An expert system is a computer program designed to emulate the decision-making capability of a human expert within a narrow domain of Artificial intelligence, typically by applying a base of encoded rules to facts supplied by a user through an inference engine. Expert systems were the most commercially significant application of Symbolic AI and drove a major wave of corporate AI investment during the late 1970s and 1980s.
Architecture
A typical expert system separates a knowledge base, a collection of domain facts and if-then rules contributed by human specialists, from an inference engine, which applies logical reasoning, commonly forward or backward chaining, to derive conclusions or recommendations. Expert systems built on reasoning techniques pioneered by Allen Newell and Herbert Simon in the 1950s, adapting general problem-solving methods to encode narrow, domain-specific expertise instead. Many systems also included explanation facilities that could show users the chain of rules used to reach a conclusion, an early form of what is now discussed as Explainable AI, and confidence factors to handle uncertainty that pure logic could not express well.
Notable examples
Early landmark systems included DENDRAL, developed at Stanford beginning in the mid-1960s to infer molecular structure from mass spectrometry data, and MYCIN, developed in the 1970s to recommend antibiotic treatment based on patient symptoms and lab results; MYCIN reportedly outperformed junior doctors on some evaluations despite never being deployed clinically due to liability and integration concerns. In industry, Digital Equipment Corporation's XCON system, deployed from 1980, configured computer orders and was credited with saving the company tens of millions of dollars annually, becoming a widely cited proof of commercial value.
The 1980s boom
XCON's success helped trigger a broader expert systems boom through the 1980s, as corporations invested heavily in specialized "AI" consultancies and dedicated Lisp machine hardware optimized for symbolic processing. Japan's Fifth Generation Computer Systems project, launched in 1982, aimed to build massively parallel logic-programming machines partly around expert-system-style reasoning, spurring competitive investment from the United States and Europe.
Decline and the second AI winter
Expert systems proved expensive to build and, more critically, expensive to maintain: rule bases grew unwieldy as domains scaled, interactions between rules became difficult to predict, and systems handled novel situations outside their encoded knowledge poorly. The specialized Lisp machine hardware market collapsed as cheaper general-purpose workstations caught up in performance by the late 1980s. These failures contributed to the second AI winter, a sharp pullback in AI funding and commercial interest that lasted into the early 1990s, mirroring the earlier funding collapse documented in the 1973 Lighthill report.
Legacy
Expert systems established practices that persist in modern AI: structured domain knowledge encoding, explanation of system outputs, and rule-based validation layers still used in some deployed pipelines. Their conceptual descendants include modern Knowledge graph systems and rule-based Guardrails (AI) layered around less predictable Large language model outputs. Unlike modern Machine learning systems that learn patterns from data, expert systems relied entirely on manually encoded expertise, a limitation that motivated the broader shift toward statistical approaches from the 1990s onward.