For years, travel agents have been treated as an endangered profession. Online booking was supposed to remove them. Now artificial intelligence is expected to finish the job: a traveler describes the holiday, an AI searches every option, completes the booking and manages whatever happens next.

Recent reporting points toward a more complicated outcome. The Wall Street Journal reported in May 2026 that demand for travel advisors is growing as travellers seek expert recommendations, personalised service and help managing complicated trips.

Cassio believes the combination of human expertise and automation could produce something more significant than the survival of the travel agent.

“I think, if anything, we’ll give them superpowers,” he told me. His prediction is that agents will become a larger part of the industry over the next five years, not a smaller one.

But those superpowers depend on more than a conversational interface. The infrastructure underneath must simplify fragmented content, control the economics of AI-generated search, automate repetitive work and know when a machine should hand a decision back to a person.

AI can simplify the experience. It cannot remove the complexity underneath

Cassio oversees Travelport’s API and travel-agent products across shopping, booking and servicing. Travelport provides infrastructure connecting airlines, hotels and other suppliers with travel agencies, corporate travel-management companies, online travel agencies and technology platforms.

Before AI can make advisors more productive, the underlying travel system must stop exposing its own complexity to them.

That infrastructure reflects decades of technical change. Traditional airline content may arrive through EDIFACT, the long-established messaging standard used across travel distribution. Newer offers increasingly arrive through NDC, or New Distribution Capability, while low-cost carriers often use different connections again. Hotels add another set of rates, rules and booking processes.

A travel advisor should not need to care which technical standard or supplier connection produced an offer. Their job is to understand the traveler, compare the relevant choices and sell the right trip.

That is the logic behind orchestration: instead of requiring an agency or technology company to build separate workflows for every source, the infrastructure coordinates the systems underneath. The seller asks for the trip; the platform determines how to search, book and service it.

Travelport formalised this direction in June 2026 with TripServices, its cloud-native API platform for flights and stays through a single connection. An API is the software interface through which one system requests data or actions from another.

The strategic shift is from distributing separate content pipes toward one connected shop–book–service experience. AI makes that orchestration more important: a conversational interface can conceal complexity from the traveler and advisor, but it cannot make the complexity disappear.

Agentic search can create more noise, not always better answers

A human traveler is limited by patience. They might test several dates, compare two airports and review a handful of destinations.

An AI agent has no such constraint. Ask it to find “somewhere warm from Munich in August” and it could explore thousands of combinations across destinations, dates, airports, fares and connections.

This puts pressure on the look-to-book ratio: the number of searches generated for each completed booking. If AI agents produce vastly more searches without proportionally more bookings, costs can rise across airlines, distributors and sellers.

Cassio’s position is direct: placing a large language model behind every request will not work economically. In his view, AI should be a smart layer above machine learning, cached results, live availability and proven distribution technology—not the engine performing every search and calculation.

A system that generates 10,000 possibilities does not give an advisor superpowers. It gives them more noise.

A more useful system first interprets the traveler’s intent: who they are, what matters to them and which compromises they may accept. Existing data and machine learning can then narrow the possibilities before live pricing and availability are requested.

The industry is preparing for much more machine-generated traffic. At Phocuswright Europe, Microsoft Advertising cited projections that agentic web traffic could overtake human traffic by 2030. The same discussion suggested that agentic interfaces would coexist with human and conventional digital channels rather than replace them outright.

The infrastructure challenge is therefore not simply processing more searches. It is understanding travellers well enough to search less wastefully.

Start with safety architecture

Most AI projects begin with the question: what can the model do?

Cassio begins somewhere else: what must be true before the model is allowed to do it?

He calls this safety architecture.

Large language models are probabilistic: they generate a likely answer rather than following a guaranteed sequence of rules. That is useful for interpreting requests, summarising information and handling repetitive work. It is riskier when an output can exchange a ticket, cancel a reservation or leave a traveler without accommodation.

Cassio made the distinction through a simple comparison.

Receiving shoes in the wrong colour is inconvenient and reversible. An incorrectly assembled family holiday could mean the wrong hotel, missing meals or a journey that takes 56 hours instead of 18.

Safety architecture means surrounding AI with trusted data, deterministic rules, validation checks and clear boundaries. Repetitive and reversible tasks are the strongest candidates for automation. Higher-risk actions require additional verification or agent approval. When the system lacks the information or confidence to act, it should escalate rather than guess.

Your trip starts after you book

Cassio illustrated the importance of context with his own travel between London and Milan. Milan is served by several airports. If a flight to Malpensa is cancelled, a system might find a replacement to Linate with roughly the same departure and arrival time. Technically, it has solved the problem: the traveler is still flying to Milan.

For Cassio, however, arriving at Linate can add around two and a half hours to the journey to his actual destination. He may be better served by a flight leaving two hours later but still arriving at Malpensa.

The technically valid answer is not necessarily the right answer.

“You need the context, you need the judgment, you need the human expertise,” he said.

This is why Cassio argues that “your trip starts after you book.”

Servicing is where complexity becomes personal: schedule changes, weather disruption, exchanges, refunds and trade-offs that were never captured in the original booking. AI can collect alternatives, check fare rules and prepare a recommendation. Selecting the best outcome may still require understanding where the traveler is going after the airport, why they are traveling and what inconvenience they would prefer. Yet agents spend much of their days on the routine work surrounding those decisions.

Cassio’s product philosophy starts with observing agents and measuring where their time goes. The strongest automation opportunities are repetitive, recognisable and reversible. When uncertainty or material consequences enter the process, the system prompts the agent.

“We want to give agents their time back so they can sell,” he said.

Travelport is applying this approach to areas such as passenger-name-record management and routine post-booking workflows, while testing broader servicing use cases with selected customers.

Cassio said adoption of the workflows already available is rising rapidly, as is the number of processes handled through the platform. Some broader capabilities are still being tested with selected customers.

The same operating logic is appearing elsewhere. Amadeus reports significantly faster execution across selected agent tasks—including fare-rule checks and booking-history work—and fewer booking errors. Egencia AI is designed to let business travellers ask, book and manage travel conversationally while preserving fast access to a travel consultant and giving travel counsellors more context.

These applications are less dramatic than a fully autonomous booking agent. However they may be more commercially important.

The golden age of travel agents

When I suggested that the industry might be approaching a golden age for travel agents, Cassio agreed.

He believes AI will eventually handle more discovery and booking. But capability alone does not determine what travellers will delegate. Expensive and complicated trips depend on trust, and disruption creates situations where someone must understand both the traveler and the consequences of each option.

Cassio’s prediction for the next three to five years is that technology and specialist knowledge will make agents more productive, capable and important.

The strongest travel AI will reduce the routine search and servicing work that consumes agents’ time, increasing the value of what remains human: context, judgment, empathy, expertise and accountability.

That is how AI could create something the industry did not initially expect:

the golden age of the travel agent.