The success of artificial intelligence implementation in companies can depend on employee involvement in training specialized models. Financial Times columnist Sarah O’Connor reported on this.
This discussion revolves around tacit knowledge—practical skills and decision-making processes developed through experience, which are rarely documented in instructions and corporate documents. The FT connects this concept to philosopher Michael Polanyi, who articulated the principle: "we know more than we can express in words."
According to the author, universal AI models perform better in areas with large volumes of open data and clear verification criteria. In tasks where a company's internal context and expert judgment are crucial, written instructions alone may not suffice.
The publication compares this situation to the work of American engineer Frederick Winslow Taylor, who, in the early 20th century, sought to systematize factory workers' experiences through observation, recording, and time measurement.
Bridgewater Fine-Tunes Model Using Expert Assessments
One example is a joint study by Bridgewater AIA Labs and Thinking Machines Lab, published on June 30. The teams tested AI on six tasks from the investment firm's workflows: the models analyzed financial news, central bank documents, and other materials potentially significant for decision-making.
Using basic instructions, Claude Opus 4.6 and 4.8, Gemini 3.1 Pro, GPT-5.4, and GPT-5.5 achieved accuracy rates between 45.6% and 50.1%. Expert-prepared prompts from Bridgewater boosted this figure to 74.3–78.2%. Automatic optimization of instructions did not yield further improvements.
Source: Bridgewater AIA Labs and Thinking Machines Lab.“An explicit prompt can only convey the part of intuition that the expert can articulate. However, the most critical judgments are often the hardest to express,” the authors noted.
The researchers then fine-tuned Qwen3-235B using data labeled by Bridgewater specialists. The system's average accuracy reached 84.7%. According to the developers, the error rate decreased by 29.8% compared to the best-tested universal model.
The cost of processing tasks was found to be 13.8 times lower. The authors referred to this approach as “differentiated intelligence,” which involves customizing individual models to fit the processes and requirements of specific organizations.
Source: Bridgewater AIA Labs and Thinking Machines Lab.Ford Expands Team of Experienced Specialists
The FT also highlighted Ford's example. Over the past three years, the automaker has hired, promoted, or re-engaged around 350 experienced technical specialists. They review designs, identify potential faults, train younger engineers, and contribute to enhancing automated quality control tools.
Ford's Vice President of Hardware Development, Charles Pun, stated that the effectiveness of AI depends on the quality of the information used for training.
“Artificial intelligence is a great tool, but it is only as good as the information used to train it,” he remarked.
Pun noted that the company has not systematically preserved the knowledge of experienced engineers. Some specialists left Ford before their expertise could be fully integrated into internal processes.
Recall that in October 2025, ForkLog explored how developers adapt large language models for office tasks.