Booking Holdings Deploys AI Coach as Hyperscalers Struggle to Measure Return on Intelligence Spending

The most important thing Booking Holdings CFO Ewout Steenbergen said on October 1, 2026 was not a number. It was an admission: even the companies spending hundreds of billions of dollars building large language models do not know what return they will generate.

Speaking at the Fortune AIQ Summit at the New York Stock Exchange, Steenbergen told Fortune senior writer Sheryl Estrada that hyperscalers deploying capital at scale into AI infrastructure are operating on assumptions and hypotheses, not proven economics. The motivation, he said, is fear of falling behind, not calculated return on investment . For institutional investors trying to price AI exposure across their portfolios, that framing resets the entire valuation conversation.

Booking Holdings operates Booking.com, Agoda, Priceline, and OpenTable, giving Steenbergen visibility into AI deployment across a multi-brand, multi-vertical travel platform serving hundreds of millions of customers. His perspective is not theoretical. The company spends between $8 billion and $9 billion annually on paid customer acquisition channels including search, social media, and metasearch sites . Any shift in how AI routes, converts, or retains those customers moves the needle on one of the largest marketing budgets in consumer internet.

"The returns will be there for those processes that are being redesigned end to end," Steenbergen said, adding that companies must ensure costs "are not going out of control" . The caution matters as much as the optimism.

The Plocamium interpretation: Steenbergen is describing a bifurcated AI economy. Model builders carry the capital risk and the uncertainty. Model users, if disciplined about cost architecture, capture the productivity gains at a fraction of the infrastructure cost. The investment thesis is not symmetric.

Where Booking Is Actually Seeing Returns: Inside the Cost Stack

The clearest productivity gains Steenbergen described at the Fortune AIQ Summit are internal, not customer-facing. Customer service was the first process Booking rebuilt from what he called "a white sheet" .

The result: bookings grew at a high-single-digit rate while customer service costs moved slightly lower, producing a decline in cost per booking that Steenbergen described as significant . Customer satisfaction improved simultaneously. That combination, volume up and unit cost down with quality rising, is the operating leverage profile every PE investor models but rarely sees delivered.

30%: Booking's roughly 9,000 engineers are getting approximately that much more code into production, measured only by merge requests that pass testing and quality control .

The mechanism Steenbergen described is worth examining in detail. Booking uses what he called "effective model cost routing," directing simple tasks to basic or open-source models and reserving expensive frontier models for complex work. Engineering teams are measured on total IT cost per merge request, a metric that combines human labor costs and AI token costs into a single unit economic figure. Token spending is permitted to rise, provided the all-in cost per production merge request falls .

That cost architecture is replicable. Any enterprise running large developer teams can adopt the same measurement framework. The implication for software and IT services sector valuations: companies that can demonstrate declining cost per productive unit of output, not just rising AI spend, deserve a premium to peers that cannot show the math.

On the customer side, results are early and modest. Customers using Booking's AI tools book slightly faster, convert at marginally higher rates, and cancel slightly less often . Referrals from large language models, paid and unpaid, were below 1% of total room nights as of the company's second-quarter earnings call in August . Steenbergen was direct: "It's very early stage, so it's not a lot of data" .

The High-Frequency Thesis: From Transaction Platform to Travel Operating System

Steenbergen's most forward-looking argument at the Fortune AIQ Summit was not about cost. It was about revenue model transformation.

Booking is, in his characterization, "more transactional" today . AI could make it "a much more high-frequency business" . The example he used is precise: a traveler has a Paris itinerary, Wednesday's forecast is poor, and Booking proactively offers to reschedule the walking tour to Thursday and move the Louvre visit to the rainy day. If a customer with known sports preferences has an empty afternoon, Booking offers to book kite surfing .

That is not a search-and-book model. That is a continuous service relationship with ongoing commercial transactions embedded in it. For institutional capital, the distinction matters enormously. A transactional platform earns revenue once per trip. A high-frequency model earns revenue multiple times per trip across a longer customer relationship window.

The early data points from Booking's second quarter support the directional thesis. Connected trip transactions, defined as bookings across more than one travel vertical for a single trip, grew at a low-double-digit rate at Booking.com . Merchant bookings reached approximately 73% of gross bookings . Level 2 and Level 3 Genius loyalty members accounted for more than 30% of active customers .

The traveler research funnel statistic Steenbergen cited adds context. Consumers visit about five platforms on average before making a booking . If Booking's AI tools can compress that research journey and keep the customer inside its own ecosystem from inspiration to in-trip management, platform switching costs rise and paid acquisition dependency falls. Reducing even a fraction of the $8 billion to $9 billion annual marketing spend through higher direct channel retention would generate a return that dwarfs most AI pilot programs.

