Introduction
In the early 1990s, while most AI research focused on the mathematical elegance of algorithms in controlled environments, researchers at NEC Corporation faced a different challenge: deploying Case-Based Reasoning (CBR) at a massive corporate scale. Their system, SQUAD (Software Quality Control Advisor), was designed to support 150,000 employees. This paper argues that successful large-scale AI requires treating organizational human processes as integral components of the system's "algorithm."
The Organizational Knowledge Loop
The core innovation of SQUAD is the "organizational algorithm." Rather than relying on a small group of experts to curate data, NEC leveraged its existing Software Quality Control (SWQC) activities to create a continuous knowledge acquisition cycle.
As shown in the workflow above, cases originate from various departments—including software factories and sales offices—and are funneled through review committees. These committees filter submissions into "Best," "Selected," and "Pool" cases. This human-led filtering ensures that only high-quality, relevant data enters the machine-executable case base, effectively solving the "knowledge engineering" bottleneck that plagues many large-scale AI projects.
Scale and Growth
The effectiveness of this organizational approach is evident in the system's growth. By 1991, the system managed over 20,000 cases, with approximately 3,000 new entries added annually.
This steady accumulation of data represents a decade of corporate experience. By formalizing this process, NEC transformed tacit individual knowledge into articulated corporate assets, providing a searchable repository of problem-solving expertise.
Technical Innovations: Handling Sparse Data
Beyond organizational structure, the authors introduced technical refinements to handle real-world data limitations. In a massive corporate database, data is often sparse; a typical case in SQUAD might only have 3.4 out of 75 possible features populated.
Traditional similarity metrics often fail in these scenarios because they are "distracted" by irrelevant or missing features. The authors developed the $Sim_X$ metric, which focuses exclusively on features specified in the user's query. This approach reduced computational costs to 1/10th of traditional methods while improving retrieval accuracy by 20%.
Impact and Conclusion
The integration of human organizational processes with machine algorithms proved highly successful. NEC estimated that the SWQC process and the SQUAD system contributed over $100 million annually in productivity and quality improvements. The SQUAD system demonstrates that for AI to scale, it must be flexible, robust against missing data, and deeply embedded within the social and organizational fabric of the company.
Summary updated