The integration of artificial intelligence into business processes is increasingly hitting the same barrier: "dry" data extracted from documents and instructions cannot convey the full depth of expert experience. This refers to so-called tacit knowledge — those practical skills and intuitive decisions that experienced specialists accumulate over years but rarely record in writing. As recent research shows, it is this resource that is becoming the main driver for creating truly effective corporate AI models.
General-purpose language models demonstrate impressive results in areas with large volumes of open data. However, in tasks requiring an understanding of a company's internal context and expert judgment, their effectiveness drops sharply. As philosopher Michael Polanyi aptly noted, "we know more than we can tell." This phrase perfectly describes the current dilemma of corporate AI.
Bridgewater: "Differentiated Intelligence" vs. General-Purpose Models
A striking example is the joint research by Bridgewater AIA Labs and Thinking Machines Lab. The teams tested leading models (Claude Opus, Gemini, GPT) on six financial tasks, including analyzing news and central bank documents. Using simple instructions, accuracy ranged from 45.6% to 50.1%. However, after Bridgewater experts prepared specialized prompts, the figure soared to 74.3–78.2%. Notably, automatic optimization of instructions did not yield additional gains.
The key breakthrough occurred at the next stage. Researchers fine-tuned the Qwen3-235B model on data labeled directly by Bridgewater specialists. The result exceeded all expectations: average accuracy reached 84.7%, and the number of errors decreased by 29.8% compared to the best general-purpose model. At the same time, the cost of processing tasks was 13.8 times lower. This approach, called "differentiated intelligence," involves configuring individual models for the unique processes and requirements of a specific organization. Essentially, it is not just fine-tuning, but the transfer of a company's collective experience into digital form.
Ford: Hunting for Veteran Expertise
A similar trend is observed in industry. Over the past three years, Ford Motor has hired, promoted, or rehired about 350 experienced technical specialists. Their task is to review designs, identify potential failures, train young engineers, and, most importantly, participate in improving automated quality control tools.
Ford's Vice President of Hardware Development, Charles Poon, stated directly: "Artificial intelligence is a great tool, but it is only as good as the information used to train it." He acknowledged that the company had not systematically preserved the knowledge of experienced engineers, and some of them left Ford before their expertise could be fully transferred into internal processes.
My analysis: We are witnessing a fundamental shift in corporate AI strategy. The race for raw data and computing power is giving way to a race for high-quality, expert-labeled "know-how." Companies that can effectively extract, structure, and transfer the tacit knowledge of their best employees into models will gain a tremendous competitive advantage. Those who continue to rely solely on general-purpose models and public data risk being left behind in the era of "differentiated intelligence."