Five conditions decide which side of that line your business sits on.
Two passengers sitting next to each other on a flight may have paid fares that differ by a factor of three. Neither is surprised. Neither writes an angry post about it.
In February 2024, the CEO of Wendy’s mentioned on an earnings call that the chain would test “dynamic pricing” on digital menu boards. Within days, the company was in a public relations crisis, stating that it would not implement surge pricing and that it had never used the phrase.
Same underlying technology. Completely different reception. Understanding why is now a board-level question, because AI pricing tools are being sold aggressively to Indian companies that have none of the conditions that make airline pricing work.
What airlines actually have.
Dynamic pricing is not a technology. It is a set of market conditions that happen to permit a technology. Airlines have five of them.
The inventory perishes absolutely. An unsold seat at departure is worth zero, forever. This is the original reason yield management was invented, and it applies to hotel rooms, cinema tickets and concert seats — not to a burger, a bag of cement or an annual software licence.
Capacity is fixed in the short run. The airline cannot add a row of seats because demand spiked this morning.
The price difference has a visible reason that is not the customer. You paid more because you booked late, chose a refundable ticket, or wanted the 7 a.m. departure. The passenger can point to the choice that produced the price. That is the single most important condition on this list.
The practice is industry-wide. No competitor can attack you for it, because all of them do it.
Marginal cost is near zero. Filling the last seat at a low fare still adds contribution.
Even with all five conditions in place, airlines remain one of the most resented categories in consumer life. Wharton’s faculty commentary on dynamic pricing makes this point directly: in the very industry where customers have had thirty years to adjust, variable pricing still fuels perceptions of unfairness. That should tell you something about what happens when the conditions are weaker.
What AI changes, and where the line sits.
Through 2025, Delta expanded AI-assisted pricing built with the Israeli firm Fetcherr. Company leadership told investors in July 2025 that the system was setting prices for about 3% of the domestic schedule, with a target of 20% by the end of that year. Fetcherr’s technology is also used by Virgin Atlantic, WestJet, Azul and Viva Aerobús.
Three US senators wrote to Delta warning that expanding AI pricing would likely mean fares rising to each customer’s personal “pain point.” Delta’s response was unequivocal: there is no fare product it uses or plans to use that targets customers with individualised offers based on personal information, and all customers see identical prices. Fetcherr likewise said its technology does not allow individualised or personalised pricing.
It is worth noting that travel journalists reported Fetcherr had earlier published marketing material describing hyper-personalisation using purchase history and customer lifetime value, which was subsequently edited. Delta and Fetcherr deny individualised pricing; that denial is the company’s position, and I am reporting it as such.
But look at where the argument landed. Delta did not defend personalised pricing. It denied doing it. Both sides accepted, without discussion, that pricing by condition is legitimate and pricing by person is not.
That is the line. Everything useful about AI pricing sits on one side of it, and almost everything dangerous on the other.
Why it backfires everywhere else.
Coca-Cola learned this in 1999, when its then-CEO mused about vending machines that would raise prices in hot weather. The backlash was immediate, and Pepsi was happy to call it exploitation. The episode is still taught as a case study.
Wendy’s learned it in 2024, as above.
A 2023 Capterra survey of 900 US restaurant customers found that more than half equated dynamic pricing with price gouging, and 65% said it would make choosing where to eat harder. Note the second number: the damage is not only anger; it is decision friction. A customer who cannot predict your price postpones the purchase.
Here is the part most pricing decks leave out. The real cost of a dynamic pricing misstep is retrospective. When customers discover prices vary by person, they do not merely object to today’s price. They reinterpret every purchase they have already made, and wonder whether they were the one being charged more. You cannot refund your way out of that.
There is a constructive version of the same insight, and Wharton’s commentators make it well: Coca-Cola could have set the standard price at the hot-weather level and offered cold-weather discounts. Identical economics, opposite reception. Dynamic discounting is usually welcomed; dynamic surcharging is usually punished.
The Indian version of this problem.
In January 2025, the Central Consumer Protection Authority issued notices to Ola and Uber after reports that identical rides were priced differently for iPhone and Android users.
The Union Consumer Affairs Minister called it prima facie an unfair trade practice and said the examination would extend to other sectors, including food delivery and online ticketing. Uber denied setting prices by phone model.
Note what triggered that intervention. Not surge pricing during rain, which Indian consumers grumble about but accept. It was pricing that appeared to vary by a proxy for who the customer is — and specifically, a proxy for how much they can afford.
For Indian businesses, that is the operative distinction. Vary price by time, scarcity, quantity or commitment, and you are on familiar ground. Vary it by device, location, browsing history or inferred wealth, and you are in regulatory and reputational territory regardless of what your model’s accuracy numbers say.
My Personal observation: If you show your desperation to purchase a particular item/model (say, a specific mobile model), amazon.in will inflate the price deliberately on your subsequent visits. Here, Amazon is determining the price for me based on my browsing history!
The regulation has already arrived — elsewhere.
Most Indian executives have not tracked how quickly this moved abroad in the last year. New York’s Algorithmic Pricing Disclosure Act took effect in November 2025. It requires any business setting prices for New York consumers using “personalised algorithmic pricing” to display a specific disclosure stating that the price was set by an algorithm using the customer’s personal data. Reports put the penalty at $1,000 per violation.
In January 2026, the California Attorney General announced an investigative sweep into surveillance pricing under the state’s privacy law. California’s AB 325, effective 1 January 2026, addresses shared pricing algorithms that use competitor data. Connecticut, Maryland and New Jersey have enacted restrictions of their own, and the New York legislature passed a broader bill in June 2026.
Separately, the US Department of Justice settled with RealPage in November 2025 over its rent-pricing software, which generated recommendations using non-public data from competing landlords. The compliance principle that emerged is worth memorising: using an algorithm does not protect from liability.
None of this binds an Indian company serving Indian customers. But if you sell to US consumers, it binds you now.
And for everyone else, it is a reliable signal of where India’s own consumer-protection and DPDP enforcement is likely to travel.
Five questions before you let a model set your prices
- Does my inventory actually perish? If unsold stock retains value, you are solving a markdown problem, not a yield problem.
- Can the customer name the reason for the difference? If the honest answer is “because the system thinks you’ll pay it,” stop.
- Would I print the pricing logic on the invoice? In New York this is no longer hypothetical. It is a decent ethical test everywhere.
- Does the price move in both directions? A system that only ever goes up is a surcharge engine, and customers will name it correctly.
- Is the data feeding my model mine alone? Pooled competitor data in a shared pricing tool is where pricing software becomes an antitrust question.
For most businesses, the productive path is narrower than vendors suggest:
- Markdown optimisation for genuinely perishable stock.
- Segment-level pricing that customers can self-select into.
- Personalised discounts rather than personalised surcharges.
- Quote-guidance tools for B2B sales with a human signing every number.
The caveats.
The evidence here is mostly cases, surveys and regulatory action rather than controlled experiments, and the famous failures are better documented than the quiet successes. The Capterra survey is US restaurant-specific. The Delta and Fetcherr positions are contested, and I have given their denials in their own terms.
Reasonable people argue the other side: the economist Tim Harford has made the case that we suffer from too little dynamic pricing, not too much, because price signals ration scarce capacity better than queues do. On electricity tariffs, I think he is plainly right.
My question for you: has your organisation been pitched an AI pricing tool in the last year — and when you asked what variable was actually driving the price difference, did you get a clear answer? I suspect the quality of that answer predicts almost everything.
Every claim in this issue is traced to a named, dated source. Where a company disputes an allegation, its denial is stated. A full source list is available on request.
— The AI Compass.
