The story behind Mon Agent Financier: the tool comparing 9 major banks in Belgium

The idea for Mon Agent Financier started as an internal challenge.

At the end of last year, Nico Vincent, co-founder and partner at Sailpeak, took part in a Trends-Tendances article titled “Et si ChatGPT remplaçait votre banquier ?” (What if ChatGPT replaced your banker?). The piece explored how artificial intelligence could reshape the way clients access banking services and what that would mean for the sector.

The discussions that followed internally quickly moved from vision to reality. If AI is becoming a new interface to access financial information, then the reliability of that information becomes critical. And that is precisely where tools like ChatGPT show their limits: they are powerful, but they are not built on structured, verified, market-specific banking data. Answers can sound convincing, even when the underlying data is incomplete, outdated or not specific to the Belgian market.

When the article was published in Trends-Tendances, one question quickly emerged internally: if this transformation is coming, what would it look like in practice? And more importantly, what would it take to make it reliable for a real user with real financial questions?

The team took this as a challenge. Instead of analysing it from the outside, they decided to build a concrete example: a system where every answer would be based on actual, comparable banking information.

That moment marked the start of Project Mon Agent Financier.

AI as a new access layer to financial services

For years, digital banking has been built around channels: websites, apps, customer portals.
Clients have had to navigate through those environments to find information, compare products and understand what applies to their situation.

Artificial intelligence introduces a different logic.

Instead of browsing, users ask.
Instead of reading multiple pages, they expect a structured answer.
Instead of adapting to the interface, the interface adapts to their question.

We do not see AI as replacing banks or human advisors. What is changing is the way information is accessed. As conversational interfaces become more common, the quality, structure and accessibility of financial data become central to the customer experience.

Mon Agent Financier was created as a practical exploration of that shift.
‍

Building on public information from verified sources

The project did not start from scratch.

Over the past months, we had conducted a large-scale benchmark on the digital maturity of banking websites in Belgium. To make that analysis possible, we had developed a full data collection and processing pipeline capable of extracting information from multiple types of public sources such as structured web pages, product descriptions and detailed PDF documents.

For Mon Agent Financier, we extended that infrastructure and reorganised it into a dedicated architecture.

Concretely, the system relies on three main layers.

First, a data acquisition layer.
This component collects publicly available content across the different banking platforms and transforms it into usable text. The challenge here is not only technical. It is also about consistency. The same type of product can be described in very different ways depending on the institution, and the relevant information is not always located in the same place or in the same format.

Then, a data treatment & storage layer.
Once the content is extracted, it is processed and organised into a coherent knowledge base. This step is essential: it allows the assistant to work with comparable information instead of raw documents. In practice, this means identifying key concepts, aligning terminology and ensuring that the data can be queried in a meaningful way.

On top of this foundation, we built a conversational interface that allows users to interact with the system in natural language.
‍

From a static RAG to an Agentic RAG System

To make the interaction more robust and closer to a real conversational experience, the assistant is built as an agentic RAG architecture rather than a simple retrieval pipeline.

Each user question first goes through a routing layer. This component identifies the language, the type of interaction and whether access to the knowledge base is required. Not every message needs a database query, for instance greetings, clarifications or follow-up questions can be handled directly, which improves both speed and fluidity.

When a factual banking question is detected, the router generates an optimised retrieval query. This is an important step: user questions are often long and ambiguous, while the database requires precise and structured queries.

The system then retrieves the relevant context from the knowledge base before generating the final answer. The response is therefore always grounded in the actual documents rather than produced from memory alone.

Measuring reliability, not only fluency

One of the key questions behind the project was whether a domain-specific architecture actually makes a measurable difference compared to general-purpose models.

To test this, we conducted an internal benchmark comparing Mon Agent Financier with widely used AI systems such as ChatGPT and Gemini across several dimensions: consistency, response speed and factual accuracy in a banking context.

The results show a clear pattern. Because the assistant relies on a structured and controlled knowledge base, its answers are significantly more consistent and less sensitive to prompt variations. This is reflected in a much lower variance score, meaning that the same question asked in slightly different ways leads to the same conclusion.

Beyond consistency, the system also performs strongly in single-shot and multi-turn interactions. Since the relevant information is retrieved before the answer is generated, the model does not need to “reconstruct” knowledge from general training data. It works directly with the actual banking content.

This illustrates a broader point: in specialised domains, performance is not only a matter of model size or reasoning capability. It is primarily a matter of data architecture and retrieval strategy.

Who it is for ?

Although the interface is simple, the tool serves two distinct audiences.

For consumers, it offers a way to understand the banking landscape without having to navigate through multiple sources. It makes comparison faster, clearer and more accessible.

For professionals in the financial sector, it provides a different type of perspective. Seeing offers through a single structured interface makes it easier to position products, analyse customer experience and identify potential areas for improvement.

In both cases, the objective is not to replace existing processes, but to demonstrate how access to information can be transformed.