A price is not a number. It is a statement about a relationship, and customers read it as one.
This is the part that pricing engines were never designed to handle. Technically, charging every customer a different amount is a solved problem: the data exists, the models work, and in most jurisdictions the practice is legal provided it does not discriminate on protected characteristics. Airlines have done it for decades. Retailers do it now. The infrastructure is not the obstacle.
The obstacle is that people find out.
Two different questions
Economics asks what a customer is willing to pay. It is a good question, and price discrimination is the rational answer: capture more of the surplus, serve segments that a single price would exclude, fund the fixed costs that make the product possible at all. On these terms, personalised pricing is not merely defensible but efficient.
Customers ask a different question. Not what is this worth to me, but what does this company think of me. And a price that varies by person answers that second question in a way no amount of efficiency can soften. The customer who pays more does not conclude that the algorithm has correctly estimated their willingness to pay. They conclude that they were the one who did not notice.
The two questions are not reconcilable by better modelling, because they are not about the same thing. One is about value. The other is about respect.
Where the line actually falls
It is tempting to conclude that price variation is simply unacceptable, but that is plainly wrong: student discounts, off-peak fares, loyalty rates and regional pricing all vary the price by person and provoke no outrage at all. Something distinguishes them from the algorithmic case, and it is worth naming precisely, because it is the whole of the practical guidance.
Accepted price variation is legible. The customer can see the rule, can see why it exists, and — crucially — can see how to qualify for the better price if they want it. A student discount says: we would like students to be able to afford this. An off-peak fare says: help us fill the empty trains and we will share the saving. Each is an offer with terms, and the terms are public.
Algorithmic personalisation inverts every one of these properties. The rule is invisible, the reason is unstated, and there is no path from the worse price to the better one, because the customer does not know what put them on the wrong side of it. What they experience is not a discount they failed to claim. It is a penalty they cannot appeal.
The asymmetry that makes it unstable
There is a further problem, and it is structural rather than ethical. Personalised prices are discovered asymmetrically. The customer who was charged less has no reason to mention it. The customer who was charged more has every reason to, and now has the tools: a screenshot, a comparison, a social account. The discovery rate is low but the consequence, when it happens, is public and one-directional.
This means the expected cost of the practice is chronically underestimated. The gains are small, certain and continuous. The losses are rare, large and reputational — and they arrive in a form that no pricing model measures, because the damage lands on brand rather than on margin.
What to do instead
The useful move is not to abandon differential pricing but to make it earn its way into the open. Three tests are worth applying before any personalised price ships.
Would you publish the rule? Not the model, the rule. If the sentence explaining why this customer pays more cannot be written on the pricing page without embarrassment, the practice will not survive discovery.
Can the customer act on it? A price that varies by behaviour the customer can change is an incentive. A price that varies by an attribute they cannot change is a tax.
Does the variation track cost or willingness? Customers accept paying more when serving them costs more. They do not accept paying more because a model inferred that they would.
None of this is an argument for a single price. It is an argument that price is a communication before it is a number, and that a pricing strategy which optimises the number while ignoring the message is optimising the smaller of the two variables.