Building AI Capability Through Gamified Learning

While interest in agentic AI continues to grow, many organizations still struggle to translate AI experimentation into sustained capability. Providing access to powerful platforms is simply not enough. What matters is the environment: teams need to learn by doing, testing and applying AI to realistic business challenges – safely. This belief shaped our decision to participate in the AWS AI League, a global, hands-on competition that brought learning, experimentation and collaboration together.

A hands-on approach to learning at scale

Replacing the well-known AWS DeepRacer competition, the AI League challenged participants to solve real-world business problems using advanced AI tooling. The event combined expert-led workshops, self-paced experimentation and a gameshow-style finale, which took place at AWS re:Invent 2025 between December 1 and 5 in Las Vegas.

Participants were invited to develop AI solutions to business challenges and their models will be evaluated. This competition not only fostered individual skill development but also encouraged collaborative problem-solving, making it a standout platform for upskilling and networking within the AI community.

The value lay not in the competition itself, but in the learning model it enabled. We saw an impressive level of engagement, with over 500 Atos experts actively participating in the workshops.

Turning experimentation into insight

The technical challenge focused on building an AI-powered insurance underwriting assistant to advise on various scenarios across auto, property, travel, life and health insurance. The objective was to create a cost-effective, fine-tuned model trained on expert underwriting knowledge, capable of delivering accurate risk assessments and coverage recommendations that align with professional industry standards, all while reducing operational costs.

To tackle this challenge, participants leveraged AWS generative AI tools to fine-tune the large language models (LLMs) and train them on realistic insurance use cases. They generated their own datasets, carefully crafting instructions and ideal responses for the AI to learn from. The strongest results came from careful experimentation and evaluation, ensuring models performed reliably beyond the specific examples they were trained on.

The competition’s gamified structure, featuring a real-time leaderboard and a finale judged by both AI and human experts, pushed participants to iterate rapidly and collaborate effectively. The top five Atos participants achieved high scores on the benchmark.

The technical journey was marked by several key insights:

• Participants discovered the importance of generating diverse and high-quality training data, often developing their own tools to supplement the AWS applications.
• Repeated testing and refinement emerged as a critical factor for achieving top leaderboard scores, showing that small adjustments to how models were trained and evaluated could significantly improve results.
• Moreover, participants learned that traditional training metrics such as loss and perplexity did not always correlate with leaderboard performance, reinforcing the importance of testing AI against realistic, real-world use cases.

These lessons highlight that successful AI adoption depends as much on process, governance and evaluation as it does on technology choices. The AWS AI League proved to be a valuable opportunity to test the latest AI technologies in realistic scenarios.

Learning, confidence and collaboration

According to an Atos survey of the internal participants, intrinsic motivation and the opportunity to learn new technology were the primary drivers for engagement, while prizes and recognition played a lesser role.

More than half of the participants rated their experience with the Atos AWS AI League as excellent, and two-thirds reported increased confidence in their AI skills. The most engaging aspects of the event were learning new technology, the gamification elements such as the leaderboard, and the collaborative spirit that developed among participants. The event fostered a healthy balance between friendly competition and knowledge sharing, with participants supporting each other while striving for the top spots.