DataGov-Factory: An Empirical Evaluation of Managed vs. Custom LLM Agent Architectures for Financial Data Governance

This paper presents the first controlled empirical comparison of three LLM deployment architectures for financial data governance: (A) platform-managed agents, (B) custom multi-agent factory, and (C) direct API calls. Evaluated on five governance tasks from the Czech Financial Dataset using claude-sonnet-4-6 with blind evaluation by claude-opus-4-6. Key finding: platform-level prompt caching reduces context costs by ~90%, making managed agents competitive on total spend while delivering 16% higher output quality than custom orchestration.