Use-Case-Driven Model Selection and Configuration for Enterprise AI Workflows

Abstract

Enterprise adoption of generative artificial intelligence (GenAI) has grown rapidly, with 78% of organizations deploying AI in at least one business function as of 2025. Yet studies indicate that 95% if enterprise GenAI pilots fail to deliver measurable business impact. A key contributor to this failure is the ad-hoc, trial-and-error approach that organization use to select and configure large language models (LLMs) for production workflows. Multiple commercial and open-source models are now available, each with different strengths, but enterprise lack a practical methodology for matching models to specific workflow needs and configuring them well. This paper proposes a reference architecture for use-case-driven LLM selection and configuration in enterprise environments. The first layer introduces a Workflow task profiling framework that breaks the workflows into characterized task units along four dimensions: task nature, data profile, output requirements and interaction pattern. The second layer presents a Task-Driven Capability Matching approach that starts from the task profile, identifies the capability categories the task requires, and then identifies candidate models that satisfy those requirements. This task-first direction ensures selection decisions are driven by what the workflow needs, not by model marketing claims or benchmark rankings. The last layer defines a two-dimensional configuration approach where each parameter is driven by the task profile, the selected model, or both. Task-driven parameters such as chunking strategy and corpus structure might remain stable across model changes, while model-specific parameters such as prompt format and message structure change with the model. This enables model transitions while also accounting for relevance of the configurations to the workflow requirements and not just the models. The framework is demonstrated through end-to-end walkthroughs of representative enterprise workflows.