A dozen eggs has a price. It sits on a shelf, it gets scanned, and everyone who walks up to that shelf pays the same number. That assumption has held for so long that most people never think about it.
In December 2025, a joint investigation found that on one of the largest grocery delivery platforms in the United States, the same dozen eggs from the same store at the same moment could carry five different prices depending on who was looking. Instacart ended the practice within days. What followed was a state attorney general demand letter, a federal regulator’s civil investigative demand, a congressional oversight inquiry, and a legislative wave that has now reached more than 20 states.
For AI governance teams, the interesting part is not that a pricing algorithm behaved badly. It is that nothing in the system was technically broken. The model did what it was built to do. What was missing was the governance layer around it: disclosure, consent, discrimination testing, and a named human owner. Every one of those is now a legal requirement somewhere in the United States, and every one of them was available as a control long before the story broke.
Watch the full case breakdown below, or keep reading for the regulatory picture and the controls that would have caught this before launch.
What the investigation actually found
Consumer Reports, working with Groundwork Collaborative and More Perfect Union, recruited roughly 400 shoppers to buy the same basket of goods at the same time. The results were not subtle. Prices for identical items from the same store diverged by as much as 23 percent, and the basket-level variation worked out to enough money that a regular shopper could pay over $1,200 more per year than someone else buying the same food.
The New York Attorney General’s office, which sent Instacart a demand letter in January 2026, put the pattern more starkly: nearly three-quarters of the items tested were offered at more than one price, with an average gap of 13 percent between the highest and lowest price for a single good.
Instacart’s response matters, and governance professionals should read it carefully rather than dismiss it. The company said retailers set the listed prices, rejected the “surveillance pricing” characterization outright, and stated that the tests were randomized A/B experiments that did not use personal, demographic, or individual-level user data. It ended the program on December 22, 2025.
Take that account at face value for a moment, because it makes the governance lesson sharper, not softer. If the experiments really were randomized rather than personalized, the company still ended up with a regulatory investigation, a legislative case study, and a brand problem. The reason is that from the shopper’s side, a randomized price and a profiled price are indistinguishable. Nobody was told an experiment was running. Nobody could opt out. Nobody could point to a person accountable for the outcome. Absent those three things, a company loses the ability to prove which kind of pricing it was doing, and the burden of proof lands on it at exactly the worst moment.
That is the failure mode worth internalizing: an undocumented pricing experiment and an undisclosed surveillance pricing scheme produce the same evidence trail, which is to say none.
Three different practices that get called the same thing
Most of the public confusion, and a fair amount of the internal confusion inside pricing teams, comes from collapsing three distinct practices into one word. They carry different risk profiles and different controls.
Dynamic pricing responds to market conditions. Airline seats climb as departure nears, hotel rates rise during a conference, ride-hailing surges in the rain. The defining property is that everyone querying the system at the same moment sees the same number. It is well understood, broadly lawful, and largely outside the new pricing statutes.
A/B price testing shows different prices to randomly assigned groups to measure elasticity. The assignment is random rather than personal, which is a real distinction. But randomization is a property of the design, not something a consumer can observe or verify, and most of the new disclosure regimes were not drafted with a clean carve-out for it.
Surveillance pricing, sometimes called personalized algorithmic pricing, uses data about the individual to estimate what that individual will tolerate paying. The FTC’s market study of pricing intermediaries found that inputs can go down to mouse movements on a page and items abandoned in a cart, alongside location, demographics, and browsing history. The same study found these intermediaries serving at least 250 client businesses across sectors from groceries to apparel, which tells you this is infrastructure, not an experiment at one company.
The governance implication is that your pricing system needs to be able to prove which category it falls into. If the only person who can answer that question is the data scientist who wrote the feature list, you do not have a control. You have a dependency.
The law moved faster than most pricing programs did
Between 2025 and 2026 this went from a policy-wonk topic to an enforcement reality across multiple jurisdictions. The compressed picture:
New York enacted the first statute of its kind. The Algorithmic Pricing Disclosure Act, codified at N.Y. General Business Law section 349-a, took effect on November 10, 2025 and requires any covered business setting prices with an algorithm informed by consumer personal data to display a specific, mandated sentence next to the price telling the consumer exactly that. It survived a First Amendment challenge from the National Retail Federation, with the court treating the compelled statement as plainly factual. Civil penalties run up to $1,000 per violation, and “personal data” is defined broadly enough to cover anything reasonably linkable to a consumer or their device.
