When Kiruba Shankar, an insurtech founder, pointed a coding agent at Etihad Airways' website in early June and let it run, the result was not a flight booking — it was a stress test. The agent returned more than 881,000 fare combinations for a single trip, crawling every date, route, and connection it could reach. That kind of query load would take a human traveler years to replicate, and it cost Etihad's servers real money to answer. Shankar was not a bad actor. He was a user who outsourced a task to AI, and in doing so, exposed the deepest structural problem in the travel industry's most-hyped technological transition: AI agents that are genuinely good at searching for a trip cannot reliably execute the transaction that ends one.

That problem moved from abstract to concrete on July 1, when Travelport expanded its TripServices application programming interface platform to serve more than 400 European travel agencies through a new integration with Travelsoft's Orchestra platform. The deployment is the first major commercial test of a two-layer architecture that Travelport, Sabre, and the industry's infrastructure incumbents all now agree is the only technically sound approach to AI-assisted booking. Understanding why requires understanding what a large language model can and cannot do — and why those limits matter the moment a traveler's credit card enters the picture.

Why AI Travel Booking Agents Cannot Confirm a Flight

A large language model generates output by sampling from a probability distribution over tokens learned during training. It does not execute database lookups. When an LLM predicts that a seat on a particular flight "is available," it is producing the most statistically probable next token given its context — not querying a live reservation system. That distinction is invisible during a conversation about where to go for a beach holiday in Thailand. It becomes financially consequential the moment the same system attempts to confirm a non-refundable fare.

In traditional online booking, deterministic systems made confirmation possible: enter origin, destination, date, and passenger count, and the reservation system returned an authoritative answer drawn from live inventory. Deterministic means the same input always produces the same output — which is the exact computational opposite of how a language model operates. A deterministic booking API, when queried for seat availability, returns the current state of a live database. An LLM, when asked the same question, produces a plausible-sounding answer that reflects its training distribution, not reality.

This is the structural reason why a June 2026 SmartCustomer analysis of consumer travel reviews found a consistent pattern of AI booking platforms assigning confirmed statuses to pending bookings, processing unauthorized card charges, and canceling trips without user consent. One reviewer lost $650 on a nonrefundable trip the AI canceled unilaterally. Another paid $300 to fix an error the AI had introduced. No clear legal framework currently assigns liability when an AI booking agent errors.

The trust data confirms the failure mode. An Expedia Group survey of more than 5,700 adults published in April 2026 found that only 8% of respondents are willing to let AI complete a booking, while 68% continue to turn to established travel brands over AI agents. A Booking.com survey found that 91% of respondents have concerns about AI-generated travel outputs. Gareth Williams, founder and former chief executive of Skyscanner, has warned that once consumer trust erodes after a booking goes wrong, it may be unusually difficult to rebuild.

How Travelport's Two-Layer Architecture Works

The engineering answer Travelport is building is a deliberate separation between what AI does well and what it structurally cannot do.

The upper layer is conversational. An AI system powered by Anthropic's Claude interprets a traveler's natural-language intent — vague coastal holidays, "something flexible in Southeast Asia," a corporate reroute that needs to comply with travel policy — synthesizes preferences, and surfaces options. This is where language models earn their place: handling ambiguity, synthesizing multiple constraints, and generating coherent recommendations.

The lower layer is deterministic. Travelport's TripServices platform, a cloud-native API system, handles actual booking execution. When the AI needs to verify live availability, confirm a fare, or issue a ticket, it calls this API — not its own internal probability model. The API returns exact, authoritative answers from live inventory systems. Machine learning models embedded in TripServices rank content dynamically, surfacing the offers most relevant to a specific trip rather than returning undifferentiated results — but they do so on top of a confirmed-accurate inventory base, not a probabilistic guess.

The bridge connecting these two layers is the Model Context Protocol, an open standard Anthropic introduced in November 2024 and subsequently donated in December 2025 to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI. MCP solves what its creators called the N×M integration problem: before MCP existed, every AI application needed bespoke connectors to every external data source. MCP defines a standard interface using JSON-RPC 2.0 messages over standard input/output, so any MCP-compatible AI client can communicate with any MCP-compatible server without custom integration work for each pairing.

In Travelport's implementation, MCP allows Claude to interact directly with TripServices' booking infrastructure without the language model ever having to infer a booking outcome. The model resolves the traveler's intent, then hands execution off to the authoritative system through the protocol. The LLM never "knows" whether a seat is available — it calls the tool and receives a guaranteed, live answer.

"AI agents need clean, structured, normalised data and deterministic APIs to operate effectively. That is what Travelport TripServices provides," said Andrew Jordan, Travelport's Chief Product and Technology Officer.

Consulting firm Cognizant is contributing engineering capacity to the deployment, using its Neuro-San multi-agent framework to help analyze Travelport's legacy codebase — a challenge in its own right, since decades of embedded fare rules and business logic must be preserved through the modernization process. The first customer-facing AI capabilities from the Anthropic-Cognizant collaboration are expected later this year.

