FinDaS AI

Article

When is charging two customers different prices defensible?

When the gap comes from what each customer costs you, and you can show the cost. A lender charging 18% instead of 12% on a €1,000 loan, because it expects to lose 8% of that loan rather than 2%, is passing on €60 of cost. The same 18% charged to a 2% customer, because they look unlikely to shop around, passes on nothing. Same price, and only one of them has an answer when the customer asks why. The legal exposure is narrower than that argument: a hidden consumer price set from personal data.

Try it

Same price, different story: move customer C

A lender and three customers, each borrowing €1,000 for a year. Every loan needs 10 points of rate, and the expected loss goes on top. A and B are priced that way. Set C’s risk and price, and see how much of the gap to A has a cost behind it.

2%, €20 on the loan
18%, €180 of interest
Funding, running costs and marginExpected lossNo cost behind itCost the price does not coverA’s rate

C's gap to A

6 pts

Explained by cost

0 pts

Not explained by cost

6 pts

Left on C after expected loss

€160

6 points of C's rate have no cost behind them. The lender keeps €160 on C, €60 more than a price set by cost leaves.

The example worked through below, on a loan repaid in one payment. The simple model’s limit is in the text.

Most price differences between customers are ordinary. The question to ask of any one of them is where it comes from: what the customer costs you, or what the customer will tolerate. The first you can explain to the customer. The second you can usually explain only to your board.

A €1,000 loan, priced three times

Let’s take a lender and three customers, each borrowing €1,000 for a year and repaying it in one payment. Every loan needs 10 points of rate for funding, running costs and the lender’s margin. On customer A the lender expects to lose 2% of the loan, €20. On customer B it expects to lose 8%, €80.

B costs €60 more to lend to, which on €1,000 is 6 points of rate. So A pays 12% and B pays 18%, and after expected losses both loans leave the lender the same €100. The whole gap is cost.

Customer C has A’s risk, 2%. The lender’s data says C rarely compares offers, so it charges C 18% too. B and C now pay exactly the same price, and the extra €60 the lender keeps on C has no cost behind it. It is there because C will tolerate it.

B and C both pay €180 of interest; the lender keeps €100 on B and €160 on C

Expected lossLeft, as a price set by cost leavesLeft, with no cost behind it
Computed when the page is built.
The same data as a table
Three customers on a €1,000 one-year loan
CustomerExpected lossRateInterestLeft after expected lossGap to A, pointsOf which cost
A€2012%€120€10000
B€8018%€180€10066
C€2018%€180€16060

6 points is the simple figure (simplistic on purpose). A loan that goes bad also stops paying interest: if a lost loan pays nothing at all, B needs 19.3% to bring in what A brings in, 7.3 points above A. So a lender can justify a little more than the simple gap, on its own repayment pattern. It has to be measured, though. The 8% has to come from how customers like B actually repaid (a gap labelled “risk” with no loss figure behind it is customer C with better paperwork).

My team built and ran this kind of pricing, as marketing work at a consumer-finance company: the model set the price a client got from that client’s risk. The credit-risk models were not ours.

Failing the test does not make a price wrong

The test is mine, and it is not a legal standard. A published student discount is a gap from tolerance too, and it is ordinary: it is on the price list, and anyone in the group gets it. So a gap with no cost behind it needs a second answer: is it open? Customer C fails both questions. Nothing about the risk explains the price, and C has no way of seeing it.

Passing does not settle it either. A real cost signal can still be off limits: the EU consumer credit rules that apply from 20 November keep social networks and special-category data out of the assessment, whatever they would predict.

Instacart: random prices, and the tests still ended

Instacart is the useful case because the prices do not look personal. In September 2025 Groundwork Collaborative, Consumer Reports and More Perfect Union ran five tests at three Safeway and two Target stores in Seattle, Washington DC, Saint Paul and North Canton, with shoppers filling the same cart at the same time and stopping before purchase. 437 shoppers reported prices and 193 submissions were analysed. It is not a representative sample.

