The foundations are the real win

Ten of Belgium's leading finance teams spent a day at the Wintercircus in Ghent for an AI hackathon, building on real problems from their own departments. Eagl and Spott organised it. We came third with Budgie, an AI-powered budget platform.

Here's why we still treat this as a win. In a single day we shipped a working app, running on live financial data, with proper access control. The reason we could move that fast says more about our AI transformation than the scoreboard does.

‍

The problem we picked

Each month, budget owners get the same questions: why are we over budget here, under there? The answers scatter across static Looker dashboards, Slack threads, and the odd spreadsheet. Then someone in finance burns an afternoon reconciling them by hand.

So we built Budgie: one place where the data, the comments, and the reporting live together. A finance view consolidates everything for the team. A My Budget view lets a budget owner sign in and see only their own budget. On top sits a chat, wired to our AI platform, so you can ask questions about the figures and have anomalies flagged rather than hunt for them.

‍

We did not build a mock-up

Most teams ended the day with a live artefact: a page or an HTML mock-up running on a schedule. We ended with something different. A working web app, both sides functioning, behind real access control, running not on a sample dataset or an Excel export but on our real, live, consolidated financial data.

On that subject. In finance, the numbers have to be right every time. Budgie does not let the language model do the sums. The calculations run on curated data and existing queries. The AI just fills in parameters like the date range or budget owner, flags anomalies, and handles the language on top. The experience feels like a conversation, but the figures are the ones we already trust.

‍

The foundations we could build on

Building a working app on live data with secure access, in a day, is a high bar. We cleared it because of two investments we had already made.

The first is our data warehouse. Our financial and other business data already flows into BigQuery as curated, structured data products, so we did not start from Excel exports. We started from a clean, governed dataset. That is the difference between an app that demos well and one you can actually run.

The second is Trellis, our internal AI platform. Trellis lets us build agents on top of a base model, Claude, Gemini, or GPT, with their own instructions, tools, and skills. The tools connect securely to our knowledge bases (like Google Drive, Confluence and Slack) and to the data warehouse. Access is governed by group membership, so anyone using an agent only ever sees the data they are allowed to see. For finance, that is not a nice to have. It is the whole game.

Better still, we can improve the agent ourselves: new instructions, different tools, an extra skill, no engineering ticket for every change. On the day, finance shaped the product while IT focused on the platform, not on the plumbing.

This is the part of AI transformation where there are no shortcuts. Anyone can spin up an agent. Far fewer organisations can hand a finance team clean, governed data and a platform that connects to it securely, then let them iterate without waiting on engineering.

‍

Why it still felt like a win

We did not win the prize. (Congratulations to Lighthouse.) But it still felt like a win. We saw what our peers are building, got to work on a use case we knew mattered, had finance and IT building side by side, and most of all: we realised we could build on solid foundations.

That is what an AI transformation actually looks like. Not a winning demo, but a finance team that owns its own problems and has the building blocks to act on them fast. The trophy went elsewhere. The working app, and the head start behind it, came home with us.

‍