The Knowledge Acquisition Bottleneck, Then and Now: From Expert Systems to Large Language Models

The knowledge acquisition bottleneck was the greatest obstacle of the second AI boom: expert systems could reason with rules, but extracting those rules from human experts proved extraordinarily difficult. This essay revisits CA Outlook, an expert system the author developed at IBM Japan in the late 1980s to forecast hardware revenue, and asks what has become of the bottleneck in the age of large language models (LLMs). By learning from vast amounts of text, LLMs appear to have cleared the wall of knowledge acquisition. Yet organization-specific tacit knowledge, the judgment of when to trust model output, and the redesign of work all remain on the organization's side. The bottleneck has not disappeared; it has moved to organizational capability, as the IT–organizational capability dependency theory predicts. The CA Outlook experience also shows that management trusted explainable multiple regression over an opaque AI, a lesson that applies to today's black-box models. Finally, the essay argues that two features of Japanese management, multi-skilling through job rotation and long-term employment, produce people who can supervise chains of AI agents and workplaces where employees have no reason to withhold know-how from AI, giving Japan a real but time-limited opportunity.