Sit in enough pricing meetings and you learn to recognise the moment a room stops thinking. It arrives with a number that sounds like knowledge. Elasticity for the segment is minus something. A point of premium costs us this many renewals. Everyone writes it down, the discussion moves to how much of the point to take, and nobody asks which of these people it is actually true of.
The answer, almost always, is none of them. An elasticity figure attached to a segment is an average over a group assembled for reporting reasons. Age band, product, channel, tenure bucket. People land in the same cell of that grid while wanting opposite things from you, and the number that comes out the other end is the arithmetic residue of their disagreement.
I keep returning to a mapping exercise I worked on with a European composite insurance group that was building a group-level analytics capability across its operating companies. The job was to lay out, on one sheet, where analytics could move the profit and loss: revenue on one arm with product volume, pricing, marketing and distribution effectiveness under it, cost and claims on another, risk on a third. Eight candidate cases were placed against that structure, from fraud and connected-home data through next best action, churn and process work on the claims value chain.
Three of the eight reach for the individual, and they reach for different things. The connected-home case wants individual behaviour as a risk signal: how often the house stands empty, whether the alarm works. The next best action case wants the individual’s position in a life cycle, so that the offer arrives when the need does. Only the pricing case makes the customer’s own response to price the quantity being modelled, and its wording is worth reading twice: understanding price sensitivity on an individual level, so that price-sensitive customers can be offered the best value for money and less price-sensitive customers can be offered the convenience and service they are actually willing to pay for. Not a better segmentation. A different quantity, measured one person at a time.
The average was hiding the opposite sign
The most useful thing in the whole file is not the pricing page. It is a retention model on an older track record page, built as a prototype with neural network technology on a deliberately small subset of about fifteen variables, trained on eighty percent of one year of data and tested on the remaining twenty. It predicted whether a business customer would cancel one policy, two policies, or none within ninety days. The reported result was 83.3 percent exact predictions across those three categories, with the 16.7 percent of errors being predictions of cancellation that did not happen, which is the harmless direction to be wrong in.
Take those figures as the sales document's own account of its work. The part I care about is not the accuracy, but the two findings recorded underneath it.
The first: loyalty improved with the number of account-management visits, up to a point, after which more visits reduced it. The second, and the one I have quoted in client rooms ever since: some customers should not be touched, because waking a quiet customer by contacting them about a long-forgotten contract raised the chance they cancelled.
Read those as pricing statements, because that is what they are. Contact is a price you charge in attention, and the response to it changes sign from person to person. A group average reports a mild positive effect and hides both the customers who wanted more of it and the customers for whom any of it was a loss. Run the campaign on the average and you spend money destroying value in a subset you never identified, while under-serving a subset who would have paid for more.
Price elasticity behaves the same way. The sensitive and the insensitive are not two ends of one distribution that a mean fairly represents. They are different populations wanting different things, sharing a cell in your reporting grid.
What makes a segment a reporting convenience
The segment did not appear because anyone believed people in it were alike. It appeared because the data could not carry anything finer, and then it stayed after that stopped being true.
Two other engagements in the same file mark the transition. One built a life-cycle model of a customer base: which products are bought in which order, and what other events, such as unpaid bills, occur along the way. It ran on a two-terabyte database, and the claim made for it is that it could be cut by segment, by area, by life-cycle stage, and down to the level of one customer. The commercial claims attached to it, again the firm's own account, are that the conversion strategy changed in a way that prevented more than twenty percent of a revenue loss, and that better customer segmentation produced an estimated improvement of more than forty percent on cross-selling.
The other is closer to pricing than anything in the pricing case. A large financial organisation was folding one of its separately operating labels into its central brand, which meant repricing and reducing a product range, with customers free to refuse the new terms and cancel. The model built for that decision projected pricing and dislocation impact on individual policy characteristics, along with assumptions about how each customer would behave in their own situation. The recorded outcome is that the original plans changed substantially once the analysis came back.
That is the whole argument in one engagement. When the decision was expensive enough that being wrong per customer mattered, nobody used a segment. They went to the policy.
- 83.3%
- Exact predictions across three cancellation categories, retention prototype
- ~15
- Variables in the subset that model was built on
- 90 days
- Cancellation window the model predicted over
- 1 customer
- Deepest level the life-cycle model could be cut to
Money on both sides of the average
A segment-level price decision loses money in two directions at once, which is why it rarely shows up as a loss.
On one side sit the customers less sensitive than the average. They stay through the increase you did not take, and they would have paid for cover, convenience or service the segment view never offered them, because that view had already decided the group was price-led. Nothing in the ledger records a premium you never charged or a product you never proposed.
On the other side sit the customers more sensitive than the average. They leave over an increase the segment number said they would tolerate, or they take a renewal discount they were never going to need. The second is the easiest to miss, because a retained customer counts as a win regardless of what the retention cost.
Both errors are the same error, and averaging is what produces them. The pricing case in that file is trying to fix both with a single move: sensitivity per person, then an offer built to what that person values. Value for money for one, more than the bare basics for another, at the same moment in the same renewal conversation.
There is an honest caveat, and the file states it by omission. The claims value chain case on the same sheet carries a hard number: fewer process steps and savings of roughly twenty-five percent, worth 3.3 million euros, at a customer support centre, found with process mining. The pricing case carries no number at all. It carries a direction, that a better read on when price is the real objection leads to more successful renewal negotiations. Directional is what the evidence supports, and directional is what I will claim.
An elasticity curve for a segment is a curve that nobody in the segment has. It is the shape left over after you add up people who wanted different things and divide by how many of them there were.
What has to be true before you can price a person
Per-customer pricing is not a modelling problem. The modelling end is close to routine now, and the first cautious language-model pilots crossing my desk absorb more attention than the whole pricing stack. The work sits underneath.
One identity per customer across products and channels, so that the same person holding motor and household is one row rather than two. Behaviour recorded as events with dates, not as a state at month end, since the whole signal here is in sequence: which product came first, what happened before the quiet period, when the unpaid bill landed. Features defined once and stored where the training job and the quoting system read the same definition, or the score in the model and the score at the point of sale drift apart within a quarter. Lineage on every input, so that a questioned price can be traced. A path from the model into the renewal flow, because a per-customer score delivered as a monthly spreadsheet becomes a segment again the moment somebody groups it to make it usable. And a written rule on which attributes may touch price at all, settled by the business rather than discovered later in an audit.
That block of work is the opening of a RealAI Platform engagement, and it is the reason we ask which decision you want to move before we ask which model you want to build. Answer it for pricing and the segment does not disappear. It goes back to being what it always should have been: a way of summarising a decision for a board pack, taken after the decision was made one customer at a time.
Drawn from a competitive proposal for an analytics mapping project at a European composite insurance group, and from the earlier insurance engagements recorded in its appendix. The candidate cases in that proposal are proposals: they were scoped and argued, not delivered. The retention, life-cycle and process figures are the proposing firm's own account of prior work, quoted as they appear. The reading of price sensitivity as a per-customer quantity rather than a segment average is ours.
“An elasticity curve for a segment is a curve that nobody in the segment has. It is the shape left over after you add up people who wanted different things and divide by how many of them there were.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production: sovereign models, governed data, and agents you can audit. Start with a value-first assessment.