Natural Language Analytics for NYU Langone Health

Industry: Healthcare

Service: Databricks Genie Implementation

Timeline: 8 weeks

Natural Language Analytics for NYU Langone Health

Team: Tempered AI, in partnership with Computomic

March 24, 2026

Executive Summary

NYU Langone Health is one of the largest academic medical centers in the United States, managing complex financial and operational data across multiple campuses and service lines. Leadership needed instant, trusted answers from that data without depending on analysts for every question. Tempered AI delivered a production-ready Databricks Genie Space in 8 weeks, giving finance and operations executives a natural language interface to query their data directly, with 95% accuracy and full HIPAA-compliant governance built in.

About the Client

NYU Langone Health is a world-class academic medical center and health system based in New York. It operates across multiple campuses and manages a broad range of inpatient and outpatient services, from surgical care to emergency medicine. Like most large health systems, NYU Langone generates enormous volumes of operational and financial data across diverse business units, and leadership depends on that data to make decisions about performance, resources, and strategy.

The Challenge

The primary bottleneck for business leadership at NYU Langone was not the availability of data, but the overreliance on technical teams to interpret it. Even straightforward questions like "What was the contribution margin by service line last quarter?" required analyst involvement, creating delays that slowed decision-making across the organization.

Data was available but not accessible

NYU Langone had robust data infrastructure, but business leaders could not query it directly. Every routine question required a ticket, a wait, and a handoff to a technical team. The bottleneck was not data, it was access.

Healthcare data is complex and regulated

Financial and operational reporting in a health system is not straightforward. Inpatient and outpatient care follow different metrics, units, and reporting standards. Contribution margin, case mix index, OR hours, and length of stay all have precise definitions that vary by context. Any analytics tool had to reflect those rules exactly, not approximate them.

A wrong number is worse than no number

In healthcare finance, an inaccurate answer that looks correct is more dangerous than no answer at all. The system had to be configured to handle ambiguous questions gracefully, communicate its limits clearly, and never guess when certainty was not possible.

What We Built

A production-ready Databricks Genie Space that lets finance and operations leaders ask questions in plain English and get accurate, governed answers in seconds. A user can ask "Show me Contribution Margin Ratio by service line for inpatient cases in FY25, broken down by campus" and Genie translates that into SQL, runs it, and returns a table or visualization, without involving an analyst.

The 8-week project ran in two phases. The first focused on discovery and foundation building, including use case definition, data architecture design, and semantic layer construction. The second phase built and productionalized the Genie Space through iterative validation, structured user acceptance testing, and deployment to production.

Key Features:

Natural Language Querying

Executives type questions in plain English and get results in seconds. Genie translates the question into optimized SQL and returns tables or visualizations without requiring any technical knowledge from the user.

Precisely Defined Semantic Layer

Metric definitions for contribution margin, case mix index, OR hours, length of stay, payor group, service line, and more are encoded directly into the system, with plain English descriptions and synonyms. The model does not guess how metrics are defined, it knows.

Inpatient and Outpatient Coverage

The health system tracks inpatient and outpatient care differently. The system handles both in a single interface with the correct logic for each, including surgical cases, ED admissions, observation stays, and ambulatory surgery encounters.

Guardrails and Behavioral Instructions

Genie is configured to handle ambiguous questions, apply time periods correctly, and communicate clearly when it cannot answer rather than return an uncertain result.

HIPAA-Compliant Security

Row-level and column-level security ensure sensitive patient data stays protected. Users access only what their role permits, with all access logged for compliance auditing.

Query History and Audit Trail

All queries are logged with user attribution, timestamps, and results, meeting healthcare compliance requirements while enabling teams to learn from each other's analytical approaches.

The Result

Production-ready Genie Space live in 8 weeks

95% accuracy across question types, validated through structured user acceptance testing

Finance and operations leaders get answers to complex questions in seconds, without analyst involvement

Full coverage across inpatient and outpatient operations including volume trends, financial performance, OR metrics, length of stay, and case mix

HIPAA-compliant with row-level and column-level security and full audit trail

Internal team fully equipped to maintain and evolve the system independently

Why It Works

Delivering a conversational AI product that performs reliably in healthcare and finance requires more than a good AI model. It requires a clean, well-governed data foundation, a precisely defined semantic layer, and careful configuration of how the model behaves. That is where the most critical work took place. The Genie interface is what users see, but the accuracy they experience is a result of the engineering beneath it. Close collaboration with NYU Langone's finance and operations teams was what made production-grade accuracy possible.

Ready to give your leadership team direct access to your data?

We can implement a governed, natural language analytics platform calibrated to your business metrics and your environment.