Automating Healthcare Documentation and Rule Management with GenAI

orizon · Healthcare · 2024 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Healthcare billing-rule documentation

Orizon needed to document and interpret tens of thousands of legacy healthcare billing rules without tying up scarce developers for days per rule update.

Constraints:

rules are proprietary and healthcare-sensitive
business rules are coded in legacy C# and C++
new-rule documentation previously required multiple developers
business users needed explanations without waiting for developers
models had to run under strong governance and access controls

Architecture

Orizon adopted the Databricks Data Intelligence Platform and a lakehouse architecture to consolidate data, govern proprietary rules, fine-tune Llama2-code and DBRX on company data and business rules, and deploy them through Databricks Model Serving. The initial user surface was a Microsoft Teams chatbot that lets business users ask questions about rules and receive text and graphical documentation without involving developers.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  legacy-rule-corpus[Legacy rule corpus] --> databricks-lakehouse[Databricks lakehouse]
  databricks-lakehouse --> mlflow-lifecycle-manager[MLflow lifecycle manager]
  databricks-lakehouse --> unity-catalog-governance[Unity Catalog governance]
  mlflow-lifecycle-manager --> fine-tuned-rule-models[Fine-tuned rule models]
  unity-catalog-governance --> model-serving-endpoint[Model serving endpoint]
  fine-tuned-rule-models --> model-serving-endpoint
  model-serving-endpoint --> microsoft-teams-chatbot[Microsoft Teams chatbot]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Legacy rule corpus`	Contains medical billing rules implemented in C# and C++.	`Databricks lakehouse`
`Databricks lakehouse`	Consolidates isolated databases and unifies data, analytics, and AI.	`Legacy rule corpus`, `MLflow lifecycle manager`, `Unity Catalog governance`
`MLflow lifecycle manager`	Tracks ingestion, feature engineering, model building, tuning, and experiments.	`Databricks lakehouse`, `Fine-tuned rule models`
`Unity Catalog governance`	Controls access and permissions for proprietary business-rule data.	`Databricks lakehouse`, `Model serving endpoint`
`Fine-tuned rule models`	Llama2-code and DBRX models customized on Orizon data and business rules.	`MLflow lifecycle manager`, `Model serving endpoint`
`Model serving endpoint`	Serves the fine-tuned models securely across the organization.	`Fine-tuned rule models`, `Microsoft Teams chatbot`
`Microsoft Teams chatbot`	Lets business users query rule explanations and documentation in natural language.	`Model serving endpoint`

Data flow

Legacy business rules and healthcare data are consolidated into Databricks, governed through Unity Catalog, used to fine-tune Llama2-code and DBRX, managed with MLflow, and served through Databricks Model Serving. Microsoft Teams exposes the models to business users for rule explanations and documentation.

4 supporting quotes

“Orizon adopted the Databricks Data Intelligence Platform to more easily integrate all necessary components required for comprehensive data warehousing”article, stated

“embedding GenAI, along with its natural language processing capabilities, into Microsoft Teams so business users could get answers to questions”article, stated

“Once the Orizon team fine-tuned the Llama2-code and DBRX GenAI models with their own data and business rules, they securely deployed the models through [Databricks Model Serving](https://www.”article, stated

“securely deployed the models through Databricks Model Serving”article, stated

Why this architecture

Orizon needed custom models because its billing rules were proprietary and could not be sent outside the company. Databricks supplied both the AI lifecycle tooling and governance substrate: MLflow for reproducibility and model lifecycle, Unity Catalog for fine-grained permissions, and Model Serving for organization-wide access. Microsoft Teams gave business users a familiar interface.

