Steenbergen: hyperscalers don't know their AI ROI

Steenbergen's remarks place Booking on the buyer side of the AI divide he describes: the company uses large language models rather than building them, which he suggested keeps its costs lower than those of the hyperscalers. That framing fits the internal numbers he cited — customer service costs slightly down while bookings rose at a high-single-digit rate, and engineering output up about 30% in code reaching production. The customer-facing side is far less developed: LLM referrals, paid and unpaid, were under 1% of total room nights as of the company's Q2 call in August.

Booking's answer to controlling costs is a routing system that sends simple tasks to basic and open-source models and complex ones to pricier models, with engineering groups judged on total IT cost per merge request rather than token spend alone. Steenbergen also described a shift from experimentation over the past two years toward rebuilding processes "from a white sheet." He said AI would change how travelers plan and book, not how travel itself happens, and pointed to the "connected trip" idea of managing disruptions across a journey.

Where this leads depends on whether Booking's modest early gains in its own apps — slightly faster booking, slightly higher conversion, slightly fewer cancellations — hold up as data accumulates. Steenbergen's own use of an "AI coach" and two agents suggests he expects leaders to demonstrate the tools they ask their organizations to adopt. For finance chiefs at companies buying rather than building models, his caution about costs going out of control is likely to land as the more actionable point.