CData ran 378 prompts through AI agents hooked up to CRMs, ERPs, project tools and data warehouses. By its own count, competing setups got 25 to 35% of them wrong, and it traces the gap to understanding, like what a field means or whose permissions apply. Its own Connect AI platform scored 98.5% on the same prompts. By that math it missed roughly 6 of the 378, against about 95 to 132 for the others. The same test found the cheapest correct answers came from economy-tier models, at about…