How We Optimized Microsoft Fabric Data Agents to Reduce Compute Unit (CU) Consumption: A Real-World Healthcare Implementation

Artificial Intelligence Is Only as Smart as the Data Model Behind It

Microsoft Fabric Data Agents are changing how organizations interact with enterprise data. Instead of relying on dashboards or writing SQL queries, business users can simply ask questions in natural language and receive meaningful answers powered by AI.

Imagine a hospital administrator asking:

• How many patients were readmitted last month?

• Which department has the highest average length of stay?

• Show admissions by physician.

  
  

The Data Agent translates these questions into DAX queries, retrieves the data from the semantic model, and returns the answer within seconds.

While this experience feels seamless to the end user, there is significant processing happening behind the scenes. Every question requires the AI to interpret business intent, understand the semantic model, generate DAX, and execute the query. The more ambiguity the AI encounters, the more reasoning it performs - and that reasoning consumes Microsoft Fabric Compute Units (CUs).

We recently implemented Microsoft Fabric Data Agents for a healthcare organization. The initial solution worked well from a functional perspective, but every question was consuming approximately 5,000 Fabric Compute Units (CUs). As the number of users increased, it became clear that improving efficiency would be just as important as improving accuracy.

Rather than increasing Fabric capacity, we focused on optimizing the Data Agent itself.

By applying Microsoft's Semantic Model best practices together with implementation-specific improvements, we achieved faster, more consistent responses while significantly reducing unnecessary AI reasoning.

This article shares our approach.

The Problem Was Not the AI

One common misconception is that poor AI performance is caused by the language model. In our experience, that was not the case.

The AI was performing exactly as expected. The challenge was that it had too many decisions to make before it could answer a question.

Consider a simple request:

"Show patient volume by department."

To answer this question, the Data Agent must determine:

• Which table contains patient admissions?

• Does "department" refer to the clinical department, billing department, or specialty?

• Which measure represents patient volume?

• Which relationships should be used?

• What DAX query should be generated?

Every one of these decisions requires additional reasoning, increasing both execution time and Compute Unit consumption.

Instead of trying to optimize the AI model, we optimized the information available to the AI.

Building the Right Foundation(Data Warehouse) Before Enabling Data Agents

Before implementing Microsoft Fabric Data Agents, we invested significant effort in understanding the client's business processes.

Rather than exposing raw operational data, we designed a structured data warehouse consisting of:

• Fact tables for admissions, encounters, and clinical activities

• Dimension tables for physicians, departments, patients, and dates

• Standardized business measures

• Well-defined relationships

  
  

Only after completing the semantic layer did we enable the Data Agent.

This proved to be one of the most important decisions in the project.

A Data Agent is only as effective as the semantic model it queries.

Applying Microsoft's Semantic Model Best Practices

Microsoft recommends preparing semantic models specifically for AI-powered experiences rather than simply exposing existing Power BI datasets. This includes organizing business-friendly metadata, defining AI Instructions, creating Verified Answers, improving naming conventions, and validating semantic models before deployment.

We adopted these recommendations and extended them with additional implementation practices learned during the project.

1. Simplifying the Semantic Model

Many organizations expose entire data warehouses to AI.

While this provides flexibility, it also introduces unnecessary complexity.

If multiple tables contain similar business concepts, the AI spends additional time determining which objects should be used.

We simplified the semantic model by:

• Removing duplicate business concepts

• Hiding technical columns

• Exposing only business-relevant measures

• Simplifying relationships

• Using meaningful names instead of technical database terminology

For example, instead of exposing internal table names such as:

• PATIENT_MASTER

• ADMISSION_HDR

• FACT_ENCOUNTER

we exposed business-friendly objects such as:

• Patients

• Admissions

• Encounters

Reducing ambiguity meant the Data Agent had fewer possible paths to evaluate before generating DAX.

2. Using Verified Answers to Eliminate Guesswork

One of the most valuable capabilities in Microsoft Fabric Data Agents is Verified Answers. Without Verified Answers, every user question is interpreted from scratch.

Consider the question:

"Show performance by territory."

Does "territory" refer to sales territory?

Product territory?

Distribution region?

The AI must determine the user's intent.

Microsoft's Verified Answers allow organizations to define trusted responses and associate multiple trigger questions with the same business concept.

We applied the same approach in healthcare.

Instead of defining only one question, we created several variations for the same business requirement.

For example:

• Show admissions by department

• Department-wise admissions

• Which department admitted the most patients?

