Booking Holdings CFO: Even AI Hyperscalers Can't Compute Their ROI

Referrals from large language models still account for less than 1% of Booking Holdings' total room nights — and the company's CFO thinks even the companies spending hundreds of billions on AI can't tell you what their returns will be.

"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," Ewout Steenbergen said Thursday at Fortune's AIQ Summit at the New York Stock Exchange. "They have maybe some assumptions and some hypothesis, but the whole point is: you don't want to fall behind."

That uncertainty is good news for everyone below the hyperscalers, Steenbergen argued. Companies that use AI models rather than build them pay far less — provided they watch the bill. "The returns will be there for those processes that are being redesigned end to end," he told Fortune senior writer Sheryl Estrada, "but you also have to make sure that those costs on the other hand are not going out of control."

Booking, which owns Booking.com, Agoda, Priceline and OpenTable, has so far found its clearest gains inside the company, not on the customer side. Steenbergen said on the Q2 earnings call in August that LLM referrals, paid and unpaid, remained under 1% of room nights.

Watching the token bill

The past two years were about experimenting, Steenbergen said. The next phase means rebuilding processes "from a white sheet." Customer service went first. Bookings are growing at a high-single-digit rate while customer service costs are slightly down, so cost per booking has fallen "by a lot" — and customer satisfaction rose at the same time.

In engineering, Booking's roughly 9,000 engineers are getting about 30% more code into production, counting only merge requests that pass testing and quality control.

To keep AI spending in check, Booking uses what Steenbergen called "effective model cost routing": basic and open-source models handle simple tasks, while expensive models take complex ones. Engineering groups are measured on total IT cost per merge request, combining human and AI-token costs. Token spending can rise, he said, as long as the cost of each merge request reaching production comes down.

The hotel is still the hotel

AI won't change how people travel, only how they plan and book, Steenbergen said. "I think fundamentally the way how travel will happen in the future is not going to change with AI, but what is going to change is the experience for travelers," he said. "The hotel is still the hotel and the airline is still the airline and the rental car is still the rental car."

Travelers visit about five platforms on average before booking, he noted. AI could compress that research and then manage disruptions — a delayed flight that breaks a restaurant reservation, for instance. Booking calls this its "connected trip."

The commercial stakes are large. About one-third of Booking's customers arrive through paid channels such as search, social media and metasearch, and the company spends $8 billion to $9 billion a year on them. Booking runs constant A/B tests and large optimization models to allocate that money. The other two-thirds come directly, and Booking is building AI tools into its own apps to keep them.

The customer-facing AI gains remain modest: users of Booking's tools take a little less time to book, convert at slightly higher rates and cancel slightly less often. "It's very early stage, so it's not a lot of data," Steenbergen said.

From transactional to high-frequency

The bigger opportunity is growth, not cost. Answering an audience question about trip bundles, Steenbergen said Booking is "more transactional" today and AI could make it "a much more high frequency business." His example: a traveler has a Paris trip and Wednesday's forecast is bad. "We can proactively say, 'Hey, shall we change your schedule around on Wednesday? It's bad, so why don't you do the Louvre tour then and your walking tour on Thursday?" An active-sports enthusiast with an empty day could get a kite-surfing offer. That, he said, builds "more loyalty, more trust and more brand value."

Booking's Q2 numbers show early signs: Connected Trip transactions, where travelers book more than one travel vertical for the same trip, grew at a low-double-digit rate at Booking.com. Merchant bookings reached about 73% of gross bookings. Level 2 and Level 3 Genius members made up more than 30% of active customers.

Steenbergen is learning too. He has an "AI coach" and talks about it inside the company on purpose. He also runs two agents: a "strategic thought partner" that helps with board presentations and strategic plans, and a "critical equity research agent" that helps him prepare for earnings calls. "No one knows this," he said. "If I have to learn and I have a coach, it's very normal. Everyone has to learn."

The message to investors: Booking's AI returns today live in cost per booking and code output — the high-frequency travel assistant is still a hypothesis, and Steenbergen is willing to say so.