Asia's travel platforms are building AI agents that stop short of the last click

Within 48 hours in mid-September 2026, two of Asia's largest online travel companies shipped agentic products with the same deliberate limit. Agoda opened a limited public beta of an AI Assistant on 15 September, running on OpenAI's GPT-5.6 and connected to roughly six million properties, 130,000 flight routes and 300,000 activities. It checks live prices and availability, and it hands the transaction back to the user. A day later in Dubai, Trip.com and Mastercard announced that TripGenie would run on Mastercard's Agent Suite for Merchants, with purchases described as consumer-authorised. Neither product completes a booking on its own.

That restraint reads as a product decision. The survey evidence suggests it is closer to a market constraint. In research published this month by Skift Research and McKinsey, 74 per cent of travellers said they still prefer to make the final booking themselves, and 2 per cent said they already trust an AI tool to book on their behalf.

What the adoption numbers actually measure

The same study, based on a survey of 1,015 US travellers with data as of June 2026 and a diary study of 26 travellers logging 211 planning sessions, records adoption that looks strong until it is broken down by task. Sixty-two per cent of travellers said AI tools have made planning easier. Fifty-two per cent expect to use them more over the next 12 months. Thirty-four per cent used AI during the research stage, more than at any other point in the journey.

Then the curve bends. Seventy-one per cent were comfortable letting AI generate initial ideas and 67 per cent were comfortable letting it compare and narrow options, but roughly half would let it place a hotel or flight booking. Fifty-one per cent would let a tool handle a disruption, the lowest score of any use case tested. And 16 per cent had actually used an AI tool to book anything travel related, against 43 per cent who booked directly with an airline or hotel and 42 per cent who booked through an OTA.

The gap between 86 per cent who would consider an AI booking in future and 16 per cent who have made one is the number worth holding onto. Willingness is not the binding constraint on the last click.

The blockers are procedural

What travellers said would raise their confidence points away from model quality. Fifty-two per cent named the ability to review and change anything before confirmation. Thirty per cent wanted a track record of accurate recommendations. Twenty per cent wanted assurance that a human could intervene if something went wrong, and the same share wanted the system to know their preferences and past trips. Only 14 per cent said nothing would persuade them.

Read against the product announcements, this is a specification rather than an objection. Review before commit, reversibility, an escalation path and retained context are engineering requirements, and three of the four are cheaper to build than a frontier model.

Naveen Manga, global chief information officer at Marriott International, locates the problem in the data layer. "Travellers are cautious about giving an open model their credit card," he says in the report. "Rates, policies, and amenities may be directionally correct, but they are not yet authoritative." Appearing in an AI answer and being bookable through one are separate engineering problems, and the second has a failure mode the first does not: in hospitality a wrong answer can mean a broken trip.

The same shape in financial services

The pattern is not confined to travel. At the Global Fintech Fest, which closed in Mumbai on 11 September 2026 with participants from more than 70 countries, agentic AI dominated the programme. Reporting from the event described executives across banking, lending and insurance settling on a similar division of labour: agents assemble the material and draft the recommendation, and a named person signs off. In lending, the reported example was the underwriting note, where the assembly work compresses sharply while the credit decision stays with the credit manager. In insurance, fully autonomous policy issuance was described as not yet on the table.

Two regulated, high-value consumer sectors, working independently, appear to have drawn the line in the same place. That convergence is consistent with a structural constraint. Travel is not the outlier here.

The complication

If the constraint were purely about trust maturing, the direction of travel would be obvious and the only question would be timing. The survey complicates that.

Fifty-six per cent of travellers said that although AI gives them answers faster, they spend more total time on the journey from inspiration to booking because of it. Speed at the level of the individual query has not produced speed at the level of the decision. Meanwhile 52 per cent still feel anxious about making the wrong booking choice and 43 per cent report doubting whether they made the right one, on a journey that now runs to 65 touchpoints against about 45 in 2018.

The checkout was solved. Sixty-one per cent say booking got easier over the past year. The part that has not been solved is the part agents were supposed to remove, and adding a faster answering layer has so far added research time.

Why that may not settle it

Three things point the other way.

First, the delegation ladder is already being climbed in the adjacent layer. Google added hotel booking to its Universal Commerce Protocol in May 2026, building the checkout that agents will need. Whether a given supplier is reachable through it is a decision suppliers are making now.

Second, the operators holding the reservation have an incentive to keep the last click. Seventy-nine per cent of travellers rank greater trust in a supplier's own website as their number one reason for booking direct, ahead of better prices, and 87 per cent place it in their top three. Trust at the transaction is currently the direct channel's main asset, which means the slow pace of autonomous booking is partly a commercial preference, and partly a consumer one.

Third, precedent. Pat Nestor, head of AI and data analytics at Hyatt, argues the discomfort is familiar: "It's not all that different from the first time you could process your credit card online and there was widespread apprehension, but now consumers wouldn't think twice about it." That comes from a company with an interest in the comparison holding, and the survey gives it partial support, since the reluctance is conditional for 86 per cent of travellers.

The real test is narrower than adoption. It is whether the share of travellers who have actually completed an AI booking moves off 16 per cent once review, reversal and escalation are built in. Nobody can answer that yet, because no major platform in the region has shipped a product that does the last step.

What this piece can and cannot show

Three limits are worth stating plainly. The survey behind the behavioural figures sampled 1,015 US travellers, so the percentages describe a source market and not Asian travellers, whose comfort with in-app payment and messaging-led commerce may differ in either direction. The account of the Global Fintech Fest comes from a single trade report carrying a branded-content byline, and this publication could not corroborate the individual remarks independently, which is why the section names no executives. And the product details for Agoda and Trip.com come from the companies' own announcements, with no third party yet able to test what the systems do at scale.

None of the three undermines the central finding, which rests on what the companies shipped and on a survey whose method is published. They do set the boundary of what the piece is claiming.

What to watch

Three markers, each checkable against a named source.

Whether Agoda's assistant moves from limited beta to full mobile rollout with the confirmation step still owned by the user, which Agoda's own release notes and company announcements will record. Whether Booking Holdings discloses a figure for Priceline's Penny beyond Glenn Fogel's description of good engagement and higher conversion on what he called a very small sample, which would appear in quarterly earnings materials. And whether the Mastercard and Trip.com deployment, currently focused on the Middle East and Africa with wider expansion stated for early 2027, carries the consumer-authorised framing into that expansion or relaxes it.

Any one of the three moving would mark the point where the last click stops being a design choice.

Sources

Skift Research and McKinsey & Company; Agoda; Mastercard; Trip.com; Booking Holdings; Skift; The Economic Times. Company figures are drawn from the companies' own announcements.