AI enablement: Claude, Power BI, and Microsoft 365


Husco International builds hydraulic and electro-mechanical components for the automotive and off-highway markets, and has done it for more than eighty years. The company is privately owned and runs on Microsoft. Power BI reports and dashboards run the business day to day.
What Husco was trying to solve
Husco wanted users to ask a question in plain language and get a decision-ready answer from the company's own PowerBI data, with their Entra ID permissions deciding which data each person sees. As a Microsoft shop they started with Copilot, which gave plug-and-play access to Power BI and Microsoft 365 and returned answers too thin for the questions people were actually asking. They moved to Claude with MCP connectors to test whether the same data access would produce better analysis, found large gaps in the off-the-shelf Power BI connector, and built a custom MCP internally. That showed real progress and stalled on two things: identity and session persistence, and a lack of report and business level context that would enable Claude to understand the underlying semantic models.

What Red Foundry built
We rebuilt session management to remove friction and confusing UX and disconnect/reconnect behavior that would have limited or prevented regular use of Claude to answer PowerBI related questions. We extended the connector to consume Power BI report definitions and additional business context alongside semantic model information, so an answers match the intel that could be derived from a direct read of a PowerBI report. Row limits and concurrency caps were also introduced to mitigate potential PowerBI overloads that the native connector had allowed to happen in the past.
One answer, two systems
On top of the connector we built a skill layer that tells Claude how to use the connectors together, so one answer can draw on Power BI data and the Microsoft 365 content that explains it. The skill classifies each question and holds a fixed answer shape. A companion glossary carries Husco's own definitions, so the terms in an answer mean what they mean in the business, and each rule traces to a live query against the estate.
A benchmark written before the build
No definition of a correct answer existed when we started, so we built one. An evaluation suite of roughly 100 test cases, drawn from Husco's business requirements, benchmarks both correctness and richness. We ran it at multiple phases of connector and skill development, and a change to the skill still gets tested against that set before anyone ships it. Husco keeps editing the skill after we leave, and the suite is what makes that safe.
One identity, all the way through
The server carries the signed-in user's own identity to Power BI and Microsoft 365, using delegated on-behalf-of authentication against Entra ID with JWT validation and no stored secrets. Claude sees what that person could already open natively. We tested it at the boundary: two users with different entitlements ask the same question, and each gets back only their own data. Across the full test set we recorded zero cross-user leaks, which was the condition for going live.
Governed before it widens
Husco uses a Team license with admin control, and every user runs one published version of the skill through Claude organization settings. Per call telemetry captures the user, tool, model, duration, rows, and outcome alongside token cost, so Husco can forecast spend before adding use cases.

Results and Outcomes
- Semantic model reach: the off-the-shelf connector lacked specific business required behavior. Husco's V1 custom behavior closed that gap, but missed required context and identity and session management needed for the connector to be successful.
- Report context: answers were correct against the model and mismatched to the report onscreen. The connector now reads report definitions alongside the semantic model.
- Answer quality: there was no definition of a correct answer. A ~100 test case suite drawn from business requirements now benchmarks correctness and richness, and gates every change to the skill.
- Permissions: untested. Zero cross-user leaks across the full boundary test set.
- Cost visibility: none. Per call telemetry on user, tool, model, duration, rows, outcome, and token cost.
- Getting an answer: an analyst opened Power BI and built it. Anyone now asks in plain language and gets a sourced answer.

Ready to get started?
We see our clients as our partners and are invested in their success. We ensure what we build for our clients will have a positive ROI. We never build anything just for the sake of it.




