Microsoft executive urges enterprises to build their own AI learning loops and avoid model lock-in

In his keynote, “From Open Source to Agentic Systems: Building the AI Native Era,” Waite described the interaction between AI and human expertise as a learning loop. AI models can process enormous amounts of information and generate insights at a scale humans cannot match, he said, but people bring context and judgement. Employees put those insights into a wider context, use them to make conclusions, and turn information into knowledge. That knowledge then becomes part of the organization’s collective expertise, and with every iteration it grows, helping the business climb further.

Waite said the knowledge generated in that loop belongs to the business. “It’s based on your data,” he said. “It’s based on your people. It shouldn’t be given off to an AI model, that’s separate. And that’s why we believe that for organizations to build their own hill-climbing machines, to build their own learning groups, they need to build control with the sort of learning loops and not be beholden to a single AI provider.”

Everything AI generates is based, or should be based, on proprietary business information, and the decisions made from those insights depend on employees and their understanding of the company, Waite said. Businesses therefore need to ensure that ownership of the knowledge accumulated through these interactions is not handed over to an external AI provider. He suggested the challenge will grow as companies become more dependent on AI tools.

Knowledge produced through AI tools could also become tied to a particular model or provider. If a business later wants to adopt better technology as it emerges, that attachment can become a problem. To future-proof their processes, Waite argued, businesses should build their own learning loops and maintain control over the information, processes, and technologies that make them work.

On model choice, Waite said many businesses begin their AI journey with proprietary frontier models, especially when tackling problems they do not fully understand. As they become better at those tasks, they may no longer need expensive frontier models and can move to specialized models that are smaller or cheaper, or even open-weight models they can operate themselves.

“Having the ability to move between different models is important, especially as you think about your algorithms. Which models are you going to use? You shouldn’t be trapped with a particular kind of model because it has your knowledge,” Waite concluded. The key, he said, is to retain the freedom to switch and move along the way. That also means controlling where AI workloads run, whether in sovereign infrastructure governed by local regulations, on-premises systems, or public cloud environments where businesses can take advantage of greater scalability and cost benefits. Open source technologies provide the foundation for that flexibility, Waite concluded.