Making business intelligence conversational
Easi builds software, cloud infrastructure and business intelligence tools for companies across Belgium. Spoom, its BI platform, is where those companies read their own numbers. Getting to one figure meant navigating dashboards and filters built for comprehensive reporting. Easi wanted to close that gap, letting users find a figure by simply asking.
The challenge
Spoom, the business intelligence platform developed by Easi, is designed to offer fast, user-friendly access to financial reporting through a comprehensive suite of dashboards and KPI summaries. That comprehensiveness came with a cost: getting to one figure meant knowing which dashboard held it, which filter isolated it, which configuration was set up to show it.
This was hardest on newer or less frequent users, who hadn't had time to learn where things lived. Easi wanted a way in that didn't depend on that knowledge, one that let any user, new or experienced, reach the same figure without first learning where it lived inside Spoom, through a more intuitive, conversational experience.
The solution
Our team built and implemented a generative AI agent that lets Spoom users ask a question in plain language and get an answer back, instead of a dashboard to search through. The agent works through a new endpoint on Spoom's own API, built by Easi specifically for this project, so its answers stay consistent with Spoom's own reporting.
The agent reads a question, works out which KPI and dimensions it's asking about, and calls the right combination of parameters to return a correct figure. Across a benchmark of diverse KPI questions, the agent reliably returned the correct answer, holding up even when the KPI or dimension wasn't stated precisely and had to be inferred from context.
The approach
Before we could build anything, we needed to know where a correct answer could actually come from. Spoom's raw data lives in its database, but the calculations that turn that data into a KPI sit in Spoom's own backend logic, so pulling numbers straight from the database would have bypassed that logic and risked wrong figures. That's why Easi built a dedicated API endpoint for this project, one designed to return figures computed the exact same way the dashboards already do.
With that endpoint in place, we turned to building the agent that would query it, testing a single agent against a multi-agent setup. The multi-agent version added more calls per question and made the reasoning behind each answer harder to trace, so we moved forward with the single agent, which proved faster, cheaper to run, and just as accurate.
We shaped that agent's reasoning around a ReAct approach: at each step, it states what it's trying to find, takes the action needed to find it, and uses the result to decide its next step, rather than guessing an answer in one leap. We integrated it with predefined queries that followed Spoom's existing data rules, reducing the risk of incorrect or unsupported queries.
