The pricing question changes when willingness to pay is created before trade
Most pricing models treat willingness to pay as something the seller discovers. Nguyen and Tan's paper, Profiling and Endogenous Valuation, changes that premise. In their setup, willingness to pay is partly made by the buyer before trade happens.
That is common in software, enterprise platforms, devices, cloud services, and data products. A customer learns the workflow, migrates data, trains staff, builds integrations, or adapts internal processes. Those actions make the vendor's product more valuable to the customer. They also make the customer easier to charge later.
That is the paper's central tension. Profiling helps the seller identify who is likely to value the product highly. But if the buyer expects the seller to use that information to raise the price after the buyer has sunk effort or money, the buyer may not make the investment in the first place.
This is not the usual price discrimination story. In the classic Bergemann, Brooks, and Morris (2015) framework, buyer values are fixed, and information mainly affects who gets the surplus. Here, information also affects whether the surplus exists.
The model: a monopolist, a buyer investment, and a profiling signal
The model is deliberately spare. A monopolist sells one unit of a product. The buyer starts with value L. Before meeting the seller, the buyer can pay a private cost c to raise his value to H. The gain from investing is H minus L, and the authors assume investment is socially efficient for every buyer type.
The seller does not observe the investment itself. She does observe a signal about the buyer's cost c before setting price. That signal is the profiling technology. It can be fine or coarse. It can separate some types and pool others. The paper asks which buyer and seller payoffs can arise across all possible profiling structures.
The timing matters. The profiling structure is known, the buyer sees his own cost and chooses whether to invest, then the seller sees the signal and posts a take-it-or-leave-it price. Since the buyer's value is either L or H, the relevant posted prices are L and H.
A low price gives an investing buyer some surplus, because he pays L for something worth H. A high price extracts the value created by investment. The buyer invests only if he expects enough chance of facing the low price after investing. That expectation is pinned down by the seller's signal and pricing rule.
One immediate result is harsh: the first-best surplus is not attainable. If every type invested, the seller would optimally charge H. Seeing that coming, no type would want to invest.
The new constraint: hold-up risk limits both surplus creation and surplus sharing
The key result is a new constraint on feasible outcomes. The seller can always ignore her signal and charge H. When many buyers invest, that outside option becomes more attractive. That is the hold-up risk.
To sustain investment, the equilibrium must give the seller at least as much as she could get by ignoring the signal and charging H. This creates an upper bound on investment. More investment produces more high-value buyers, but it also increases the seller's temptation to capture the investment after the fact.
This is where the model departs sharply from fixed-valuation price discrimination. With fixed values, information can move surplus between buyer and seller while leaving total surplus alone. With endogenous value, surplus creation and surplus division are tied together.
Nguyen and Tan show that on the Pareto frontier, if investment rises, the seller's required payoff rises faster than the extra surplus created by the investment. That is the surprising comparative statics result. The seller must be compensated for not holding up the buyer, and that compensation grows quickly as more buyers invest.
So protecting buyer welfare can require less investment, even though investment is efficient type by type. That sounds perverse, but the mechanism is clean. To give buyers more payoff, the market must reduce the seller's temptation to charge H. One way to do that is to reduce the mass of buyers who invest. Less created surplus can make more buyer surplus feasible.
Why better profiling can make the seller worse off
The paper also shows that sharper profiling need not help the seller. Perfect information about the buyer's cost gives the seller her lowest individually rational payoff.
The reason is commitment. The seller would like buyers to invest first and then pay a high price. Buyers see this coming. If the signal tells the seller exactly which types have low investment costs, those buyers expect to be priced against. Their investment incentive falls. The seller becomes better at extracting surplus that no longer gets created.
Coarser information can help the seller because it softens this pricing problem. If the seller cannot fully separate low-cost, likely-to-invest buyers from others, she may sometimes charge L. That leaves enough rent for investment to happen. The seller may then earn more from a larger pie, even with less precise extraction.
The authors' dual-cutoff construction makes this point more concrete. Information about high-cost types, who are unlikely to invest, can help both sides because it prevents inefficient trade breakdown and expands the surplus that can be sustained. Information about low-cost types is more distributive. Those are the customers who can create a lot of value. Profiling them too well tells the seller where the surplus is, which can destroy the reason to create it.
For practitioners, the distinction is not "data good" or "data bad." It is which data, about which customers, used at which point in the commercial relationship.
Why we could not backtest this on our data
We could not backtest the paper in a clean way. The model is about counterfactual profiling structures and pre-trade investment incentives. Our transaction data show prices, purchases, and some customer attributes. They do not show each buyer's private cost of learning, migrating, integrating, or adapting before purchase.
A real test would need exogenous variation in what the seller knows before pricing, plus measures of buyer investment that occur before trade. It would also need the prices that would have been offered under different profiling rules. We do not observe those counterfactuals. Standard panel variation would mix the profiling effect with product changes, sales effort, budget cycles, and account selection.
That limitation does not weaken the paper's contribution. It means the contribution is conceptual rather than a ready-made empirical signal.
What this implies for platforms, vendors, and data-policy design
The most relevant use case is any product where the customer must sink effort before the product becomes valuable. Think enterprise SaaS implementation, cloud migration, payments integration, developer platforms, analytics tooling, or hardware ecosystems. In these markets, customer success work and pricing policy are linked. If the customer believes adoption will be used against them at renewal, adoption slows.
For vendors, the model points to a practical pricing problem. Fine customer scoring can raise price extraction, but it can also make customers reluctant to invest in the relationship. Commitment devices may matter here: long contracts, renewal caps, posted price schedules, or limits on account-level price discretion. The paper does not model those tools directly. They are natural ways a seller might try to make investment feel safer for the buyer.
For platforms, the result speaks to data governance. A platform may want detailed data to improve matching, support, and onboarding. But using the same data for personalized price increases can poison the investment channel. Separating data used for support from data used for pricing is not just a privacy stance. It may help preserve demand creation.
For policy design, the paper argues against blanket claims about profiling. Restricting all information can destroy useful segmentation for high-cost customers. Allowing all information can make hold-up worse for low-cost customers who are ready to invest. The better question is narrower: does this information help create surplus, or does it mainly identify sunk customer effort after the fact?
The model is stylized. It has one buyer, one seller, binary investment, and no competition or reputation. Real vendors face churn, procurement teams, reference customers, and sales capacity constraints. Still, the mechanism is easy to recognize. A vendor that wants adoption may need to avoid profiling the customers most ready to adopt, or make a credible promise not to price against that readiness later.