3 supporting quotes

“Guisse expanded, “We couldn’t send this data outside the company because our business rules are confidential. To develop our LLM securely within Orizon, we needed a partner like Databricks to empower our users to create custom models based on our own data.”article, stated

“Using MLflow, an open source MLOps platform developed by Databricks, Orizon could seamlessly manage the entire machine learning lifecycle — including data ingestion, feature engineering, model building, tuning and experiment tracking — to ensure reproducibility.”article, stated

“Thanks to Unity Catalog’s ability to define and enforce granular permissions, Orizon could maintain the necessary security and compliance, ensuring that sensitive data remained protected yet accessible to authorized personnel for critical operations.”article, stated

Models used: Llama2-code · DBRX

Scale

Orizon manages about 40,000 medical rules.
“Guilherme Guisse, Head of Data and Analytics at Orizon, explained, “We have around 40,000 medical rules in our system and add around 1,500 new rules each month.”article, stated
Orizon adds around 1,500 new rules each month.
“Guilherme Guisse, Head of Data and Analytics at Orizon, explained, “We have around 40,000 medical rules in our system and add around 1,500 new rules each month.”article, stated
The new workflow processes 63% of tasks automatically.
“According to Guisse, “With our brand-new GenAI features, we are currently processing 63% of tasks automatically.”article, stated
Documentation time fell to less than five minutes.
“Even more impressive, the documentation process takes less than five minutes, dramatically accelerating workflows, with the potential to increase the rate of new rules from 1,500 to 40,000 per month — translating to 1 billion Brazilian reals (BRL) in added productivity.”article, stated
Resource savings were about $30K per month.
“This shift has also translated to significant cost savings for Orizon — approximately $30K per month in better-used resources.”article, stated

Alternatives considered (roads not taken)

Manual developer documentation

Rejected because: New-rule documentation took several days and required developers to interpret legacy code.

“Since adding a new rule took several days to complete, business analysts often had to request C++ developers to interpret code”article, stated

External data/model processing

Rejected because: Business rules were confidential and needed to stay inside Orizon's governed environment.

Fragmented databases and legacy workflows

Rejected because: Fragmentation made fraud-detection and documentation workflows inefficient.

“the company was burdened by their internal data management, which hindered their fraud detection capabilities.”article, stated

Patterns

`fine-tuning-adaptation` (modeling)

Orizon fine-tuned code-oriented and general GenAI models on proprietary healthcare rule data.

“fine-tuned the Llama2-code and DBRX GenAI models with their own data and business rules”article, stated

`privacy-preserving-processing` (governance)

Confidential business-rule data stayed inside a governed Databricks environment with Unity Catalog permissions.

“We couldn’t send this data outside the company because our business rules are confidential.”article, stated

`observability-tracing` (operations)

MLflow manages the machine-learning lifecycle including experiment tracking and reproducibility.

“could seamlessly manage the entire machine learning lifecycle”article, stated

`tool-use-orchestration` (workflow)

The user-facing workflow embeds rule-querying GenAI into Microsoft Teams for business users.

“embedding GenAI, along with its natural language processing capabilities, into Microsoft Teams”article, stated

Tradeoffs

Fine-tune custom models inside Databricks

Chose Llama2-code/DBRX fine-tuned on Orizon data over sending proprietary rules to external services.

Rationale: Custom governed models protected confidential rules while supporting domain-specific explanations.

“We couldn’t send this data outside the company because our business rules are confidential.”article, stated

“create custom models based on our own data.”article, stated

Expose GenAI through Microsoft Teams

Chose business-user chatbot interface over developer-mediated rule explanations.

Rationale: The Teams interface let business users ask rule questions without waiting on developers.

“business users could get answers to questions without relying on developers.”article, stated

Gotchas (what bit them)

Confidential rule data blocked naive external-model use.

Root cause: Orizon's medical billing rules are proprietary and healthcare-sensitive.

Resolution: The company used Databricks governance and custom models trained on internal data.

“our business rules are confidential. To develop our LLM securely within Orizon, we needed a partner like Databricks”article, stated

Legacy medical rules created a developer bottleneck.

Root cause: Rules were written in C# and C++ and needed manual interpretation and flowchart documentation.

Resolution: GenAI produced text and graphical documentation so developers could focus on development.

“These rules — coded in legacy languages like C# and C++ — were necessary to determine a member’s healthcare eligibility”article, stated

“our developers can focus on development, not just documentation.”article, stated

Evaluation

Orizon measured automation rate, documentation time, developer capacity freed, cost savings, and potential rule throughput increase.

Metrics: tasks automated · documentation time · developer capacity freed · monthly resource savings · potential rules per month