Booking Holdings CFO says even AI hyperscalers don't know their ROI — but he's learning from his 'AI coach'

Booking Holdings’ finance chief thinks the loudest AI spenders are still making educated guesses. Ewout Steenbergen said at Fortune’s AIQ Summit at the New York Stock Exchange that even the hyperscalers pouring hundreds of billions into large language models don’t really know the ROI. They have assumptions, he said, but the real motivation is simple: nobody wants to fall behind.

That caution matters for a company like Booking because it changes the frame. If you’re not trying to build the models, you can use them more cheaply — as long as the token bill doesn’t run away. Steenbergen said the payoff comes when processes are redesigned end to end, not when AI is bolted on like a shiny sticker.

Booking’s biggest wins so far are inside the company, not in customer referrals. On the company’s Q2 call in August, Steenbergen said referrals from large language models were still under 1% of total room nights. But customer service is already showing cleaner economics: bookings are up at a high-single-digit rate, customer service costs are slightly down, and the cost per booking has fallen a lot while satisfaction has improved. In engineering, Booking’s roughly 9,000 engineers are pushing about 30% more code into production, measured only after merge requests pass testing and quality control.

The company is managing AI spend with what Steenbergen called effective model cost routing. Simple tasks go to basic or open-source models; harder ones get the pricier systems. Engineering teams are judged on total IT cost per merge request, which includes both human work and AI token costs. The rule is blunt: token spending can rise if the cost of getting code into production falls.

Steenbergen’s bigger ambition is growth. He said travel itself won’t change — the hotel is still the hotel, the airline is still the airline — but the experience around it can. Booking says travelers visit about five platforms on average before booking, and AI could reduce that search, then help manage disruptions like a delayed flight that ruins a restaurant reservation. The company already spends $8 billion to $9 billion a year on paid channels, while the other two-thirds of customers arrive directly, so keeping people inside Booking’s own apps matters a lot.

The early signs are modest but real. Customers using Booking’s AI tools take a little less time to book, convert slightly better and cancel slightly less often. Connected Trip transactions grew at a low-double-digit rate at Booking.com in Q2, merchant bookings reached about 73% of gross bookings, and Level 2 and Level 3 Genius members topped 30% of active customers. Steenbergen says he’s learning the same way everyone else is: with an AI coach, plus two agents of his own, one for board decks and one for earnings-call prep.