Workplaces may need employees to train custom AI tools

In short: More companies are finding that to make AI useful at work, they may need employees to help train it using their own hard-won know-how.

What's going on

A new Financial Times column argues that many of the skills needed to make workplace AI accurate are not written down anywhere. They live in employees’ heads as “tacit knowledge”, meaning know-how people use every day but find hard to explain in steps.

The piece compares this to an older idea from factory management. In the early 1900s, consultant Frederick Winslow Taylor tried to capture workers’ methods by watching and timing tasks, then turning that into official rules. The column suggests a modern version is happening now, except the goal is to feed that knowledge into AI systems.

The article points to examples where off-the-shelf AI still struggles with job-specific judgment. An investment firm, Bridgewater Associates, tested large language models, which are chat-style AI systems trained on lots of text (like a very fast reader that tries to guess the next word). Bridgewater found that versions of Gemini, Claude, and GPT were only right about half the time for certain finance tasks. The firm said better instructions, also called “prompts”, did not fix the problem. It had more success by fine-tuning a smaller model using its own experts and data, raising accuracy to about 85%.

What to watch

If companies need workers’ help to train these custom tools, employees may gain leverage. They can ask for clearer promises about pay, job security, and how AI will be used. The column warns that low-trust workplaces may struggle most, since people may not want to “pour what they know” into a system that could later be used to monitor or replace them.

Source: Financial Times