Enterprise AI Hits a Wall: Knowledge About How Work Gets Done Lives Only in People's Heads

Companies are pouring billions into enterprise AI, but they face a fundamental problem: most of the knowledge about how work really gets done lives not in databases and documents, but in employees' heads. Extracting that knowledge is much harder than it seems.

The company's invisible knowledge

Formal corporate systems such as CRM, ERP, Jira and Confluence record outcomes and formal processes. But they almost never capture what really matters: how exactly an experienced manager resolved a disputed issue with a key client, why a team has worked around a pipeline bottleneck in one particular way for years, which nuances are critical when dealing with a specific partner.

This phenomenon is called "tacit knowledge", a term introduced by the philosopher Michael Polanyi. The idea is that an expert often cannot explain why they make a given decision; they simply know. An experienced cook can tell by touch that the dough is ready. An experienced lawyer instantly spots the problematic clause in a contract. Without that kind of context, AI stays blind, no matter how powerful its architecture.

Why it is hard to digitize

Enterprise AI runs into this problem from several sides at once:

• Articulation: employees often cannot describe what they do themselves; they act on intuition built up over years
• Volatility: informal knowledge is constantly updated by new events, and documentation can't keep up
• Incentives: an employee has no direct motivation to explain how exactly they work, because it is their competitive advantage inside the company
• Context: the same rule works completely differently depending on the client, the deadline and the team's makeup
• Scale: a large organization has thousands of holders of unique knowledge, and each has their own version of the same processes

Even a powerful language model trained on a company's official documents will give advice that works on paper but diverges from reality. Bloomberg calls this the main unsolved challenge of enterprise AI today.

What companies are already trying

Several approaches are gaining popularity. Some companies run "knowledge mining" programs: series of structured interviews with key experts, whose recordings are then used to fine-tune internal models or fill corporate knowledge bases. It is labor-intensive but yields high quality.

Others deploy process mining tools, which automatically analyze ERP and CRM logs and reconstruct real workflows: what actually happens, not what the regulations say. The gap between the two can be striking.

Finally, a whole class of software has appeared: AI "shadows" that watch an employee's actions in the background and gradually build a model of their behavior. This raises serious questions about privacy and about ownership of the knowledge a person has accumulated over years of work.

"The real challenge is not training a model on company data.

The challenge is first getting that data out of people."

What this means

Enterprise AI is constrained less by technology than by organizational anthropology. Companies that learn to systematically extract and digitize their employees' informal knowledge will gain a real competitive advantage, not from smarter models but from better data on how their business actually works. That shifts competition from the question "whose AI is more powerful" to "who knows itself better".