For a brief, intoxicating stretch in the 1980s, artificial intelligence had a business model. Not a research grant, not a thought experiment, but a product line with a sales force and a quarterly forecast. The product was the expert system, and for a few years it convinced corporate America that human expertise could be bottled, licensed, and run on a server in the basement. The story of how that worked, and then spectacularly didn’t, is one of the most instructive chapters in the history of the field.
Knowledge in a Box
An expert system is, at its core, a deceptively simple idea. Take the accumulated judgment of a human specialist, encode it as a pile of IF-THEN rules, and let a piece of software reason over them. The architecture splits neatly into two halves. First, the knowledge base: hundreds or thousands of rules of the form “IF the patient has a fever AND the culture is gram-positive, THEN consider a streptococcal infection.” Second, the inference engine: a domain-independent reasoner that chains those rules together, either forward from known facts toward conclusions or backward from a hypothesis toward the evidence that would support it.
The elegance was that the two halves were separable. The same inference engine could, in principle, diagnose an infection on Monday and configure a mainframe on Tuesday, provided you swapped out the rules. Expertise had been decoupled from reasoning. That decoupling was the whole pitch.
The Stanford Pedigree
The intellectual groundwork came out of Stanford. DENDRAL, begun in the mid-1960s, was the first serious attempt: it inferred molecular structures from mass-spectrometry data, encoding the heuristics that organic chemists used in their heads. It proved that a narrow, deep slice of human reasoning could be captured in code.
MYCIN followed in the 1970s, and it remains the canonical example. Built to diagnose blood infections and recommend antibiotics, it carried roughly 600 rules and, crucially, handled uncertainty through “certainty factors” rather than pretending the world was binary. In trials it reportedly performed on par with infectious-disease specialists. It was never deployed clinically, partly over the thorny question of who is liable when software prescribes the wrong drug, but it established the template. Its inference engine was later stripped of medical rules to create EMYCIN, “Essential MYCIN,” an empty shell you could fill with expertise of your choosing. The shell, not the diagnosis, turned out to be the commercial product.
The System That Paid for Itself
If MYCIN was the proof of concept, XCON was the business case. Developed at Carnegie Mellon by John McDermott and deployed at Digital Equipment Corporation around 1980, XCON, originally called R1, did something gloriously unglamorous: it configured VAX computers. Ordering a minicomputer in that era meant specifying compatible cabinets, cables, processors, and boards from a catalog of thousands of components, and getting it wrong meant shipping a customer a machine that didn’t work. Human technicians made errors constantly.
XCON grew to thousands of rules and checked orders automatically. DEC reportedly saved tens of millions of dollars a year by reducing configuration errors and the costly reshipments they caused. Here, at last, was an AI program with an unambiguous return on investment, and it sent a clear signal to every CIO in the country: this stuff is real, and it makes money.
The Gold Rush
What followed was a genuine industry. Companies sprang up to sell “expert system shells,” the EMYCIN idea generalized into commercial toolkits. A new job title appeared, the knowledge engineer, whose work was to sit with a human expert, extract the rules locked in their intuition, and translate them into code. It was part interviewer, part programmer, part therapist.
The hardware came too. Because most of this work was done in LISP, an elegant but memory-hungry language, vendors like Symbolics and Lisp Machines Inc. sold dedicated LISP machines: expensive single-user workstations optimized to run the language at speed. For a moment, owning a roomful of Symbolics boxes was the mark of a serious AI shop. Estimates of the period put the broader AI industry’s revenues in the billions by the late 1980s.
Where the Magic Failed
The cracks were structural, not incidental. Three problems proved fatal:
- Brittleness. An expert system was confident and competent right up to the edge of its rules, and then it fell off a cliff. Ask it anything outside its narrow domain and it produced nonsense with total assurance. It had no common sense and no idea what it didn’t know.
- The knowledge-acquisition bottleneck. Extracting rules from experts was slow, expensive, and maddening. Experts often couldn’t articulate what they knew; much of real expertise is tacit. Building a system could take years, and the rules conflicted, overlapped, and multiplied.
- No learning. The systems could not improve from experience. Every new case had to be anticipated by a human and hand-coded. As rule bases swelled into the thousands, maintaining them became a nightmare; XCON itself reportedly required a standing team just to keep it current as DEC’s product line changed.
The Winter
The reckoning arrived at the end of the decade. The specialized LISP-machine market was the first domino: cheaper, general-purpose workstations from Sun and the arrival of capable desktop hardware made dedicated AI machines look like overpriced relics, and that market collapsed around 1987. Corporate buyers, having spent heavily, found that maintenance costs ballooned while the systems stayed stubbornly narrow. Promises had outrun delivery. Funding dried up, startups folded, and the term “artificial intelligence” became something companies quietly stopped saying in pitch meetings. This was the second great AI winter, and it lasted into the 1990s.
It is worth being precise about what failed. The systems often worked exactly as designed. What collapsed was the belief that this design was a path to general intelligence, and the inflated economics built on that belief.
The Lesson That Carried Forward
The expert-systems era was not a dead end so much as a hard-won lesson in humility. Its central flaw, the inability to learn and the need to hand-author every scrap of knowledge, is precisely the problem modern machine learning set out to solve: rather than asking experts to write down their rules, we now let systems infer them from data. The IF-THEN cathedral fell, but the questions it raised about brittleness, tacit knowledge, and the gap between competence and understanding are very much still open. Anyone building AI today would do well to remember that the last generation also thought it had the magic in a box.