The global travel and tourism sector is projected to contribute nearly US$12 trillion to global GDP in 2026, supporting 376 million jobs—roughly one in every nine jobs worldwide. Yet beneath this buoyant statistic lies a troubling paradox for the travel trade sector. The platforms that connect travellers with airlines, hotels, and experiences remain, by their own CTO’s admission, fundamentally limited in what they understand about the journeys they facilitate.
“Our platform today is somewhat dumb,” said Idan Zalzberg, chief technology officer at Agoda. “It doesn’t think about your trip. It thinks, ‘oh, you’re searching for Bangkok.’ It tries implicitly to give you the best results based on machine learning models and everything we built. But it doesn’t say, ‘Allan is having this weekend in Singapore with the family.’”
This candid assessment from one of Asia’s largest digital travel platforms signals a reckoning. The architecture that has defined online travel for two decades—optimised for discrete, transactional searches—is being forced to evolve. The new organising principle is not the search query, nor the booking, but the trip itself. And the technology forcing this change is artificial intelligence.
The shift from transaction to journey
The behavioural data confirms what Agoda’s CTO describes internally. Phocuswright’s latest research shows that 56% of travellers used AI for planning, booking, or in-destination assistance in the past 12 months—up from 43% just six months earlier. Every generation posted double-digit gains. Among those who have used AI, 94% have applied it specifically to travel.
More revealing is where in the journey AI is gaining traction. McKinsey’s March 2026 ConsumerWise survey found that roughly half of AI users rely on it to discover destinations and brands, about 40% use it to compare options, but fewer than one in four use it to actually book. AI is becoming the top of the funnel. If a hotel or airline is not surfaced, understood, or recommended at that stage, it may never enter consideration at all.
Skift Research’s State of Travel 2026 captures the implication starkly: only 6% of hotels currently surface in AI search results. The other 94% are invisible to the customer’s first question.
Zalzberg’s diagnosis of why legacy platform architecture struggles with this shift is precise. A traditional OTA processes each search as an isolated event. It has no memory of the trip being assembled across days or weeks. It cannot reason that a flight has been booked but the hotel has not, or that a gap remains in the itinerary.
“You can’t start from scratch every time,” he said. “It should be like, ‘oh, Allan, you’re back. Look, what about that flight? We still don’t have a way to get there. Let’s close this flight, and then we can move on to something else.’”
This is not a user-experience tweak. It is an architectural inversion.
Making the trip the atomic unit
Zalzberg describes the required transformation in unambiguous terms: “The trip has to be a first-level citizen in everything we do, not just a current search.” The platform must maintain persistent context across sessions, across devices, and across the fragmented components of a journey—flights, accommodation, ground transport, activities, and contingency planning.
The vision he articulates is to become a “personal travel agent” for every traveller, not just the ultra-wealthy who employ human concierges. “Imagine one of those billionaires,” he said. “They don’t go on Booking.com or Agoda to search by themselves. They have someone that thinks about their experience end to end.” That someone anticipates needs, adapts when weather disrupts plans, and provides options before the traveller knows to ask.
Delivering this to mass-market travellers requires infrastructure that today’s OTAs were not built to provide. Zalzberg acknowledges the scale of the challenge: “It requires you to rebuild or at least rearchitect parts of the platform to be much more concentrated on the trip.”
The technical demands are formidable. Persistent context means session state must survive across days and devices. Real-time inventory accuracy must be maintained across a supply chain that Zalzberg describes as riddled with “APIs that are not documented or don’t behave when they’re supposed to behave.”
And the system must orchestrate multi-system calls—airline availability, hotel inventory, payment processing—without the latency that would make the experience unusable.
Why this matters for airlines, hotels, and airports
The travel trade sector has a direct stake in whether platforms succeed in this rearchitecture. Airlines, hotels, and airports are increasingly dependent on digital platforms not merely for distribution, but for demand forecasting, inventory management, and dynamic pricing.
The timing is precarious. IATA’s June 2026 Global Outlook paints a sobering picture: global airline net profits are forecast to fall to US$23 billion, a margin of just 2.0%, as jet fuel prices surge 70% to an average of US$152 per barrel. In this environment, every percentage point of distribution efficiency matters.
Deloitte’s Travel Weekly Annual Report 2026 identifies AI and technology as one of the defining forces shaping the industry, noting that “competitive advantage will come from blending cutting-edge technology with human expertise”. But the report also warns that organisations must balance innovation with the “enduring need for human connection in what is considered a high-value travel purchase.”
