#1000 1980 · Digital Equipment Corporation (John McDermott, Carnegie Mellon) · Computer manufacturing
DEC put its scarce order-configuration experts' rules into software any clerk could run
the problem
DEC's VAX had thousands of components, and only a few senior staff could configure an order that actually worked
background
A wrongly configured VAX order meant DEC shipping incompatible parts, then absorbing the cost of replacements and delayed installs, so every order effectively needed a scarce senior configuration specialist's sign-off. That specialist's judgment had taken years to build and could not simply be duplicated by hiring, since duplicating it meant duplicating years of experience.
Carnegie Mellon researcher John McDermott set out to see whether that judgment was really an inexpressible intuition or a large but learnable set of consistent compatibility rules, by systematically interviewing DEC's configuration experts about exactly how they reasoned through an order.
what everyone would do
The standard fix for a configuration bottleneck was to hire and train more senior configuration specialists, accepting the years each one took to learn DEC's ever-expanding VAX component catalog well enough to be trusted with a customer order.
what they saw
McDermott saw DEC's experts weren't using ineffable intuition; they applied a large, finite set of consistent rules, learnable over years, that interviews could extract into a program a novice could run.
the move
McDermott built R1, later renamed XCON ('expert configurer'), a rule-based program written in OPS5 that encoded DEC's experts' own if-then reasoning, eventually around 2,500 rules, so an ordinary order-entry clerk could run a proposed order through the program and get back a fully compatible, buildable parts list.
why it works
A junior clerk failed at configuration not because the reasoning was too subtle to execute, but because they didn't yet know which of thousands of possible rules applied to a given order — a large-but-bounded lookup problem software is good at and humans are slow to memorize. Once interviews had captured DEC's experts' actual decision rules as an explicit rule base, the clerk's job changed from 'recall the right rule from memory' to 'enter the order and read the answer,' which is why accuracy jumped immediately rather than improving gradually the way a human trainee's would.
the payoff
By 1986 XCON processed roughly 80,000 orders a year at 95–98% accuracy, saving DEC an estimated $25 million a year in parts and rework.
where it breaks
Rule-based encoding only works while the underlying product line changes slowly enough for someone to keep rewriting the rules; DEC eventually needed a dedicated team just to maintain XCON's rule base as VAX options kept expanding. A domain where the right answer depends on judgment calls that genuinely vary case to case, rather than fixed compatibility logic, will defeat this approach.
what came after
XCON became the most cited commercial success of the 1980s expert-systems boom, spawning sibling systems for the sales force and site planning, and setting the template other manufacturers copied for encoding scarce configuration expertise into software.
references
- [1]R1: An Expert in the Computer Systems DomainProceedings of the First AAAI Conference on Artificial Intelligence (AAAI-80), 1980cdn.aaai.org