74% of items had more than one price, and among those the gap averaged 13%

193 submissions analysed, from five tests. Not a representative sample, and shoppers stopped before purchase. Source: Groundwork Collaborative, Consumer Reports and More Perfect Union, Same Cart, Different Price.
The same data as a table
Instacart price tests, September 2025
MeasureFigure
Items with more than one price74%
Average gap, lowest to highest, among items that varied13%
Largest gap, lowest to highest, among items that varied23%
Submissions analysed193
Shoppers who reported437

An item could carry up to five prices. Basket totals differed by about 7% on average. (The report’s $1,200 a year is that 7% applied to $17,424, Instacart’s own estimate for a four-adult household. It is what could arise, not a cost anyone measured.)

The differences look random. Instacart told the report’s authors that assignment was random by category and location, and the authors found demographic differences small and not statistically meaningful. On 22 December 2025 Instacart ended all item price tests anyway and barred Eversight, its pricing tool, from running them. Store-by-store pricing and promotions continue.

My read: even random differences cost trust. A random gap fails the test as surely as C’s, since no cost sits behind it and a shopper sees only their own price, and “it was random” did not save the programme. Nobody measured the trust lost, so the evidence is one company’s decision, not a rate. That is enough for me to count the cost of being found out before running a hidden price test on consumers, and not enough to put a number on it.

The law as it stands

These are my readings of the texts as of 4 November 2026, not a regulator’s position, and this is not legal advice.

Pricing on cost, risk or segment is lawful and ordinary under the texts below, with one limit on segments: Connecticut’s ban reaches groups as well as individuals, so a segment built from personal data is not automatically safe there.

The exposure is a hidden consumer price set from personal data, mostly in food and retail

Ordinary

Lawful and ordinary under these texts

  • A price from what the customer costsMaryland and Connecticut exempt cost differences by name.
  • Credit priced on credit-report dataCarved out of Connecticut’s ban and New York’s bill.
  • Published or loyalty discountsExempt in Connecticut; promotions and loyalty in Maryland.
  • A moving price that is not personalisedOutside the EU’s disclosure duty (Recital 45).

Needs disclosure

The customer has to be told

  • New York, General Business Law 349-aDynamic pricing set by an algorithm that uses personal data.Since 10 November 2025
  • EU, Directive 2019/2161A price personalised by automated decision-making.Since 28 May 2022
  • EU credit, Directive 2023/2225A personalised offer or price from automated processing.Applies from 20 November
  • US credit, 15 USC 1681m(h)A notice rule on risk-based pricing, or a credit-score disclosure to all applicants. It does not restrict the pricing.In force

Banned

Off limits, in the scope named

  • Maryland, Chapter 154A higher price from personal data, at food retailers of 15,000 sq ft and up and third-party food delivery.Since 1 October 2026
  • Connecticut, Public Act 26-130Surveillance pricing by retail sellers of tangible goods and delivery services, individual or group.From 1 July next year
  • EU credit, Directive 2023/2225Social networks and special-category data in the credit assessment.Applies from 20 November

Not law: New York bill A9349-B. It would turn 349-a from disclosure into a ban, 180 days after enactment. See what is not settled, below.

