The Citibank Expert System
The Citibank expert system was Sawyer's 1984 AI application — the first AI system deployed by a major money-center bank. Using natural language processing and rule-based inference, it assisted in international banking decisions with a reliability that human operators could not match within its specific domain. The system represented the state of the art in symbolic AI applications of its era and demonstrated the genuine capabilities of expert systems when applied to well-defined problem domains. Its limitations — the decomposition of complex tasks into predetermined components, the operation by fixed logic, the absence of anything resembling emergence — were what led Sawyer to leave technology in the late 1980s for the study of human creativity. Forty years later, his framework for understanding what the expert system lacked turns out to be precisely the framework needed for analyzing what large language models provide.
In the AI Story
The expert system architecture Sawyer worked within was the dominant AI paradigm of the 1980s. Expert systems encoded domain knowledge as rules and used inference engines to apply the rules to specific cases. The Citibank system applied this architecture to international banking, producing reliable assistance within its programmed domain.
The system worked as designed. It was commercially valuable. It demonstrated what symbolic AI could do in well-defined domains. But it operated as a solitary processor — a single program applying fixed logic to inputs, with no capacity for emergence, no ability to produce outputs that surprised its programmers, no mechanism by which the processing could become something other than what the rules specified.
Sawyer's trajectory from this achievement to the study of jazz improvisation was not the career swerve it appeared to be. It was the same question approached from the other direction. The expert system could process banking decisions with more reliability than any individual human. The jazz ensemble could produce music that no individual musician — not the most reliable, not the most skilled — could have produced alone. The two capabilities were different in kind, and Sawyer's framework would eventually make the difference articulable.
Forty years later, the systems that followed the expert systems — large language models, neural networks, the full architecture of modern AI — can do things the Citibank system could not dream of. They can hold extended conversations. They can draw connections across the entirety of recorded human thought. They can produce outputs that surprise the humans interacting with them. The question Sawyer's framework forces is whether these new capabilities amount to the kind of emergence his jazz ensembles produced, or whether they represent a remarkably sophisticated version of what the Citibank system did — with all the same fundamental limitations.
Origin
Sawyer built the system while working in the financial technology sector in the early 1980s, applying his MIT training in computer science and electrical engineering. Citibank deployed it in 1984.
Key Ideas
First bank AI deployment. A specific historical achievement in applied AI.
Expert system paradigm. Representative of the dominant AI approach of its era.
Solitary processor. The system exemplified the limitations Sawyer would later articulate through his ensemble research.
Trigger for career shift. The limitations the system embodied led Sawyer to study human creative processes.
Retrospective significance. The framework developed to understand what the expert system lacked applies directly to modern AI.
Appears in the Orange Pill Cycle
Further reading
- Edward Feigenbaum and Pamela McCorduck, The Fifth Generation (Addison-Wesley, 1983)
- Pamela McCorduck, Machines Who Think (Freeman, 1979)
- Daniel Crevier, AI: The Tumultuous History of the Search for Artificial Intelligence (Basic Books, 1993)