In 1980, Digital Equipment Corporation had a problem that was getting worse by the quarter. Every time a customer ordered a VAX minicomputer system, a human engineer had to figure out how to configure it — which components, which cables, which backplanes, in what arrangement. The configurations were fiendishly complex, the error rate was embarrassingly high, and the engineers who could do this well were expensive and scarce. DEC was one of the most successful computer companies on earth, and it was being quietly strangled by its own product complexity.
The solution came out of Carnegie Mellon, from a collaboration between DEC and a researcher named John McDermott. The system was called XCON — eXpert CONfigurer — and it was built using a production rule architecture called OPS5. No neural networks, no gradient descent, no learned representations. Just thousands of if-then rules, painstakingly extracted from the brains of DEC’s best engineers and encoded into a system that could apply them at a speed and consistency no human team could match.
By 1986, XCON was processing roughly 80,000 orders per year. Its accuracy on configurations exceeded 95 percent, compared to the 65–70 percent rate DEC had been achieving manually. The system was eventually credited with saving the company somewhere around $40 million per year. Over its operational lifetime, estimates for its total value to DEC ran into the hundreds of millions, and by some analyses, crossed a billion dollars. For a technology that the mainstream world had never heard of, that was a staggering number.
What makes XCON so worth revisiting now is what it reveals about the shape of AI progress. It was, in the language of the time, an expert system — a branch of AI that spent much of the 1980s dominating the field, collecting investment, and drawing genuine excitement before the so-called AI Winter arrived and froze most of it solid. The lesson most people took from that collapse was that expert systems were a dead end. Brittle. Unmaintainable. A footnote on the way to the real thing.
But XCON worked. It worked for years, in production, at scale, handling real business-critical decisions. The brittleness was real — maintaining the rule base as DEC’s products evolved was a genuine engineering burden, and eventually the system couldn’t keep pace with product complexity — but the core idea, that you could capture specialist reasoning and deploy it reliably across thousands of decisions, was completely correct. It just needed better machinery to run on.
What’s striking from today’s vantage point is how directly that idea flows into the large language models now being used for reasoning and planning tasks. The ambition is the same: take knowledge that lives in expert heads, make it available systematically, apply it at scale. The knowledge representation is radically different — learned from data rather than hand-coded from interviews — but the aspiration XCON embodied is essentially the aspiration behind every AI coding assistant, medical reasoning system, and legal analysis tool being deployed right now.
XCON also established something the field sometimes still struggles to communicate: that applied AI doesn’t need to be philosophically complete to be genuinely useful. It doesn’t need to pass a Turing test or exhibit general intelligence. It needs to solve the specific problem in front of it, reliably, at a cost that makes economic sense. XCON did that before the internet existed, before backpropagation was widely understood, before anyone had a GPU.
The trajectory from XCON to today’s reasoning models isn’t a straight line, but it’s a real one. And when the next generation of systems makes expert-level judgment available to everyone on earth, the long arc back to that VAX configuration tool will be clearer than ever.