My reading of the texts as of 4 November 2026, not a regulator’s position and not legal advice. The exemptions are in the table.
What each text does, as of 4 November 2026
TextWhat it doesIn force
New York, General Business Law 349-aDisclosure only, for dynamic pricing set by an algorithm that uses personal data, including data linked to a device. Insurers and Gramm-Leach-Bliley financial institutions are exempt.Since 10 November 2025
Maryland, Chapter 154Food retailers of 15,000 sq ft and up and third-party food delivery may not use personal data to set a higher price for a consumer. Promotions, loyalty, cost and supply differences are exempt.Since 1 October 2026
Connecticut, Public Act 26-130, section 11Bans surveillance pricing by retail sellers of tangible goods and delivery services, individual or group. Exempt: retention and acquisition discounts, cost and supply-demand differences, published or loyalty discounts, insurers, and credit pricing based on FCRA or ECOA data.From 1 July next year
EU, Directive 2019/2161Requires disclosure that a price was personalised by automated decision-making. Dynamic pricing without personalisation is excluded (Recital 45), and financial services are outside the Consumer Rights Directive.Since 28 May 2022
US credit, 15 USC 1681m(h), 12 CFR 1022.72 and 1022.74A notice rule on risk-based pricing that does not restrict the pricing. A credit-score disclosure to all applicants is the alternative.In force
EU credit, Directive 2023/2225No risk-based pricing notice. Consumers must be told of a personalised offer or price from automated processing (Art. 13, Recital 46). Human review (Art. 18(8)), and a provision on excessive rates (Art. 31).Applies from 20 November

GDPR Article 22, on automated decisions, says nothing about prices. An ordinary personalised price is not automatically caught, though one high enough to shut someone out could be. SCHUFA (C-634/21) treats a credit score as such a decision where the lender leans on it strongly, so a lender that prices straight off a score should read that one closely. And New York’s law does not reach a random test with no personal data, though assignment keyed to an account or a device could be contested.

One caution, and a narrow one. A price set from how badly someone needs the item sits closest to the edges of these texts, next to that shut-out price under Article 22 and the excessive rates of Article 31 of the EU credit directive. I read that as a caution, not a ban. How far it reaches depends on the sector, the product and the countries, so it gets scoped client by client.

What is not settled

  • New York’s ban. Bill A9349-B is not law on any official page: it passed both houses on 4 June 2026, and the bill page showed no signature on 4 November 2026. It would turn 349-a into a ban 180 days after enactment, exempting insurers, consumer-report credit pricing and certain discounts. I would plan for it and not describe it as in force.
  • New York’s disclosure law itself. A Second Circuit appeal was reportedly argued on 19 October 2026, and I found no decision.
  • The EU’s Digital Fairness Act. No proposal is tabled. The Parliament’s briefing of 5 October 2026 says the fourth quarter of 2026.
  • How each EU country transposes the consumer credit directive, which I have not checked.

How FinDaS does it

Pricing that differs by customer is part of the pricing work, risk-based pricing and personalised offers included. It starts by splitting each gap into the part your cost data explains and the part it does not, on your own figures. For a lender, the expected-loss estimates stay your risk team’s. AI agents build the analysis, automated checks and a review round check it, and what you act on goes to a panel of reviewer agents, and then I sign it. The legal reading stays with your counsel. Price tests are on pricing analysis, and the lender’s version of this work on risk-based pricing. A change you want to test first is sized in How many customers do you need to test a price change?

To set a rate for each risk band on a lender and watch margin move, try the risk-based pricing demo; to size a price test, the price-test calculator. When an AI agent recommends the price, the gap can come from the agent too: Who does your AI agent work for?

Sources

  1. Groundwork Collaborative, Consumer Reports and More Perfect Union, Same Cart, Different Price.
  2. Instacart, Ending item price tests on Instacart, archive copy of 9 October 2026.
  3. New York General Business Law 349-a.
  4. New York Assembly Bill A9349-B, bill page read on 4 November 2026.
  5. Maryland, 2026 Regular Session, Chapter 154.
  6. Connecticut, Public Act 26-130, section 11.
  7. Directive (EU) 2019/2161.
  8. Directive (EU) 2023/2225, Articles 13, 18 and 31, Recital 46.
  9. General Data Protection Regulation, Article 22; SCHUFA, case C-634/21.
  10. 15 USC 1681m(h); 12 CFR 1022.72 and 1022.74.
  11. European Parliament briefing on the Digital Fairness Act, 5 October 2026.

Start here

Up to thirty minutes, one question, a fixed proposal.

Book a free call