California is running two tracks. The Attorney General has framed undisclosed personalized pricing as a CCPA problem under the purpose limitation principle, on the theory that using data collected for one stated purpose to decide what someone pays is not a use consumers would reasonably expect, and announced an investigative sweep across retail, grocery, and hospitality. In parallel, AB 2564, the Surveillance Pricing Prohibition Act, passed the Assembly in May 2026 and would ban the practice outright for retailers rather than merely require disclosure.
Federal activity remains investigative rather than statutory. The FTC’s Section 6(b) study is ongoing, the agency issued a civil investigative demand to Instacart, and the House Oversight Committee opened an inquiry covering several platforms. Several bills to ban surveillance pricing have been introduced without clear paths to passage.
The EU got here first and is worth watching as a template. Article 6(1)(ea) of the Consumer Rights Directive, added by the 2019 Omnibus Directive, has required traders to tell consumers when a price was personalized through automated decision-making since May 2022. GDPR Article 22 and its restrictions on solely automated decisions with significant effects sit alongside it, and Article 9 imposes a consent requirement wherever sensitive categories of data enter the calculation.
For a multi-state or multi-region retailer, the practical consequence is that the compliance floor is now set by the strictest jurisdiction you operate in, and the jurisdictions are diverging rather than converging.
Four controls that would have caught this before launch
The four controls below map cleanly onto ISO/IEC 42001 and the NIST AI Risk Management Framework. None of them require you to abandon algorithmic pricing. They require you to be able to explain it.
1. Pricing transparency. If a price is shaped by an individual’s data, that individual has a right to know it happened and what categories of data were involved. In ISO/IEC 42001 terms this sits under Annex A.8, information for interested parties, and it is now a hard legal requirement in New York and across the EU. The evidence artifact is the disclosure itself plus a record of where and when it renders.
2. Consent and purpose limitation over personal data. Data collected to fulfil orders and personalize recommendations is not automatically available to set prices. Annex A.7 covers data provenance and purpose, and the CCPA enforcement theory California is advancing turns on exactly this gap. The test to run internally: for every feature in the pricing model, can you name the original collection purpose and show that pricing was disclosed as a compatible use?
3. Continuous discrimination testing. A pricing model does not need a protected attribute in its feature set to produce a discriminatory outcome. Location, device type, browsing history, and inferred household size are all proxies capable of correlating with race, income, disability, or age. NIST AI RMF MEASURE 2.11 asks for fairness and bias to be evaluated and the results documented, and ISO/IEC 42001 Clause 6.1.4 requires an AI system impact assessment on affected individuals and groups. Run this before deployment and on a defined cadence afterwards, because drift in customer data changes outcomes without anyone touching the model.
4. Named human accountability. Someone with a job title must own what the pricing model does, be able to explain its logic in plain language to a regulator, and be answerable for its outcomes. ISO/IEC 42001 Annex A.3.2 and NIST AI RMF GOVERN both make this explicit. “The retailer sets the price” and “the algorithm decided” are the two answers that guarantee escalation, and both were on display in this case.
A pre-deployment checklist for any pricing model
Before a pricing system that touches consumer data ships, an AI governance function should be able to answer these in writing:
- Which of the three practices is this: dynamic, randomized test, or personalized? What documented evidence supports that classification?
- What is the complete feature list, and what was the original stated collection purpose for each feature?
- Which proxy variables could correlate with protected characteristics, and what were the measured outcome differences across groups in the last test cycle?
- What disclosure does the consumer see, in what medium, at what point relative to the price?
- Can a consumer decline personalized pricing and still transact? What price do they see if they do?
- Who is the named owner, and can they explain the model’s behaviour without engineering support?
- What is the maximum price spread the system can produce for one item, and who approved that ceiling?
- If a regulator demands the pricing logic for a specific transaction on a specific date, how long does reconstruction take?
The last question is the one most programs fail. Pricing decisions are usually logged as outputs rather than as decisions with inputs attached, which makes retrospective explanation nearly impossible once a model has been retrained.
The point of governance is to be early
By the time this became a news story, the only options left were reactive. End the program, issue a statement, respond to the demand letters. None of that was a governance decision. It was damage control after the governance window had closed.
The window was open much earlier: at the design review where someone should have asked which category of pricing this was, at the data review where someone should have traced each feature back to its collection purpose, and at the launch approval where someone should have signed their name to it. Those are ordinary, unglamorous meetings. They are also the entire job.
An AI governance professional’s value is measured in the incidents that never reach a headline, and pricing is now a permanent item on that list. If you are building the skills to run these reviews inside your own organization, the ISO/IEC 42001 Lead Implementer certification is where the management system behind these controls is taught end to end.
One question worth sitting with, because regulators are actively debating it: is it ever fair for a company to charge you a different price based on your personal data?
