Case Study 2 — The Price-Optimization Bans: When Data-Driven Pricing Severs Price from Risk

A real, public regulatory episode, examined from the chapter's complementary angle: not a model that discriminates by a protected proxy, but a profitable, "legal-sounding," data-driven practice that crosses the other fairness line — the one that says price must reflect cost, not the insurer's ability to extract more from a captive customer. The bulletins, the multistate regulatory response, and the practice itself are matters of public record; this study keeps every specific qualitative.

Background

Sometime in the 2010s, advances in data and modeling made possible a kind of pricing the rate-manual era could not have implemented at scale. Insurers had always known, in a rough way, that some customers shop aggressively and others renew year after year without comparing prices. What changed was the ability to model that behavior at the individual level — to predict, for a given policyholder, their price elasticity: how much a premium could rise before they would shop, switch, or lapse. With that prediction in hand, an insurer could set a premium based not only on the customer's expected loss but on what the customer was willing to pay.

This practice acquired the name price optimization (§35.7). In its benign-sounding framing, it was described as sophisticated demand modeling, retention management, and revenue optimization — the same toolkit used across retail and travel, finally arriving in insurance. Charge the loyal, inelastic customer a little more (they won't leave); charge the price-sensitive shopper a little less (you need to win them). To its proponents it was simply good business, powered by better data. To regulators, it was something else.

The insurance / underwriting issue

The issue is the one §35.2 makes central: in insurance, price differences between insureds are supposed to reflect cost (risk) differences. That is the principle that separates fair discrimination from unfair discrimination. Price optimization breaks it. Under price optimization, two customers with identical expected loss can be charged different premiums — not because one is a worse risk, but because one is predicted to tolerate a higher price. The premium now encodes willingness to pay rather than risk. The "same class and essentially the same hazard" are being charged differently, on a basis that has nothing to do with hazard at all.

Several distinct objections converge here, and they are worth separating because they show how many fairness principles a single practice can offend:

• It severs price from cost. This is the core §35.2 violation: the price difference reflects the insurer's market power over the customer, not the customer's risk. Many regulators concluded this is the definition of unfairly discriminatory — arbitrary sorting by something other than expected loss.
• It can have a regressive, proxy-like effect. Price sensitivity is not randomly distributed. The customers least able to shop effectively — those with less time, less financial sophistication, or fewer alternatives — may be precisely the ones the model identifies as inelastic and therefore chargeable more. Loyalty and inertia can correlate with vulnerability, so an "optimized" price can quietly fall hardest on those least equipped to escape it. This is a different harm from §35.3's protected-class proxy, but it rhymes with it: a neutral-sounding behavioral variable producing a socially troubling distribution.
• It is invisible to the customer. The policyholder cannot tell whether their increase reflects a worsening risk or merely the model's judgment that they won't notice. The information asymmetry that underwriting is supposed to correct (Chapter 1) is here turned around and used against the insured.

What it shows

Price optimization is the chapter's cleanest example of a practice that is profitable, data-driven, and legal-sounding, yet crosses a fairness line that has nothing to do with protected classes (§35.7). It matters precisely because it is not the redlining story. There is no racial map, no protected variable, no historical-injustice feedback loop. It is, on its surface, neutral demand modeling. And it is still unfairly discriminatory — because fairness in insurance pricing is not only about protected classes; it is also about the foundational requirement that price track risk. A model can be scrupulously colorblind and still violate fairness by pricing what a person will tolerate instead of what they cost.

This is the §35.7 lesson in its sharpest form: the line between underwriting and extraction. When an insurer prices the peril a customer brings, it is underwriting. When it prices the customer's resignation — their predicted unwillingness to fight back — it has stopped underwriting and started extracting, and the law has repeatedly said so. The case also shows why "the data supports it" and "it improves the combined ratio" are not sufficient justifications for a pricing practice (themes 3 and 4 of the book, held in tension with theme 6): a practice can do both and still be impermissible, because adequacy and profitability are necessary conditions for a legitimately priced book, not a license to price by anything that boosts the margin.

Outcome

Beginning in the mid-2010s, a series of state insurance regulators issued bulletins prohibiting or restricting price optimization in personal-lines rating, on the ground that it produces premiums not based on risk and therefore runs afoul of unfair-discrimination and rating-law standards. The response built into a substantial multistate pattern, with many states taking action over a relatively short period. The practical effect was to push insurers to demonstrate that their rates are grounded in cost-based factors and to disentangle legitimate retention and demand considerations from impermissible willingness-to-pay pricing.

As with most fairness developments, the boundary is not perfectly crisp and the debate continues at the edges: legitimate competitive pricing, lawful underwriting judgment, and prohibited price optimization are not always cleanly separable, and reasonable disputes remain about where exactly each begins. But the central holding is settled and widely adopted: an insurer may not set an individual's premium based on their predicted price sensitivity rather than their expected cost. Consistent with this book's discipline, the precise list of states and the exact terms of each bulletin should be read from current regulatory sources; the qualitative outcome — a broad regulatory rejection of price optimization in personal lines — is well attested.

Lesson

The lesson pairs with Case Study 1 to bracket the whole chapter. Case Study 1 (Colorado SB21-169) defends the fairness line on the protected-class / disparate-impact side — the danger that neutral data discriminates by proxy. This case defends the fairness line on the price-must-reflect-cost side — the danger that neutral data prices captivity rather than risk. Both are fairness failures; neither involves an explicitly prohibited variable; both can look like sophisticated, profitable, defensible data science. Together they teach the underwriter the most important habit in the subject: a practice's profitability, predictive power, and apparent neutrality tell you nothing about its fairness. You have to ask the harder questions every time — is the price tracking cost? is a neutral factor carrying protected information? — because the data, by itself, will never tell you when it has crossed a line it cannot see.

Discussion questions

1. Explain precisely why price optimization violates the §35.2 principle that price differences between insureds must reflect cost differences. How is this a different fairness failure from proxy discrimination (§35.3)?
2. The chapter says price optimization marks "the line between underwriting and extraction." Define that line in your own words, and give one example of a pricing decision on each side of it.
3. A defender argues that price optimization is no different from the demand-based pricing used by airlines and hotels, and that insurance should be allowed to do the same. Why do regulators treat insurance differently? (Consider insurance's social function and the cost-based-rating principle.)
4. Price optimization can fall hardest on customers who shop least — who may be the most vulnerable. Explain how a behavioral variable (price sensitivity) can produce a socially troubling distribution even though it is not a protected class. How does this "rhyme with" the proxy problem without being identical to it?
5. Both case studies show a profitable, data-driven, legal-sounding practice crossing a fairness line. State, in one sentence each, the two different fairness lines the two cases defend — and explain why an underwriter needs to watch both lines at once, not just the protected-class one.