Challenge
Financial figures within reach, behind a language the business does not speak.
The House's finance teams keep their figures in an on-premises OLAP cube running on Analysis Services. Querying it requires writing MDX queries, a language business users do not practice.
Every question therefore went through a technical intermediary. Finance expected three things.
- Ask questions in natural language. Without going through a query or an expert.
- Get readable answers. Tables or charts, ready to use.
- Compute on the result. Apply everyday financial calculations, such as growth or shares.
The need: make financial data accessible to those who steer it, without moving the cube out of the data center.
Solution
An agentic workflow that turns the business question into a query, then into an answer.
We built an assistant that speaks the language of finance. Three principles shape the solution.
- From question to query. An agentic workflow on LangGraph translates the question into MDX, runs it on the cube and applies growth and share calculations.
- Business rules at the core of the model. We encode management rules, definitions and the cube structure in the prompts, based on workshops with the finance team.
- Data stays on premises. A custom connector links the cloud application to the data center, without moving the cube.
Set-up
Understand finance, encode its rules, connect the cube, deliver the application.
- Business workshops. Collection of management rules, definitions and the cube structure with the finance team.
- Agentic workflow. Translation of questions into MDX, execution on the cube and financial calculations on the result, orchestrated with LangGraph.
- On-premises connector. A custom connector, hosted on an Azure virtual machine, between the cloud application and the data center.
- Web application. Python back end and React front end deployed on Azure.
- Continuous improvement. User feedback and usage tracking built into the application.
Tech stack: LangGraph, Azure, Analysis Services, MDX, Python, React.
Results
In production since January 2026, used every day by the CFO.
Finance teams ask their questions in natural language and get tables and charts, growth and share calculations included. The cube stays in the House's data center.
The assistant has been in production since January 2026 and the CFO uses it every day.
Two effects extend the project.
- A finance function autonomous on its data. Business users access their figures directly, without a technical intermediary.
- An assistant that improves with use. User feedback and usage tracking guide its evolution.
Expertises mobilisées
Savoir-faire
- AI agents
- Talk-to-your-data
- Context engineering
- Text-to-MDX
- Finance
Technologies
- LangGraph
- Azure
- Analysis Services
- Python
- React
Signature
Four mechanisms that make the difference.
- An assistant that knows finance. We collect management rules and definitions with the teams rather than assume them.
- Reliable calculations on the result. The workflow computes growth and shares; the model never improvises them.
- Data that never leaves the site. The custom connector links the cloud to the cube without moving it.
- Usage as the measure of success. Feedback and usage tracking are built in from go-live.
For you
Your financial data, accessible in natural language.
This approach addresses four situations.
- Your figures live in an OLAP cube or a data warehouse that only experts can query.
- Your business teams expect answers rather than queries.
- Your data must stay on premises, but you want a modern application.
- You want an AI assistant that respects your management rules and definitions.
We collect your business rules with your teams, connect the assistant to your data where it lives and put it into production with usage tracking.
