From Black Box to Teammate: A Practical Guide to Building with AI

Every company seems to be in a frantic race to integrate AI. But there's a huge gap between a flashy tech demo and a useful feature that actually solves problems for customers.

In a recent talk, Tom De Keyser and Levi Vandenbempt from the software firm Kunlabora laid out their company's three-year journey with AI. Their story is a refreshingly honest look at evolving from treating AI as a mysterious, separate project to making it a supportive, integrated part of their development process.

The "Bolted-On" Era

Tom began by making one thing clear: Kunlabora's team are software developers, not data scientists. Their job is to build great applications. Three years ago, their first foray into AI reflected that distinction.

For a project called “Quwa,” built for civil society organizations, the AI was treated as a completely separate entity. Tom called it a feature that was "bolted on to the side." The main app was built with one tech stack, while a standalone Python API handled AI-powered tasks like topic analysis and summarization.

This siloed approach created immediate problems. Maintenance was a huge headache; since Tom built the AI component, he was the only one on the team who could maintain it. But the bigger issue was that some features were driven by tech novelty and not by an immediate customer need. This created a wall between the core development team and the "AI thing," a lesson that forced a change in strategy.

The Risky Integration

Learning from that first attempt, the team took the opposite approach on their next project for Helan, a health fund. The goal was to create a digital assistant to help users find services they were entitled to. This time, they built the AI, a Retrieval-Augmented Generation (RAG) system, directly into the main application using Java.

There was no separate service and no single "AI expert" responsible for it. The entire team could now understand and work on the whole product, a major improvement in their workflow.

However, this tight integration exposed a much more serious problem: non-determinism. With the AI now so central to the app, its unpredictable nature could break the entire user experience. Tom shared a stark example where the system suggested palliative care products to a user who was not in palliative care. An error like that, he explained, instantly shatters user trust. The lesson was clear. Making AI the core of the application was too risky and unpredictable. The solution wasn't better AI, but a better-defined role for it.

AI as the Supportive Partner

This realization led to their third and current approach, which Levi explained using their latest project: a tool for social workers. The goal is to help them quickly find the right social benefits for their clients. Here, the team's philosophy shifted entirely.

Instead of being the main event, the AI acts as a smart assistant. The application's core function is a powerful, filter-based search that works perfectly on its own. A social worker can use these filters to find exactly what they need. As an optional enhancement, they can also describe their client's situation in simple language, and the AI will suggest relevant benefits or filters to apply.

As Levi noted, if you remove the AI, you're still left with a fully functional and useful application. It's an enhancement, not the foundation. The same team builds both the core app and the AI layer, which ensures the AI serves a practical purpose without the risk of derailing the product. This is their blueprint for the future: AI that supports the user, not tries to replace them.

Conclusion

Kunlabora's journey provides a clear roadmap for any company getting serious about AI. The team moved from a fragile, bolted-on feature to an all-in, high-risk system, before finally landing on a balanced approach. Their model treats AI as an assistive feature that's owned and understood by the whole development team.

Looking ahead, Tom said the next big challenge is figuring out how to maintain these inherently unpredictable systems over the long term. In a field moving at breakneck speed, his final advice felt universal. We should all approach this technology the same way we would prompt an AI, or talk to an anxious child: "Take a deep breath and work on this step by step."

Inspired by this practical approach to building with AI? Contact us to explore what’s possible.