The Infinite Search Problem Breaking Airline Booking Systems

The Etihad query that returned 881,000 fares is not an anomaly. It is the expected behavior of an AI agent operating against a booking system built for humans, and it points to a second structural problem that runs alongside the confirmation flaw.

Airline fare data increasingly operates under the International Air Transport Association's New Distribution Capability standard, an XML-based protocol that replaced legacy EDIFACT messaging. EDIFACT limited airlines to 26 booking codes — 26 discrete fare classes per flight. NDC enables continuous dynamic pricing: an airline can theoretically offer an unlimited number of price points based on personalized factors. That flexibility is excellent for revenue management. It is catastrophic for AI search costs. Every AI agent query against an NDC-enabled system triggers a live fare calculation that must account for the full continuous pricing space. According to analysis cited by PhocusWire, look-to-book ratios for AI-generated queries could reach 200,000:1 — 200,000 search requests for every single completed booking. At that ratio, the cost of serving search traffic exceeds the revenue from bookings.

Airline websites compound the problem further. Most are built around JavaScript rendering, meaning the page assembles itself in a browser. An AI agent that cannot execute JavaScript — which covers most current agents — sees an empty page where a human browser would show flight results. A March 2026 Bain & Company study that tested three major AI models against real airline websites found that AI agents accessed airline websites directly only about 5% of the time. The other 95% of discovery went through online travel agencies and intermediaries with structured, machine-readable data feeds.

Amadeus, which holds approximately 40% of global GDS market share, has proposed addressing the cost problem through precomputed fare caching. Rather than answering every AI agent query with a live NDC calculation, Amadeus would precompute a large set of likely queries and cache the answers. AI agents would then query the cache rather than triggering live fare calculations — solving the cost problem while keeping Amadeus as the intermediary that every AI booking system must route through. Decius Valmorbida, president of travel at Amadeus, outlined the proposal in a June 23 interview, describing it as the company's answer to the "infinite search" problem.

The GDS Infrastructure War and Who Will Control AI-Era Booking

The deeper story in the AI travel booking transition is not purely technical — it is structural. The same companies that built the global distribution system infrastructure dominating travel booking since the 1960s are now racing to ensure that AI-era booking routes through them rather than around them.

The global distribution system originated when American Airlines and IBM built SABRE in 1960 as an internal reservation system. By the 1980s, competing carriers began sharing inventory across systems, creating the multi-supplier network that became today's GDS. Amadeus, Sabre, and Travelport collectively dominate that network, charging airlines and travel agencies per-booking fees. Airlines spent years attempting to reduce that dependence through direct distribution channels. NDC was designed partly to enable that bypass. AI agents, it turns out, may be driving the distribution layer's next iteration — and the incumbent GDS providers have a head start in building the infrastructure AI agents need.

The emerging tension surfaced publicly in May 2026, when Sabre chief executive Kurt Ekert used the company's strongest financial quarter in more than two years to accuse Amadeus of holding "a dominant monopoly position" in airline technology and announced that Sabre is exploring "regulatory and legal approaches." The specific dispute concerned Amadeus's Altéa passenger service system and alleged restrictions that make it difficult for airlines to switch to competing providers. Ekert's accusation followed Amadeus's February 2026 announcement that it would shut down its self-service developer API portal on July 17, 2026 — a move that blocks independent developers from experimenting with Amadeus APIs without an enterprise contract, timed precisely as AI agents are beginning to reshape how travel is distributed.

Timothy O'Neil-Dunne, principal at travel and aviation consultancy T2Impact, was direct about the implications: "Amadeus is so dominant that they must provide access or they will face the wrath of the regulators." Louis-Hippolyte Bouchayer, head of lodging strategy and supplier management for SAP Concur, was sharper: "AI agents are about to redefine how travel is discovered, priced and booked — and the response is to gate access? That won't stop innovation. It just pushes it outside the walls."

In February 2026, Amadeus also acquired SkyLink, a Y Combinator-backed startup whose multilayer orchestration engine routes complex travel decisions — search, policy compliance, booking execution — through a natural language interface. The acquisition gave Amadeus a production AI booking system that had already completed tens of thousands of bookings. The pattern is consistent with the precomputed fare cache strategy: Amadeus is positioning itself as the obligatory infrastructure layer that AI booking agents must connect to, rather than a network that agents might route around.

Travelport's own data reflects the demand these infrastructure bids are competing for. API-driven transactions through Travelport climbed from 43% of all customer transactions in 2022 to 63% in 2026 — a reflection of rising demand from AI-native platforms that require machine-readable, structured booking access. Travelport shareholders committed $50 million to support TripServices' expansion and future AI development.

Sabre and Mindtrip: A Live Blueprint for Agentic Travel Commerce

Travelport is not the only GDS building the two-layer architecture. Sabre's answer was to partner with AI travel startup Mindtrip and payment platform PayPal on what the companies described as the first end-to-end agentic AI flight booking experience, launched May 6, 2026.

