- In 2025 the FTC collected $2.56B in algorithmic pricing settlements — from just two companies. If your prices run on a third-party algorithm, consumer data, or reinforcement learning, the question isn't whether regulators look. It's whether you can answer them when they do. 🧵
- The $2.56B breaks down as Instacart $60M and Amazon $2.5B. One is discrimination. One is collusion. Most teams have a defense for neither — and the FTC charged both in the same year.
- Theory 1 — pricing discrimination. Instacart's Eversight tool generated up to 5 different prices for the same item at the same store. Variation hit 23%. The FTC didn't need to prove intent. The outcome was enough.
- The trap is proxy variables. Your algorithm never sees race or income. It sees ZIP code, device type, browsing hour. An old Android in a low-income ZIP at 11PM is priced differently than an iPhone in a wealthy suburb. Census data shows those clusters track demographics.
- Theory 2 — algorithmic collusion. Amazon's Project Nessie extracted $1.4B by predicting when competitors would match a price hike, then raising prices on 8M items. No meeting. No agreement. Two algorithms reaching the same supra-competitive equilibrium.
- The vendor trap makes it worse. If your pricing vendor serves your competitors and pools data across clients, you can have hub-and-spoke exposure without ever talking to one. California's Cartwright Act (Jan 2026) wrote this into statute: a "common algorithm" with 2+ users.
- The penalties stopped being theoretical. CA: $6M or 2x gain, plus treble damages in private suits. NY: $1,000 per violation. EU AI Act (Aug 2026): €35M or 7% of global turnover. 51 bills across 24 states were filed in 2025 alone.
- RealPage's DOJ consent decree (Nov 2025) shows where this heads: no competitor data under 12 months old, no sub-state geography, symmetric guardrails, an antitrust compliance officer, 180 days to implement. That's the template regulators now reach for.
- Here's the real gap. Pricefx, PROS, Zilliant, Competera — every major pricing platform optimizes. Not one ships fairness testing or collusion monitoring. Law firms advise but don't instrument your engine. So nobody logs the data needed to prove a price wasn't discriminatory.
- Which means when the FTC's civil investigative demand lands, most companies spend 6-12 months in forensic data extraction — reconstructing decisions their pricing engine never recorded. Compliance is cheaper as architecture than as archaeology.
- Question for pricing and legal teams: when a buyer's AI agent negotiates with a seller's AI agent and the agreed price is discriminatory, who is liable? No regulator has answered that. If your pricing agent strikes that deal tonight, who do you name? #AntitrustAI #AICompliance
- We mapped every active law, settlement, and safe harbor into one compliance framework for pricing engines: https://veriprajna.com/solutions/ai-pricing-compliance