Business questions were waiting in the BI queue
Premier's warehouse already brought operational and accounting data together. Getting an answer still meant asking the data team to write SQL, build a dashboard or stitch spreadsheets together. Leaders could wait about a week for the answer they needed, while ad hoc requests interrupted the team's strategic work.
- New questions depended on analyst time and the dashboard queue
- Follow-up questions created more one-off reporting work
- Spreadsheet assembly added friction between a question and a decision
- BI staff had to switch between strategic projects and ad hoc requests
A trusted path to warehouse answers
The goal was to give teams direct access to warehouse answers in plain English, grounded in Premier's reviewed business definitions, with scoped data access and evidence they could inspect.
Trustworthy answers started with semantic modeling
Premier had already unified data from across its systems in a Microsoft Azure warehouse. Connecting an agent to that warehouse gave it access to the data. To retrieve the right information and answer reliably, the agent also needed a shared model of Premier's business concepts: what its metrics meant, how records related across systems, and which data belonged in each calculation.
As the engagement developed, we reviewed real question history and tested repeated questions. The answer interface needed more than saved SQL examples: it needed explicit rules for sources, formulas, dates and populations. “Revenue” could mean different things in the CRM and accounting system. “GM” needed an agreed revenue and cost basis as well as an acronym definition.
Faster answers. More time for strategic work.
The value was a shorter path from a business question to a usable answer. Premier reduced dashboard build time and gave teams a way to explore warehouse data without putting every question into the BI queue.
1. Dashboard build time reduced by 98.4%
Producing dashboards took substantially less time in the Premier engagement.
2. Answers moved from about a week to seconds
Teams could get warehouse answers grounded in approved business definitions, with the sources and query available for inspection.
3. Business users could explore without waiting on BI
Teams could ask follow-up questions without requesting another one-off report.
4. BI had more room for strategic initiatives
Less ad hoc answer production meant fewer interruptions to the data team's broader work.
We engineered the definitions and the answer path
Fabric's semantic modeling process mapped Premier's business concepts to the warehouse's sources, relationships and calculation rules. We captured that model in a Context Layer, giving the agent approved building blocks for a question: sources, measures, dimensions, filters and business guidance. Each definition can be inspected, reviewed and versioned.
Model business concepts
Declare the source, calculation, counting grain and date rules behind each measure.
Compile and query
Compose approved definitions into SQL and execute through the customer's scoped connection.
Keep the evidence
Connect the answer to its definitions, query and returned data, with charts and export.
Premier's data team kept authority over the meaning
We worked with Premier's data engineer to inspect the layer, challenge it against warehouse and BI knowledge, and stage reviewable changes. Proposed definitions passed through customer review. That collaboration made the business rules explicit and gave the team a way to maintain them.
Inspect together
Read the definition and compare it with the customer's source evidence and business knowledge.
Review the change
Show what is being proposed, why it is needed, and which questions it affects.
Test after publishing
Rerun approved question families to check for changes in meaning, results and answer behavior.
Data engineer partnership
Fabric MCP made the governed layer available through Claude Cowork for operator review, alongside the hosted answer interface. Shared knowledge preserved the decisions and evidence behind the work.
A business question, followed all the way to its sources
“How much backlog revenue do we have by market?” needs an agreed backlog population, a market mapping and a revenue measure. It also needs a view of priced and unpriced jobs, so a revenue total does not hide missing information. The example below illustrates that answer path.
Validation behind the example
In a September 1, 2026 checkpoint, three hosted runs of the backlog question selected the same approved definitions and returned the same full result. The check covered that question and warehouse snapshot, including the backlog population, market grouping, revenue and pricing status.
The review followed the answer through its selected definitions, executed query and returned rows. This made both the business meaning and the result available for inspection.
Ongoing ownership
A maintained capability starts with clear responsibilities
The engagement established an operating approach for customer review and regression. Maintaining that capability means keeping accepted questions, business definitions, source freshness and operator responsibilities explicit as the warehouse and the business change.
- Keep an agreed set of business questions and their acceptance criteria
- Give business definitions and warehouse refresh a named customer owner
- Review proposed changes with their evidence and version history
- Rerun the question suite after changes to the data or the answer path
