AI Agent Projects Fail on Data Prep, Not Model Choice

A developer demonstrates that AI agents return wildly wrong answers when customer data is inconsistent, e.g., $120,000 vs $1,345,000 vs $275,000 for the same pipeline. The article argues that data preparation—not model selection—determines agent success, with two distinct data problems: knowledge data (format/consistency) and operational data (identity/authority). It warns that unprepared data causes recurring delays and recommends testing if a human can answer correctly from the data before building an agent.