Build-your-own GPT: Field lessons about configuring a proprietary generative AI solution

We study how a large German automotive manufacturer built an in-house, proprietary generative artificial intelligence (GenAI) solution to assist their employees with everyday work tasks. Using Simon's view of design as problem-solving as our interpretive scaffold, we examine the GenAI development initiative as a configurational search process in which both the problem space—use cases and user needs—and the solution space—the capabilities of the GenAI solution—are initially unknown and iteratively co-evolve as developers and users explore and gradually align both spaces. We suggest this process begins with affordance-led decomposition and unfolds through a heuristic search loop that involves probing through provisional solutions, capability re-basing, and governance-bounded configuration. Through this loop, a satisficing and bounded design configuration emerges in the form of a GenAI solution that is relevant to some but not all potential use cases and leverages some but not all GenAI capabilities. Our analysis draws attention to decomposition as a pragmatic organizational configuration mechanism for GenAI solutions and identifies transferable configurational heuristics that explain how developers and users jointly search for satisficing alignments between organizational problem and technological solution spaces when both are initially unknown and dynamically evolving.