Mistral CEO Arthur Mensch said companies that do not keep AI under their own control should not expect the resulting growth to truly belong to them. He directed his criticism at closed-model vendors, saying they are now enforcing data retention policies. Once a company connects a model to its business context, the vendor can see that information and learn from it. He also said such providers have a track record of using what they learn to go after their most successful customers.
Open models as the starting point for enterprise control
Mensch’s position is blunt: business leaders should use open models. His argument is not framed around price alone. It is about control. If a company keeps feeding internal knowledge, workflows, and user interactions into a closed system, the model grows stronger while the company may lose leverage over a core layer of its own operations. For firms that want AI embedded deeply into daily business, that dependency can become substantial.
Store data and logs in open systems
He said switching to open models is still not enough. Companies also need to store data and records in open systems, or software vendors may block them from building AI systems outside the walled gardens those vendors control. If an enterprise cannot secure full access to data managed on its behalf, migration becomes harder and more constrained. Mensch added that AI itself can help accelerate that migration, giving companies a way to rebuild their data foundations faster.
Access controls must cover both hard and soft rules
Once companies regain control of their data, they still need to manage how AI systems access that data on behalf of human users. Not every employee should see the same information. In the example given, Bob should not be able to see what Alice is doing inside the company. The difficulty is that AI models are very good at finding weaknesses in need-to-know boundaries, so enterprises need systems that can enforce hard access rules along with models that can evaluate soft access rules.
The training flywheel is where defensibility is built
Mensch described a continuous training flywheel as the most important step. He said companies need AI systems that keep improving through interactions with employees and users, turning business-specific strengths into capabilities that vendors and rivals cannot copy. He also linked this to deployment costs, saying businesses can shrink models based on the distribution of inputs they actually see. With AI bills becoming increasingly significant, efficiency is no longer a side issue.
How Mistral says it approaches deployment
Mistral said it offers the required building blocks through its Studio control plane and Forge training platform, while its applied AI engineers and scientists work alongside customers to transfer knowledge before stepping back. On deployment, the company said its systems can run on customer-owned infrastructure or through a hosted service with zero data retention. In Mensch’s framing, a company’s edge should remain its own, and the switch should stay in the customer’s hands.