Probability with money attached

Dean Lee

markets are probability with money attached.

<- all essays

AI Economics / No. 093

The Zero-Marginal-Cost Shopper and the Airline Seat

United Airlines moved to ban third-party AI booking agents on November 11. Behind the security and regulatory warnings sits a harder economic constraint: unconstrained machine search breaks airline look-to-book economics and dynamic yield management.

United Airlines updated its Digital Trust Center policies on November 11, 2026 to prohibit third-party artificial intelligence agents and automated browser assistants from searching, booking, modifying itineraries, or redeeming MileagePlus miles across both direct channels and travel aggregators. As CBS News reported, the carrier warned travelers that sharing credentials exposes sensitive data, triggers unintended purchases, and risks violating Federal Aviation Administration and Department of Transportation rules, including the check-in attestation confirming bags contain no explosives or other banned substances. Speaking at a Chicago industry event the same week, United chief executive Scott Kirby stated the commercial objection plainly, telling the audience he did not want to create another middleman.

The policy lands as consumer model vendors push directly into travel execution. Meta has promoted flight reservations as a core use case for its Muse assistant, OpenAI acquired travel agent startup Odessia ahead of the rollout, and Southwest Airlines integrated its inventory so ChatGPT users can query fares inside the chat interface. Network carriers face a different set of incentives. Airline distribution has spent three decades defending direct booking channels against intermediaries. When an online travel agency or metasearch portal sits between the airline and the passenger, the carrier loses ancillary merchandising opportunities, yields customer data, and pays transaction fees. Allowing an external AI assistant to log into a traveler’s account or scrape checkout flows inserts a new interface layer that bypasses seat maps, paid upgrades, and co-branded credit card offers.

There is a second pressure point inside the reservation stack itself. Airline revenue management relies on third-degree price discrimination, separating price-sensitive leisure travelers from time-sensitive corporate buyers through fare rules and continuous seat-class repricing. That pricing structure assumes human shoppers face positive search costs. A traveler compares perhaps five to twenty date and airport combinations before picking a fare and closing the browser tab because attention is finite.

Client-side software has no fatigue constraint. Instructed to minimize total ticket price across flexible dates, an autonomous browser agent systematically enumerates every combination of departure time, layover, and fare bucket. In June 2026, an insurtech founder pointed a coding agent at Etihad Airways’ website and pulled 881,076 fare combinations for a single trip. The user was not attacking the airline; he had simply outsourced price comparison to a script with zero marginal search effort.

Reservation servers were not priced for that query load. According to a November 2025 IATA distribution whitepaper, major travel IT providers already process 3 billion to 13 billion daily search requests, while large metasearch engines reach 100 billion, pushing conventional look-to-book ratios into the tens of thousands of queries per ticket sold. Under New Distribution Capability protocols, each search triggers a live pricing calculation across continuous fare tiers rather than static booking classes. Industry estimates reported by PhocusWire and TravelersToday put agentic look-to-book ratios above 200,000 searches per completed booking. At that ratio, the compute and server expense of answering unmonetized queries begins to erode the margin on the seat itself.

Carriers and distribution vendors have already started rationing access. United previously introduced penalty fees on travel agencies generating excess search traffic through NDC feeds, and Amadeus deployed machine-learning filters that blocked more than 70 percent of low-intent queries on Air France-KLM’s network while testing precomputed fare caches. Caching solves compute load at the expense of price accuracy, since dynamic inventory can change before a stale cached offer reaches checkout. Blocking unauthorized personal agents across its booking, ticketing, and distribution policies attacks the problem from the demand side by removing the payoff for unthrottled scraping.

United left open the option of authorized integrations once security, accountability, and commercial terms are settled. Until platforms settle who pays for look-to-book compute and who owns the checkout relationship, automated assistants can help travelers plan routes, while final pricing and payment stay behind authenticated airline walls.