For hotels, the risk is existential. BCG’s AI-First Hotels report argues that as more travellers use AI assistants to plan trips, hotels risk losing visibility entirely “if their inventory, content and pricing are not ready for conversational search” . The report recommends that hotel companies regain control of first-party guest data and connect directly with AI interfaces.
Zalzberg’s advice to the trade sector is implicit in his description of Agoda’s own infrastructure challenge. The platform works with a vast network of suppliers, “possibly the largest supply in the world,” and handles whatever they can provide. But the quality is uneven. “Many of them just break apart at high scale,” he said.
Agoda’s response has been to build predictive caching: machine learning models that anticipate what travellers will search and which suppliers will have the best availability, so that responses are pre-fetched before the user completes their query. “You, as a customer, can still get a good kind of performance,” Zalzberg explained. “So that is definitely todays on us.”
For airlines and airports considering their own AI investments, the lesson is that integration is not a one-way street. Platforms are already absorbing the cost of supplier-side fragmentation. The question is whether suppliers will invest in standardised, well-documented, scalable APIs that reduce that burden—or continue to rely on platforms to normalise their inconsistencies.
The engineering reality: Non-deterministic software at scale
Zalzberg’s candour about the engineering challenges of this transition is striking. When Agoda began deploying generative AI internally, his team confronted a fundamental problem: none of the basic assumptions of software engineering held true.
“When I write code, I know what this code will do,” he said. “When I run the code, I can check if it did what it’s supposed to do. If I run it again, it will do the same thing. None of them are true for AI.”
This forced Agoda to develop new practices for what Zalzberg calls “predictable software with unpredictable, underlying components.” The most important is the use of evaluations—a sophisticated form of testing where a separate LLM judges whether outputs meet quality and factual standards.
“You can use another LLM to see if the result is what we call LLM-as-a-judge,” he explained, “to look at the output of the model and say, ‘do you think this output matches what you expect? Do you think this output is based on facts that’s been used in the conversation?’ So, to prevent hallucination.”
The architecture must also enforce strict boundaries. “You want to have clear rules, separation of roles,” Zalzberg said. “This agent is in charge to do this and that. This is the scope in which it lives. Anything else, it should either pass to another agent or say it cannot do it, so the scope doesn’t become infinite.”
These constraints matter enormously when the software is managing live travel bookings. A hallucinated hotel confirmation or a misinterpreted cancellation request is not an abstract failure; it is a traveller stranded at an airport, or a hotel charging the wrong amount. “We play with people’s travel,” Zalzberg said. “It’s not life and death, but it’s a very big thing for many people.”
The cost of intelligence: Optimising without sacrificing quality
The infrastructure burden of running AI at Agoda’s scale—processing half a million tokens per second—is significant. Zalzberg is dismissive of the “token maxing” culture that treats higher token usage as inherently better. “I would say almost insulting as an engineer,” he said. “It all has to be based on what outcome you’re trying to achieve and how much you want to spend on that outcome.”
Agoda’s solution is architectural flexibility. The company built a unified endpoint that abstracts away API differences between models from Google, OpenAI, Anthropic, and open-source providers. Developers can switch models by changing a single name, enabling rapid evaluation of quality, latency, and cost trade-offs.
For the travel trade sector, this approach offers a template. Airlines and hotels facing margin compression cannot afford to treat AI infrastructure as an unlimited budget line. The discipline of measuring cost-per-outcome, rather than cost-per-token, will separate sustainable AI deployments from expensive experiments.
The platform as journey orchestrator
The architectural rewrite that Zalzberg describes is not a distant roadmap item. It is already underway, driven by traveller behaviour that is shifting faster than the industry’s ability to respond. Phocuswright’s data shows that AI use in travel has crossed the majority mark in a single year . Expectations are rising in tandem.
For the travel trade sector—airlines, hotels, airports, and travel retailers—the implication is clear. The platform layer is being rebuilt around the trip, not the transaction. Those who build their data infrastructure, API standards, and content strategies to integrate with this new architecture will find themselves in the consideration set. Those who do not will be invisible.
“The problem is how do we take what we have today, which is a search and book, and convert it to that,” Zalzberg said. The answer, for Agoda and for the industry, lies in making the trip itself the centre of the technological universe. It is a fundamental change—and it is happening now.
