Making data talk

Rather than rushing into development, we took a moment to zoom out first. Where were bottlenecks slowing down CrediBill’s operations? What challenges frustrated clients the most? One task stood out like a sore thumb: creating custom dashboards on demand. While the task itself seemed straightforward — extracting data, coding the query, formatting the output — it consumed far too much time. Not only for CrediBill’s team but also for their clients, who were left waiting for results before they could act.

So, we started reimagining how CrediBill’s user base could interact with their financial data. Instead of overwhelming users with endless dashboards and filters, we designed a system that allows users to interact with their data in the most intuitive way possible: by asking questions.

Our proof of concept brought this idea to life. Using a powerful large language model (GPT-4o mini through Microsoft Azure), CrediBill’s new system lets users interact with their data as if they’re talking to a colleague. Think of questions like:

• “Which customers were late with payments in 2024?”
• “What’s the average payment term for customers in Antwerp?”
• “Which accounts are exceeding their credit limits every month?”

The LLM handles the heavy lifting, instantly transforming these queries into clear, actionable insights, presented as tables, charts, or detailed summaries. It’s fast, intuitive, and puts decision-making back into the hands of financial professionals.

‍

‍

A multi-agent approach

‍At the core of the solution lies a multi-agent architecture, mimicking the efficiency of a well-oiled team. Instead of relying on a single AI model to handle all tasks, this multi-agent approach divides responsibilities among specialized agents. Here’s how it works:

1. Main agent — Think of this as the project manager of the system. When a user submits a query, the main agent breaks it down, assigns tasks to the right agents, and ensures everything runs smoothly from start to finish.
2. Database agent — This agent acts as the systems ‘data expert’. This agent understands the structure of CrediBill’s databases like the back of its hand. It identifies which datasets are relevant to the query, so only the most relevant information is processed.
3. Query agent — Once the database agent identifies the relevant data, the query agent takes over. This agent is responsible for translating the user’s natural language query into a precise SQL query.
4. Chart agent — Once the data is retrieved, the chart agent brings it to life. Whether it’s a table, pie chart, or bar graph, this agent knows the best way to make the answer easy to digest.

By breaking the system into these smaller, specialized agents, the architecture ensures that each part of the process is optimized for accuracy and efficiency. And because it’s modular, it can easily evolve with new features, data sources, or industry-specific requirements.

‍From proof of concept to production

‍Great ideas are one thing. Turning them into tools that people can rely on every day? That’s a whole different story. For CrediBill, getting this solution production-ready meant more than just building a functional system. Our goal was simple: refining the system until it worked seamlessly in the real world.Together with CrediBill, we tested the system in real-world scenarios and invited clients to ask their own questions during live demos. These sessions weren’t just about validation. They were about learning. For example, when the AI initially struggled to distinguish between “accounts” as customers and “accounts” as ledgers, it was a clear call to fine-tune. With every iteration, the system became sharper, faster, and more intuitive.The next chapter is about scaling. Security and trust are key priorities as the tool moves closer to production. Features like authentication and data segregation will ensure sensitive information remains protected.

By bringing generative AI into the heart of financial software, CrediBill is definitely leading the way. And this is just the beginning.