Mindtrip Flights runs the same structural logic as TripServices: a conversational AI front end handles intent resolution and recommendation, while Sabre's Mosaic platform — with agentic-ready APIs connecting more than 420 airlines and 2 million lodging options, including 150 low-cost carriers — handles live inventory access and booking execution. PayPal provides identity verification and payment infrastructure, embedding checkout directly into the conversational flow so travelers never leave the chat to complete a transaction. The architecture eliminates the point at which a language model would otherwise have to guess at inventory state: Sabre's APIs supply the authoritative answer, and the LLM routes the user's natural language through to that answer.

"For the first time, we've truly bridged the gap between inspiration and purchase, allowing travelers to move from intent to transaction entirely within the same chat," said Garry Wiseman, President of Product and Engineering at Sabre.

The Mindtrip Flights launch arrived against the backdrop of Sabre's strongest financial quarter in more than two years, giving the company both the commercial momentum and the public platform to escalate its dispute with Amadeus. Whether the regulatory and legal approaches Ekert described will materialize into formal antitrust action remains to be seen — no enforcement body has opened a formal investigation as of the publication of this article.

Will Travelers Actually Trust AI to Book for Them?

The infrastructure architecture may be sound. The consumer trust gap remains the harder problem.

The Expedia survey found that while 53% of consumers are comfortable letting AI suggest options, only 8% are willing to let it complete a booking. Two-thirds said they would not trust an AI assistant to make a reservation on their behalf at all. A Dune7 survey of 1,000 US adults identified the primary barriers as fear of irreversible AI errors, unclear accountability when something goes wrong, and data privacy concerns. The liability question remains unresolved: no regulatory framework currently assigns responsibility when an AI booking agent charges a card without authorization or cancels a non-refundable booking.

Andrew Jordan's framework for the industry maps these concerns onto a practical spectrum: at the discovery end, where a loose prompt about a Thai holiday is useful and 80% accuracy is good enough, AI performs well today. At the execution end — where a corporate traveler needs a policy-compliant reroute with zero error tolerance — the infrastructure layer is what closes the gap between useful and trustworthy. That is the bet Travelport made on July 1 when it connected TripServices to more than 400 European agencies through Orchestra. If the architecture delivers, the scenario Jordan describes — an AI routing a traveler to the wrong destination because a prompt was ambiguous — becomes a problem the infrastructure prevents, not one the traveler discovers at the airport.

Frequently Asked Questions

Why can't AI travel booking agents confirm a flight the way a regular booking system can?

Large language models generate output by predicting the most statistically probable next response based on training data — a fundamentally different process from querying a live database. When an LLM states that a seat is available, it is producing a plausible-sounding answer shaped by its training, not executing a real-time lookup against an airline's reservation system. Confirmed booking requires a deterministic system that returns an authoritative answer from live inventory. Until AI agents are architecturally connected to those systems through a reliable interface — such as Anthropic's Model Context Protocol — their booking confirmations are probabilistic inferences, not guarantees.

What is the look-to-book problem and why does it matter for travelers?

The look-to-book ratio is the number of search queries generated per completed booking. Traditional online travelers generate dozens to hundreds of searches before booking. AI agents, which can exhaustively enumerate every fare combination without tiring, have been documented generating look-to-book ratios as high as 200,000:1. At that scale, the cost to airlines and booking systems of serving AI-generated search traffic far exceeds booking revenue. Airlines including United have already introduced penalty fees for agencies generating excessive NDC search volume, and Amadeus is proposing precomputed fare caches to absorb the load. For travelers, the practical consequence is that the cost of AI-driven search will eventually be built into fares if the infrastructure does not adapt.

Who is responsible when an AI booking agent makes an error — charges your card, cancels your trip, or books the wrong flight?

No regulatory framework currently assigns liability clearly when an AI booking agent errors. Consumer protection laws in most jurisdictions were written for human agents and deterministic booking platforms, not autonomous AI systems. Industry observers including Gareth Williams, founder of Skyscanner, have warned that once consumer trust in AI booking erodes after a publicized error, it may be unusually difficult to rebuild. Until a legal or regulatory framework is established, travelers who allow AI agents to complete bookings — rather than only recommend options — bear most of the financial risk of those errors.

Is Amadeus trying to become a mandatory gatekeeper for AI travel booking?

Amadeus, which holds approximately 40% of global GDS market share, has taken a series of steps that its competitors and industry observers interpret as an attempt to position itself as the obligatory infrastructure layer for AI-era booking. These include the February 2026 closure of its self-service developer API portal effective July 17, 2026 — leaving only enterprise-contracted access — the February 2026 acquisition of SkyLink, an AI-native booking orchestration system, and the proposal to precompute fare caches that AI agents would query instead of triggering live NDC calculations. Sabre chief executive Kurt Ekert has publicly described Amadeus's position as a "dominant monopoly" and stated that Sabre is pursuing regulatory and legal remedies. No formal antitrust proceeding has been opened as of publication.