The original title is "Your AI Agent Project Is Really a Data Project: The Data-Prep Tax"

Short summary

AI agent projects fail not because of model choice but because of unresolved data problems — the 'data-prep tax' teams underestimate when scoping. The article splits data prep into knowledge data (format and terminology drift) and operational data (identity resolution and access authority), arguing both require ongoing staffed upkeep, not one-time fixes. Before building an agent, test whether a competent person could answer its intended questions from your current data — if not, fix the data first, because the agent may be unnecessary.

• •Model is commodity; data quality determines agent success in production
• •Knowledge data fails on format/terminology; operational data fails on identity/access — estimate separately
• •Data prep is a standing operating cost, not a one-time phase; assign an owner or the agent silently degrades

Generated with AI, which can make mistakes.