Over the summer, Israeli tech company Fetcherr found itself in the middle of a public controversy after remarks made by Delta president Glen Hauenstein led to accusations that the airline was using Fetcherr's AI-powered pricing technology to target customers with individualized prices based on personal data. Delta sharply denied those claims. Airlines editor Robert Silk spoke with Fetcherr chief AI officer Uri Yerushalmi to learn more about Fetcherr's pricing engine.
Uri Yerushalmi
Q: Can you explain how your AI engine analyzes and suggests airfares for airlines?
A: It's a Large Market Model (LMM). It's based on the same techniques as other GenAI models like ChatGPT, but it is not trained on textual data like Large Language Models, and it's not trained on videos or pictures. Instead, it is trained on aggregated and public information of the airline, and that allows it to be a super analyst and to make actual decisions. It is done in a way that it can predict very granularly on the product side. Things like seats and departure date, but not on the customer side, meaning we are focusing on price predictions and optimizations of the product, not of the customer. The product can be a certain departure date in a given cabin; what would be the demand at a certain price point. It enables us to simulate what would happen with every price point, and that enables the airlines to pinpoint the price that would eventually fill the airplane and get to very high load factors, which is very difficult to do without such a predictive system.
Q: Airlines do want to push out personalized offers, though. Some are doing it. That doesn't mean that they price discriminate based on personal preferences or demographics. But they do aspire to tailor offers to customers' tastes. Does Fetcherr assist with those types of capabilities?
A: Absolutely not. No information about the specific customer is being touched or leaked into the Fetcherr system. We are very strict about that. We do, however, differentiate different departure dates. Customers who have longer stay dates might be priced differently than customers who stay for a shorter time. But that's based on the ticket, not based on the customer.
Q: How many airlines are you working with?
A: The list includes Delta, Virgin Atlantic, Azul, Jet2, WestJet and also some airlines that are not published yet. The airlines have the capability to dictate different restrictions to the model -- meaning, don't go to certain areas of prices or having some relationship between the hierarchy of prices. So, the airlines have a wide flexibility to tell the model to search for different things, but the core system is the same.
Q: What are some other data points the Fetcherr engine can consider more comprehensively or efficiently than humans or systems without GenAI capabilities?
A: It considers whether there are any events in the destination you are traveling to. We also look at some public data; for example the conditions in the market and weather information. All that is fed into the LMM. Decisions can regard price, inventory, flight schedules and ancillary and bundling policies. All of these can be made based on the model's predictions.
Q: The LMM helps airlines optimize revenue by more accurately forecasting demand. But do passengers end up paying more?
A: One way for an airline to increase revenue is by increasing load factor. Increasing the load factor usually means reducing the price. On average, we are seeing airlines reduce prices more to get to higher load factors and make more revenue. The system encourages competition because it notes much faster what is happening in the market. Eventually, encouraging competition reduces prices.