex clinical trial documents and extract structured data points without writing custom code, reportedly completing the task in under an hour. A similar pattern could let a hospital extract structured points from decades of historical unstructured records. By accelerating ingestion with Fivetran and layering analytics and AI on Databricks, health systems can move from fragmented IT toward a more connected delivery network.
Limitations and what to watch
The architecture described here reflects vendor product positioning, and outcomes vary by organization. Claims about lowest total cost of ownership and reduced lock-in are vendor assertions that depend on existing infrastructure, contracts, and migration effort. Deploying AI in clinical settings raises real obligations around HIPAA, PHI handling, model validation, bias, and clinical safety that technology alone does not resolve. The AstraZeneca example demonstrates document extraction at scale, not autonomous clinical decision-making, and any referral or care workflow should keep qualified humans in the loop. Organizations evaluating these tools should pilot carefully, measure accuracy on their own data, and confirm compliance before production use.
Related coverage on this site includes how Databricks applies LLMs to PII detection and governance. Further detail on the agent tooling referenced above is available on the official Agent Bricks page.
