Expert systems: when AI became a business (1980s)

For most of the 1970s, artificial intelligence was a university discipline with modest budgets and lofty promises. By the mid-1980s, it had become a billion-dollar industry. The technology that drove this transformation was the expert system, and the story of its rise and fall remains one of AI’s most instructive cautionary tales.

What expert systems were

An expert system is a program that mimics the decision-making of a human specialist. It has two core parts:

• A knowledge base, hundreds or thousands of “if-then” rules encoding domain expertise (if the patient has fever and a stiff neck, consider meningitis).
• An inference engine, software that chains the rules together to reach conclusions, sometimes with confidence scores to handle uncertainty.

The idea was coined by Edward Feigenbaum at Stanford, who argued that “intelligent systems derive their power from the knowledge they possess rather than from the specific formalisms and inference schemes they use.” This was a deliberate break from earlier AI research, which had tried to build general-purpose problem solvers. Expert systems were narrow, practical, and, crucially, commercially viable.

The pioneers

The first notable expert systems came out of Stanford in the 1970s:

• DENDRAL (1965–) identified organic molecules from spectrographic data, the first system to demonstrate that encoding specialist knowledge could outperform general heuristics.
• MYCIN (1972–) diagnosed bacterial infections and recommended antibiotics. It was never deployed clinically, but it showed that reasoning about uncertainty could be formalised using confidence factors. In blind tests, MYCIN matched or exceeded the accuracy of infectious disease specialists.

These academic successes caught the attention of industry.

The boom

The commercial explosion began around 1980, when Digital Equipment Corporation (DEC) deployed XCON (also known as R1), an expert system that configured orders for VAX minicomputers. By the mid-1980s, XCON contained thousands of rules and was credited with saving DEC an estimated $40 million per year in avoided errors and labour.

Similar systems appeared in oil exploration (Shell’s PROSPECTOR), finance (credit assessment), manufacturing, and logistics. A new profession emerged, the knowledge engineer, whose job was to extract expertise from human specialists and encode it as rules. Feigenbaum and journalist Pamela McCorduck captured the spirit of the era: knowledge had to be “mined out of experts’ heads painstakingly, one jewel at a time.”

Startups multiplied. Companies like Teknowledge, IntelliCorp, and Carnegie Group sold expert system shells, generic inference engines that customers could fill with their own rules. Two-thirds of the Fortune 500 reported using expert systems by the late 1980s.

The geopolitical dimension

The boom was amplified by international competition. In 1982, Japan launched the Fifth Generation Computer Systems project (FGCS), a ten-year government initiative to build machines capable of logical inference and natural language understanding. Western governments panicked. The US responded with the Strategic Computing Initiative; Britain launched the Alvey Programme. Funding poured into AI research at a pace not seen before, or again until the deep learning era.

The bust

By the late 1980s, the cracks were visible:

• Knowledge acquisition bottleneck: extracting rules from experts was slow, expensive, and fragile. Experts often could not articulate how they actually made decisions.
• Brittleness: expert systems worked within their encoded rules but failed unpredictably at the edges. They had no common sense and no ability to learn from new data.
• Maintenance: as rules accumulated, systems became difficult to update and debug. Contradictions crept in.
• Hardware costs: many expert systems ran on expensive dedicated Lisp machines that could not compete with increasingly powerful general-purpose workstations.

Japan’s Fifth Generation project quietly wound down without delivering on its promises. The Lisp machine market collapsed. AI companies went bankrupt or pivoted. Funding dried up. The second AI winter had arrived.

Why it matters

Expert systems proved that domain-specific AI could deliver real commercial value, XCON alone justified the entire field’s investment. But they also demonstrated the limits of symbolic AI: hand-coded rules cannot scale to the messiness of the real world.

The lessons shaped what came next. When machine learning returned in the 1990s and 2000s, it succeeded precisely where expert systems had failed: learning patterns from data rather than having them manually encoded. Today’s large language models are, in a sense, the answer to the knowledge acquisition bottleneck, they absorb expertise from text at a scale no knowledge engineer could match.