80% Less Time on Data Requests with Self-Learning AI

The Problem

At a large enterprise insurance company, the data analytics team had become an unintentional bottleneck. Every time someone in marketing wanted campaign performance numbers, or operations needed claims data for a quarterly review, they filed a request with the data team. Analysts spent the vast majority of their time writing one-off SQL queries and assembling spreadsheets for other departments instead of doing the analytical work they were hired for.

The problem ran deeper than just volume. Non-technical users in marketing, operations, and executive leadership had no way to pull data themselves. They depended entirely on a small team of analysts who were already stretched thin. Requests piled up, turnaround times ballooned, and business decisions were delayed because the people who needed data simply could not access it without a specialist acting as intermediary.

Leadership knew they needed a way to democratise data access without sacrificing accuracy or security. They had tried dashboards and BI tools before, but the questions people asked were too varied and unpredictable to anticipate with pre-built reports. They needed something that could handle novel, ad hoc questions and get smarter over time.

The Solution

BetterBrain built a self-learning AI agent that sits between the company's data warehouse and its users. When someone asks a question in plain English, the agent writes and executes SQL or Python code to retrieve the answer. But it does not just fire off a query and hope for the best. The agent reflects on the results, checks whether the output makes sense, plans next steps if the answer is incomplete, and asks clarifying questions when the request is ambiguous. Every interaction is human-in-the-loop, meaning an analyst reviews and approves responses before they reach the end user.

Under the hood, the system uses retrieval-augmented generation (RAG) with both BM25 and vector embeddings to find relevant context from previous queries, database schemas, and documentation. OpenAI o1-mini handles code generation because of its strength in reasoning through multi-step data transformations. The self-learning layer is what makes the system truly powerful: every approved query-answer pair gets fed back into the system's knowledge base, so the agent learns the company's data, terminology, and common patterns with every interaction.

The result is a system that gets noticeably smarter week over week. Queries that once required analyst intervention are handled autonomously. Non-technical users can ask questions in their own words and get accurate, contextualised answers without needing to understand SQL, table structures, or data modelling concepts.

The Number

• 80% reduction in analyst time spent on ad hoc data requests
• 75% reduction in time for non-technical users to pull data
• Self-learning loops mean the system improves with every query
• Data team refocused on high-value analytical work
• Non-technical teams gained direct, self-serve data access