GENAI4dev, 6 months later: REX about deploying coPilot and Cursor to 80 devs

Introduction to AI Deployment in Large Teams

In the session "GENAI4dev 6 Months Later: REX About Deploying Copilot and Cursor to 80 Devs Since Oct 2024 with Hub," Hubert from Xfabrica provided an insightful overview of his experience deploying AI coding assistants to enhance developer productivity. The session focused on the practical aspects of deploying GitHub Copilot and Cursor to a large team of developers, offering reflections and actionable insights from real-world experimentation.

Challenges and Strategic Planning

One of the main challenges Hubert highlighted was the issue of shadow IT, where developers were independently experimenting with AI tools without proper oversight. This led to the necessity of a controlled deployment to prevent risks associated with sensitive code exposure. Hubert emphasized the importance of involving the Chief Information Security Officer (CISO) and security teams early in the planning phase to manage these risks. He noted, “These tools are new, they can be scary, they can perhaps share some secret code all over the internet so we have to manage that.”

Phased Deployment and Community Support

Hubert's strategy involved a phased deployment approach, starting with mature teams that had established processes and metrics. This method minimized the risk, as stated by Hubert, “The risk is too dangerous. I repeat, it is an experimentation. We don’t want to break everything.” AI experts were embedded with the teams, and regular metrics were tracked to assess progress. Community support channels were crucial for fostering peer-to-peer problem-solving.

Tool Selection and Implementation Success

The tools selected for this deployment included GitHub Copilot and JetBrains AI. Local running of large language models (LLMs) was avoided due to insufficient computing power. The onboarding process was iterative, with feedback sessions to address skepticism and enhance the deployment process. Hubert credited the project's success to a dedicated core team and substantial vendor support. Key outcomes included a productivity boost, faster feature delivery, and improved code quality, with Hubert stating, “The money we spent here and the money we save there, the return on invest for six months is 150%.”

Managing Risks and Ethical Considerations

Hubert stressed the importance of managing AI-related risks such as hallucinations and over-reliance by junior developers. Maintaining strong testing and code review practices was essential. Ethical guidelines and security measures were strictly enforced, with Hubert noting, “All the libraries used and everything suggested by the AI have to be approved by the company. So we stay in the frame. It was not the open bar.”

Recommendations and Cultural Impact

Hubert concluded with practical recommendations for organizations looking to implement AI tools: start small, measure often, and prefer flexible licensing to stay agile. Emphasizing human oversight and peer support is crucial. Hubert highlighted the cultural impact of the project, which improved developer morale and became a source of company pride. He emphasized, “Keep your mind in the loop. Don’t let the guy alone on the screen. And don’t only deploy licenses. They have to be accompanied.”

About The Speaker

Hubert Gregoire

CTO, exFabrica

Merging deep tech and strategy, Hubert drives global AI adoption and modern software practices with real-world impact.