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

Booking Holdings says AI is reducing internal operating costs and improving customer-service efficiency: bookings are growing at a high-single-digit rate while customer-service costs are slightly lower, materially reducing cost per booking. Its roughly 9,000 engineers are deploying about 30% more code that passes quality controls, supported by model routing designed to lower total IT cost per production merge request. Customer-facing AI remains early stage—LLM referrals accounted for under 1% of room nights—but users show slightly faster booking, higher conversion and lower cancellation rates; connected-trip transactions grew at a low-double-digit rate in Q2.

Analysis

BKNG’s relevant AI edge is not consumer-facing novelty but the ability to convert fixed service and engineering expense into operating leverage while protecting conversion quality. If management can sustain lower cost-to-serve alongside higher booking completion and fewer cancellations, incremental gross profit should fall through at an unusually high rate given the platform’s mature cost base. The key KPI is therefore not AI referral volume; it is quarterly movement in customer-service expense per transaction, marketing efficiency, cancellation rates, and adjusted EBITDA margin.

The more consequential medium-term question is whether AI changes distribution economics. Better in-app trip management can raise direct repeat usage and reduce reliance on auction-priced traffic, pressuring Google’s travel-query monetization at the margin; however, AI assistants built into Google Search, ChatGPT, or supplier-owned apps could equally remove the OTA from discovery. BKNG has supplier breadth, payments infrastructure, and post-booking servicing advantages, but those advantages only matter if its interfaces retain the customer relationship when conversational search becomes the front door.

A 1-3 month re-rating is unlikely from qualitative commentary alone: investors need evidence that productivity gains are translating into margin guidance rather than being reinvested into promotion or product development. Over 6-18 months, the upside case is higher ancillary attach and loyalty-driven purchase frequency, which would justify multiple durability; the downside is that AI lowers comparison friction and makes hotel inventory more substitutable, intensifying take-rate competition. Falsify the constructive view if direct mix stalls, paid marketing as a percentage of gross bookings rises, or service-cost savings fail to appear despite continued AI investment.

Key Decisions for Investors

• Maintain or initiate a modest long BKNG only on evidence of margin conversion at the next two earnings reports: add if adjusted EBITDA margin expands while room-night growth remains stable. Target a 8-12% relative upside versus travel peers over 6-12 months; exit if management guides to rising marketing intensity without a corresponding acceleration in bookings.
• Express execution differentiation as long BKNG / short EXPE over 3-6 months, sized market-neutral. BKNG’s scale and merchant/payment ecosystem should monetize automation and cross-sell earlier, while EXPE faces greater need to prove that AI product investment improves unit economics; cover if EXPE demonstrates faster direct-booking growth or materially greater margin expansion.
• Do not position against GOOGL solely on this development. Set an alert for BKNG direct-channel growth, paid-acquisition efficiency, and any disclosed traffic shift from AI search products; a measurable deterioration in paid-search conversion or referral terms would turn BKNG from an AI beneficiary into a distribution-risk short candidate.
• Watch merchant penetration and multi-product transaction growth over the next 2-4 quarters as the cleanest validation of the higher-frequency thesis. If these measures accelerate without higher incentives, increase BKNG exposure; if they require heavier discounting, treat the revenue opportunity as low-quality and avoid multiple expansion assumptions.