The CFO as AI Adopter: Governance Signal for Enterprise AI Rollouts

Steenbergen disclosed at the Fortune AIQ Summit that he uses an AI coach and has two AI agents working for him directly . He discussed this internally on purpose.

One agent, which he calls a "strategic thought partner," assists with board presentations and strategic planning. A second, which he labeled a "critical equity research agent," prepares him for earnings calls . His framing was intentional: "No one knows this. If I have to learn, and I have a coach, it's very normal. Everyone has to learn" .

For governance-focused institutional investors, this disclosure carries weight beyond the novelty. A CFO using an AI agent to prepare earnings call responses sits at the intersection of disclosure quality, material information handling, and fiduciary responsibility. The practice is almost certainly more widespread than disclosed. Steenbergen is the exception in admitting it publicly, not in doing it.

The safety dimension is not theoretical. David Robinson, a former OpenAI employee who wrote the safety reports accompanying major model releases, resigned this week and published a critique arguing that AI industry culture is "broken," that firms operate with "extreme confidence" and "perpetual sprints," and that frontier AI labs require safety infrastructure comparable to nuclear power regulation . Robinson's concerns center on model development, but the governance gap he identifies at the builder level flows downstream to enterprise deployers making material decisions with AI assistance.

The Alibaba Qwen case adds a second dimension to enterprise AI governance risk. Research published by Israeli cybersecurity firm Hirundo, reported by CBS News on October 2, 2026, identified systematic censorship embedded in Qwen, an AI model that has been downloaded more than 3 billion times and is used by U.S. firms including Airbnb and Uber . The model does not acknowledge the 1989 Tiananmen Square massacre and produces Chinese government-aligned responses on sensitive geopolitical topics . Alibaba did not respond to a request for comment .

For enterprise risk officers and compliance functions, the Qwen findings reframe the model sourcing question. Which models are inside the enterprise stack? Who built them? What are the embedded biases? A CFO using an AI agent to prepare earnings call materials needs defensible answers to those questions before a regulator asks.

The Plocamium View

The market is pricing AI investment as if capital deployment itself were the return. Steenbergen's admission that even hyperscalers cannot quantify ROI on their infrastructure spend is not a warning to sell AI. It is a signal to reposition within the AI value chain.

The asymmetric trade is not in model builders. It is in disciplined model users with three characteristics: large, measurable cost pools that AI can address; existing customer relationships that AI can deepen rather than build from scratch; and the operational discipline to measure AI spend at the unit economic level rather than as a line item in a technology budget. Booking Holdings fits all three criteria.

The high-frequency travel thesis deserves a specific valuation overlay. If Booking converts even a portion of its two-thirds direct customer base from single-transaction to multi-transaction relationships per trip, the revenue per customer metric expands without a proportionate increase in acquisition cost. That is multiple expansion driven by business model evolution, the highest-quality form of earnings growth.

The governance risk embedded in enterprise AI adoption is underpriced. The Robinson resignation from OpenAI and the Qwen bias findings, both surfacing in the same week as Steenbergen's remarks, form a pattern. AI tools are inside the enterprise decision stack at the CFO level. The regulatory and disclosure frameworks governing that reality do not yet exist. Companies that build internal AI governance infrastructure before regulators mandate it will carry lower compliance risk and higher institutional investor confidence. Booking's explicit cost routing architecture and unit economic measurement framework suggest it is ahead of most peers on this dimension.

The second-order play: enterprise software vendors and management consultancies that can deliver the process redesign Steenbergen described, rebuilding workflows "from a white sheet" rather than overlaying AI on legacy processes, are positioned to capture the implementation budget that follows AI pilot success. That services layer is where the predictable, recurring revenue lives in an AI economy where model ROI remains uncertain.

The Bottom Line

Ewout Steenbergen gave institutional investors a cleaner framework for evaluating AI exposure than most sell-side research published this year. The hyperscalers are spending on faith. The disciplined enterprise users are measuring in unit economics. The travel platform spending $8 billion to $9 billion a year on customer acquisition, with two-thirds of customers arriving direct, has a structural incentive to make AI work that pure technology companies do not. Watch the cost per booking metric and the connected trip attach rate in Booking's Q3 results. If the internal productivity gains Steenbergen described are sustained and the high-frequency transaction model begins scaling, the current multiple on Booking Holdings does not reflect the business it is becoming.