Unfair personalised pricing

What it is

Pricing models can estimate, for each customer, not only what it costs to serve them but how much they will tolerate paying. Optimising on that estimate raises prices for customers who are loyal, hurried, less able to compare or less likely to complain. Those traits correlate with age, vulnerability and other characteristics. The practice may involve no protected characteristic and no false statement. It is unfair because the price reflects the customer's weakness and not the value received. Where the differences track a protected characteristic it is also a discrimination risk, covered in proxy discrimination via correlated features.

How it happens

• The model is optimised for margin, with willingness to pay as an input or a learned feature.
• Behavioural signals, such as time of day or device, feed the price.
• Renewal prices rise for customers predicted not to shop around.
• Nobody reviews the distribution of prices across customer groups.

Why it matters

Regulated firms must be able to show that the price a customer pays is reasonable relative to the benefits, for each group of customers. Charging existing customers more than new ones for the same thing is restricted in some markets. Discovery tends to be public and damaging.

Warning signs

• Price differences between similar customers cannot be explained by cost or risk.
• Features describing customer behaviour, not risk, are strong drivers of price.
• Longer-standing or older customers pay more.
• Fair value assessments do not examine the model's outputs.

Who owns it

The pricing owner is accountable. Product governance assesses fair value. Compliance and the second line test the outcomes by customer group.

In practice

A home insurer's renewal model learns that customers who have renewed three times without calling will accept larger increases. Their premiums drift well above those quoted to new customers with the same risk. The same pattern generalises to any pricing model free to learn who will not object.