The trend of fine-tuning low-cost AI models using an enterprise’s proprietary data and expertise is emerging as a counterweight to advanced AI models from OpenAI and Anthropic.
Previously, specialized AI models in fields like law, healthcare, or education could only keep pace with older AI generations, but recent trials show the gap is narrowing significantly.
Harvey AI fine-tuned the Kimi 2.6 model for the legal sector, improving performance by nearly 40% and achieving results on par with advanced AI models at only about 1/11 of the cost.
However, this advantage quickly narrowed when Anthropic released a new version of Opus with higher performance on the same test.
Bridgewater Associates partnered with Thinking Machines Lab to fine-tune the Qwen3-235B model using internal data, workflows, and the experience of their investment team.
The fine-tuned model achieved an accuracy of approximately 85%, reducing errors by nearly 30% compared to advanced AI models in financial analysis tasks, at only about 1/14 of the cost.
This success stems from the model learning not just the correct answers, but also how Bridgewater’s experts actually make decisions in their daily work.
The article predicts the emergence of enterprises specialized in providing services to build and maintain custom AI models for individual organizations, rather than just selling access to foundational models.
Microsoft also warned enterprises against relying on a small number of closed AI models due to the risk of losing control over their data and intellectual property value.
The article simultaneously raises the question of whether employees are willing to transfer years of accumulated knowledge to train AI for the enterprise if doing so could diminish their own roles.
📌 Conclusion: The AI competition is shifting from building the most powerful models to exploiting the proprietary knowledge of each enterprise. Trials like Harvey AI or Bridgewater prove that low-cost models, once fine-tuned with internal data and experience, can match or exceed the performance of advanced AI models at only 1/11 to 1/14 of the cost. If this trend continues, the advantage may shift from AI model providers to enterprises that own proprietary data, workflows, and specialized expertise.
