Many organisations want to introduce AI-driven analytics, enabling business users to query data in natural language rather than SQL. However, adopting commercial solutions often requires costly migrations to new platforms and vendor-specific ecosystems.

To explore an alternative approach, we developed a proof of concept of an AI agent capable of querying ERP data warehouses using natural language.

The challenge

Large Language Models can generate SQL queries, but without business context, they often misunderstand relationships between tables, metrics, or business terminology. In enterprise environments, this leads to inaccurate results and low trust in generated insights.

At the same time, many organisations already operate mature relational data warehouses and are not ready to redesign their entire analytical landscape to adopt AI features.

The solution

Our PoC combines a Large Language Model with a semantic description layer explaining the business meaning of ERP data structures.

The AI agent:

• interprets natural language questions,
• generates SQL queries,
• executes queries against the existing data warehouse,
• returns structured analytical results.

Unlike platform-specific semantic solutions, the approach works on top of existing relational database management systems, enabling faster and cheaper adoption without large-scale migration projects.

Key observations

The project confirmed that semantic context is more important than raw schema metadata when generating reliable SQL queries with AI.

We also observed that many organisations already possess the required business knowledge — but it is scattered across documentation, BI layers, and expert knowledge rather than structured for AI consumption.

Business value

The solution demonstrates that organisations can experiment with AI-powered self-service analytics incrementally, leveraging existing infrastructure instead of replacing it.

For companies looking for a pragmatic entry point into AI analytics, this approach may significantly reduce adoption barriers while accelerating experimentation and business validation.

Authors

Artur Łapczyński
Tech Lead
Holisticon Insight
AWS Certified Solutions Architect, ML Specialist and AI Practitioner

Daniel Urbaniak
Senior Java Developer