> The Expert Systems Boom_

In 1980, Digital Equipment Corporation (DEC) deployed XCON (also known as R1), an expert system that configured VAX computer orders, and the commercial AI boom officially began. XCON was built using the OPS5 production rule language and contained over 10,000 rules encoding the arcane knowledge of how to configure a complete VAX system — which components were compatible, which cables connected what, which power supplies were needed for which configurations. Before XCON, this work was done by human technicians, and the error rate was staggeringly high: roughly 30% of VAX orders shipped with configuration errors, leading to costly returns and customer frustration. XCON reduced the error rate to under 2% and saved DEC an estimated $25 million per year.

The success of XCON triggered a gold rush. Corporations across every industry rushed to build expert systems — for medical diagnosis, oil exploration, financial analysis, manufacturing quality control, legal reasoning, and dozens of other domains. The most famous medical expert system, MYCIN, had been developed at Stanford in the 1970s to diagnose bacterial infections and recommend antibiotics. In controlled tests, MYCIN achieved a 69% accuracy rate — which sounds mediocre until you learn that infectious disease specialists scored only 80%, and general practitioners scored just 65%. MYCIN was actually better than most doctors. Yet it was never deployed clinically, partly due to liability concerns and partly due to the medical establishment's resistance to being second-guessed by software.

The expert systems industry grew at a torrid pace. Companies like Teknowledge, IntelliCorp, and Applied Intelligence Systems raised millions in venture capital. Dedicated AI hardware manufacturers — most notably Symbolics, which built specialized LISP machines — saw their stock prices soar. By the mid-1980s, the global expert systems market was valued at over $450 million. Fortune 500 companies established internal AI departments. The Japanese government launched its Fifth Generation Computer Systems project (FGCS) in response, and the U.S. and UK created their own national AI programs to avoid falling behind.

But expert systems had a fatal flaw: they were brittle. They worked well within their narrow domain of expertise but failed catastrophically when confronted with situations that fell even slightly outside their rules. XCON could configure a VAX with superhuman accuracy, but if DEC introduced a new component, a human had to manually add rules to accommodate it. Maintaining the knowledge base became an ever-growing burden — XCON eventually grew to over 31,000 rules, and keeping those rules consistent and up-to-date required a full-time team of knowledge engineers. This was the "knowledge bottleneck": extracting expertise from human experts and encoding it in rules was slow, expensive, and error-prone. The experts themselves often could not articulate why they made certain decisions — their knowledge was intuitive, not rule-based. This fundamental limitation would eventually bring the entire expert systems industry crashing down.