Booking's CFO says AI pays off for companies that use models and watch the bill

"I even think that the hyperscalers that spend hundreds of billions on the development of their large language models, they don't really know what is going to be the ROI," Steenbergen said at Fortune's AIQ Summit at the New York Stock Exchange ^[1]. His condition for everyone else is narrower. "The returns will be there for those processes that are being redesigned end to end," he said, "but you also have to make sure that those costs on the other hand are not going out of control." ^[3]

The work Steenbergen described at Booking is choosing models and metering them. Basic and open-source models handle simple tasks and more expensive models take the complex ones, a practice he called "effective model cost routing" ^[7]. Engineering groups are then measured on total IT cost per merge request, a number that adds human and AI-token costs together ^[8]. Token spending can go up, he said, as long as the cost of each merge request reaching production comes down ^[8].

The customer side rests on one number. On the August call for the second quarter, Steenbergen said referrals from large language models, paid and unpaid, were under 1% of total room nights ^[4]. Paid search, social media and metasearch sites bring about a third of customers, and Booking spends $8 billion to $9 billion a year on them ^[9]. On those two measures the paid channels are more than 33 times the size of the chatbot channel, although customers and room nights are not the same unit ^[2]. Booking decides where the paid money goes with constant A/B tests and large optimization models ^[10].

The under-1% figure can move three ways. Chatbots can stay a small source of room nights, leaving the internal savings as most of what AI returns to Booking for now ^[4]. They can grow as a paid channel and add another seller of traffic to a marketing bill that Booking is telling investors AI will make cheaper ^[19][4]. Or assistants can book hotels directly, and the referral figure would never register it, because it counts only referrals that end in a Booking room night ^[4].

Travelers visit about five platforms on average before booking, Steenbergen said, and AI could cut that research ^[14]. His case against the third outcome is about supply. "The hotel is still the hotel and the airline is still the airline and the rental car is still the rental car," he said ^[15]. It does not settle who keeps the customer if five platforms of research shrink to one assistant ^[14].

I think the evidence so far supports the cost half of Steenbergen's case and leaves the revenue half open. The counter-case is his own. He said the bigger opportunity is growth, and that AI could turn a "more transactional" Booking into "a much more high frequency business" ^[16]. Of the in-app tools that would have to deliver it, he said, "It's very early stage, so it's not a lot of data." ^[13] If customers using those tools start booking trips more often, the revenue half is proven and the cost-only reading is wrong.