- A certified minority-owned supplier with a 98.1% on-time delivery rate scored BELOW an incumbent at 97.2%. Same metric. Worse score for the better number. Not a bug — it's how confidence-weighted procurement AI works. 🧵
- The mechanism: delivery is weighted by transaction count. Incumbent: 4,200 transactions → 24.1 of 25. The MBE: 180 transactions → 16.8 of 25. The newer supplier delivers better and gets penalized for lacking a long enough history to "prove" it.
- Then it compounds. Lower score → fewer awards → fewer transactions → even lower confidence next cycle. The supplier never accumulates the data that would make the algorithm trust them. A self-reinforcing wall no one coded on purpose.
- There's a legal test for exactly this. The EEOC's four-fifths rule (29 CFR 1607.4): any group's selection rate must be at least 80% of the top group's. Written for hiring — but the same statistics apply to supplier selection.
- Run the numbers. If your AI advances 60% of non-diverse suppliers, it must advance at least 48% of MBE/WBE suppliers. Volume-weighted scoring routinely advances ~22%. Disparity ratio: 0.37. That is prima facie evidence of adverse impact.
- Now the part leadership doesn't want to hear. SAP Ariba, Coupa, GEP, Ivalua — all shipping agentic supplier scoring in 2026. Number that publish disparate-impact testing or fairness metrics on that scoring: 0 of 4.
- And they won't fix it for you. Their AI is tuned for cost across the whole customer base; per-customer fairness configs aren't how platform economics work. The speed is theirs, the fairness is yours — and courts are expanding vendor liability as contracts push it back to you.
- Meanwhile regulation pulls both ways. FAR Part 19 mandates small-business and diverse-supplier goals. EO 14319 (July 2025) bans "woke AI" in federal procurement. The only way through both is provable mathematical fairness — not a DEI program a court can second-guess.
- This is why procurement AI is stuck. 49% of teams are piloting; just 4% have deployed (ProcureAbility, 2026). A black-box score you can't defend to a stakeholder — or an auditor — doesn't make it past the pilot.
- So we built the missing layer: vendor-agnostic fairness auditing. It connects to Ariba, Coupa, GEP or Ivalua, runs four-fifths and disparate-impact tests on supplier scoring, and produces the mathematical proof your AI treats every supplier category equitably.
- Honest question for procurement leaders: if a supplier filed an adverse-impact complaint tomorrow, could you produce the math showing your AI scored them fairly — or only the vendor's marketing page? #ProcurementAI
- We wrote up exactly how the bias forms and how to audit for it: https://veriprajna.com/solutions/procurement-ai-fairness