Expert System
An expert system is a deployed knowledge-based program that captures the specialized judgment of human experts in a narrow domain, typically as a knowledge base of if-then rules plus an inference engine that applies them to a case, in order to give advice or make decisions that would otherwise require a specialist. The term names a product category, not a research paradigm: expert systems were the commercial face of symbolic AI during the late 1970s and 1980s, and they were the first form of artificial intelligence to be sold to businesses at scale.
The architecture separated what the system knows from how it reasons. A typical expert system had a knowledge base of rules elicited from specialists, a working memory holding the facts of the current case, an inference engine that chained rules forward from data or backward from a goal, and an explanation facility that could answer "why do you ask?" and "how did you conclude that?" by replaying the rules used. This separation meant a new domain could be built by replacing the knowledge base, which is how EMYCIN emerged as a reusable shell from MYCIN and how later tools such as OPS5 and NASA's CLIPS (released in 1985) let organizations build their own systems. Knowledge engineers, a new job title, interviewed experts and translated their rules of thumb into code.
MYCIN and XCON are the two landmark examples, and they succeeded for different reasons. MYCIN, developed at Stanford in the 1970s by Edward Shortliffe with Bruce Buchanan and colleagues, recommended antibiotic therapy for blood infections using roughly six hundred rules and "certainty factors" to handle uncertain evidence; in a 1979 evaluation published in JAMA, its recommendations were rated as good as those of infectious disease specialists, yet it was never used routinely in clinics, owing to liability, integration and acceptance issues. It descended from DENDRAL (Edward Feigenbaum, Joshua Lederberg and colleagues), which inferred molecular structures from mass spectrometry data. XCON, also called R1, built by John McDermott at Carnegie Mellon for Digital Equipment Corporation starting in 1978, checked and configured VAX computer orders, and it was an actual production success that grew to thousands of rules and was credited with large savings.
The knowledge acquisition bottleneck is what ultimately limited the category. Experts often cannot articulate their tacit skill as rules, disagree with one another, and drift out of date as the domain changes. As rule bases grew, rules interacted in unforeseen ways, so adding one could break another, and large systems became expensive to maintain. Expert systems were also brittle at the edges of their domain: faced with a case slightly outside its rules, a system did not degrade gracefully but gave confident nonsense or none at all. These problems, together with the collapse of the specialized hardware market, contributed to the disillusionment of the late 1980s covered under AI winter.
The idea survived in less glamorous forms and in successors. Tax and benefits eligibility engines, insurance underwriting, equipment diagnostics, loan decisions and clinical decision support alerts are still built as curated knowledge bases, often under the label "business rules" or "decision support," and regulated settings value their auditable reasoning. Modern machine learning solved the acquisition problem by learning from data rather than from interviews, yet it gave up the explanation facility, which is one reason hybrid designs, including language models consulting rule engines or knowledge graphs, are attractive. A common misconception is that an expert system means any AI that gives expert-level answers; the term properly refers to this explicit, rule-driven architecture, and a neural network that outperforms specialists is not an expert system.
Read expert system as symbolic AI turned into a product: it showed that captured human know-how could create real commercial value and be explained step by step, and its fate showed that hand-authored knowledge is the costly, fragile part, which is exactly the part later learning methods set out to automate.