Tacit Knowledge Is the Missing Layer in Enterprise AI

Enterprise AI systems often retrieve documents accurately but fail to capture tacit knowledge: the thresholds, exceptions, sensory cues and team judgments that experts use in real-world decisions. Better search ranking cannot recover expertise that was never recorded. The article recommends a structured knowledge-elicitation protocol focused on four questions: What threshold triggers action? When does the standard process fail? What evidence supports the decision? When should the issue be escalated? Language models can suggest follow-up questions, but experts must provide and independently validate the answers. A tacit-aware architecture should include four planes: capture, representation, serving and transmission. Systems should preserve provenance, confidence and operating context, cite evidence, abstain when information is missing, and refer users to experienced employees. Demonstration, shadowing and apprenticeships remain necessary for perceptual and collective knowledge that cannot be fully expressed in text. The article cites Michael Polanyi’s concept that people “know more than they can tell,” David Autor’s “Polanyi’s paradox,” and Gabriel Szulanski’s research on 271 best-practice transfers across 122 cases in eight companies. It recommends incident replay, bus-factor audits, abstention testing and transfer metrics such as faster proficiency and fewer repeat incidents. For traders, the message is relevant to enterprise AI, automation and cybersecurity software. Companies with stronger proprietary knowledge systems may gain a durable advantage, while vendors that rely only on generic retrieval may face limits in reliability and adoption.