A supplier with a better on-time delivery rate scored 21 points lower than the incumbent. The AI wasn't broken — it was working exactly as designed. A sourcing event for industrial fasteners. Supplier A — large incumbent, 4,200 transactions — scores 92. Supplier B — certified minority-owned business, 180 transactions — scores 71. On merit? Look closer. Delivery is 25% of the score, and the AI weights it by transaction count. Supplier B's 98.1% on-time rate beats Supplier A's 97.2%. But B gets dragged down — not for being worse, just for having less history to be "confident" about. The same penalty hits quality and financial stability — and it's self-reinforcing: score lower, win fewer contracts, log fewer transactions, score lower again. Run the EEOC four-fifths test and it turns legal. If your AI advances 60% of non-diverse suppliers but only 22% of MBE/WBE suppliers, that's a 0.37 disparity ratio — well under the 0.80 line for prima facie adverse impact. For a federal contractor with FAR Part 19 goals, that's an audit finding. Worse, EO 14319 now bars federally-procured AI with a "social agenda" — so the only defensible answer is math, not a DEI policy. None of the four major platforms — SAP Ariba, Coupa, GEP, Ivalua — publish fairness metrics on their supplier scoring. The speed is theirs. The proof of fairness is yours to build. If you run supplier scoring AI: have you ever tested it for disparate impact, or just trusted the score? #AIGovernance #ProcurementAI