• Patient volume by department

Although users phrase questions differently, they all expect the same business answer.

By mapping these questions to Verified Answers, the Data Agent immediately recognized user intent instead of reasoning through multiple possibilities.

The benefits were immediate:

• More consistent responses

• Faster query execution

• Improved user confidence

• Reduced AI reasoning

3. Teaching Business Language Through AI Instructions

Every organization has its own terminology.

Healthcare is especially rich in business-specific definitions.

For example:

What exactly is a "High Risk Patient"?

Different hospitals define this differently.

Similarly:

• Readmission

• Occupancy Rate

• Active Patient

• Average Length of Stay

• Critical Care

may all have organization-specific meanings.

Without additional guidance, AI attempts to infer these definitions.

Instead, Microsoft Fabric allows organizations to define AI Instructions.

These instructions become part of the context provided to the Data Agent whenever it generates DAX. Rather than leaving the AI to guess, we explicitly defined business rules.

For example:

• A readmission means a patient admitted again within 30 days.

• Average Length of Stay should always be calculated using discharge date minus admission date.

• Occupancy Rate is based on occupied beds divided by available beds.

Once these instructions were defined, the Data Agent consistently interpreted business terminology exactly as the client expected.

This significantly reduced unnecessary clarification and reasoning.

4. Reducing Ambiguity Through Better Naming

Technical database names make sense to developers.

Business users think differently.

If users ask:

"Show admissions by physician"

but the semantic model contains a table named

PHYS_MASTER

the AI must infer that both refer to the same concept.

Instead, we renamed objects using business terminology familiar to users.

Examples included:

• Physician

• Department

• Patient

• Admission

• Diagnosis

• Length of Stay

Business-friendly naming improves both usability and AI interpretation.

Why These Changes Reduced Compute Unit Consumption

The most important lesson from this implementation was simple.

Large Language Models consume more compute when they must reason.

Reasoning increases whenever the AI needs to:

• interpret business terminology

• identify the correct measures

• resolve ambiguity

• choose among multiple possible tables

• infer business definitions

By reducing ambiguity before the question reaches the AI, we reduced the amount of reasoning required.

Instead of asking:

"What does this user mean?" the Data Agent already knew.

Although Compute Unit consumption depends on several factors-including semantic model complexity, query execution, and dataset size-we observed that improving the semantic layer and reducing ambiguity contributed to lower overall compute usage while improving response quality.

Our focus was not simply reducing cost.

It was improving efficiency.

Business Outcomes

After implementing these optimizations, the healthcare client experienced several important benefits.

Improved Accuracy

Business terminology was consistently interpreted across different users and departments.

Faster Responses

The Data Agent spent less time resolving ambiguity before generating DAX.

Greater User Adoption

Business users gained confidence because answers remained consistent regardless of how questions were phrased.

More Efficient Compute Utilization

By improving the semantic model instead of increasing Fabric capacity, the organization made better use of its existing Microsoft Fabric resources.

Best Practices for Optimizing Microsoft Fabric Data Agents

Based on our implementation experience, we recommend the following approach for organizations planning to deploy Microsoft Fabric Data Agents:

• Design a business-focused semantic model before enabling AI.

• Remove unnecessary tables, columns, and duplicate concepts.

• Use meaningful business-friendly object names.

• Define AI Instructions for organization-specific terminology.

• Create Verified Answers for frequently asked business questions.

• Validate semantic models with business users before deployment.

• Continuously monitor user questions and refine Verified Answers over time.

Treat the Data Agent as an extension of your business knowledge-not simply another AI feature.

Final Thoughts

Microsoft Fabric Data Agents represent a major step forward in self-service analytics. However, the quality of the experience depends far less on the AI model itself than on the quality of the semantic model supporting it.

Our healthcare implementation reinforced a valuable lesson: optimizing AI starts with optimizing data.

By investing time in understanding the client's business processes, building a well-structured semantic model, defining AI Instructions, and creating Verified Answers, we created a Data Agent that delivered more accurate responses while making more efficient use of Fabric Compute Units.

References

Microsoft Learn. Semantic model best practices for Microsoft Fabric Data Agents. Available at: https://learn.microsoft.com/en-us/fabric/data-science/semantic-model-best-practices

Aptocoiner Analytics helps organizations implement Microsoft Fabric, Data Warehouses, Data Engineering, Power BI, and AI-powered analytics solutions. Our approach combines modern data architecture with practical business understanding to deliver scalable, cost-effective analytics platforms that enable confident, data-driven decision-making.