If there is one thing the evidence on artificial intelligence in travel agrees on, it is where the technology has actually landed. Almost all of the AI that travellers use sits in the planning of a trip, in the dreaming and the shaping and the shortlisting, and almost none of it sits in the booking. The usual way to explain that gap is a story about capability, as though booking were simply a harder task that the models will grow into once they improve. I do not think that is right. The models are probably capable of booking already, or close to it; what they lack is not intelligence but trust, and the ecosystem that would let trust form, a payment flow built for agents and the standards, liability and accountability that an autonomous purchase demands. The more useful question, then, is not why AI is worse at booking but why we are willing to let it plan and not to let it buy. Answering it locates the real barrier, and the barrier is not in the model.
Where the AI actually is
Start with the shape of the thing, because it is stark. In Simon-Kucher’s 2026 research, forty-two per cent of travellers had used generative AI to build an itinerary and thirty-one per cent to search for flights and hotels, while only twenty-eight per cent had used a chatbot directly on a booking site. Phocuswright put the share of United States leisure travellers who used AI for at least one trip at fifty-six per cent in 2026, up from forty-three per cent nine months earlier, the fastest behavioural shift it had measured in over a decade, and a Skift and McKinsey study found the share using AI tools extensively for planning had more than doubled year on year. The peer-reviewed literature locates use in exactly the same place: heavily concentrated in pre-trip search and planning, thin and largely illustrative at booking, and reappearing at the post-trip stage in the writing of reviews and the sharing of what happened. Wherever you look, the trust extends upstream of the transaction and stops there.
Why planning earns trust so readily
The first reason travellers extend their trust to planning is that generative systems changed what the planning tool actually does, and changed it into something worth trusting. For two decades, travel recommendation was a retrieval problem. Content-based, collaborative, knowledge-based and hybrid filtering matched a traveller to catalogued options and returned a ranked list, and even recent non-generative tools continue that logic, scoring and searching over a fixed set of points of interest. The traveller’s job was to choose from what the system surfaced. Generative AI removes the list. Instead of selecting among ranked options, the traveller co-constructs a bespoke itinerary through conversation, negotiating budget, duration and taste in dialogue until the plan is theirs rather than the catalogue’s. Several of the studies treat this shift, from selecting to co-creating, as the reason the technology reshapes planning rather than merely speeding it up, and the conversational form brings an affordance the old tools never had: the ability to prompt the system to answer as a local resident, a tour guide or a travel blogger, so that the same request yields a walking route, a logistics plan or an atmosphere piece depending on the voice asked for.
The co-creation has a psychology, and it is what turns trust into a habit. Qualitative work describes an aha moment, a threshold the traveller crosses when the interaction feels situationally normal, the AI’s responses feel real, and, above all, the traveller feels intimately understood. At that point the traveller stops being a passive recipient of itineraries and becomes an active co-creator who iterates on the suggestions rather than merely accepting them. The market data shows the same from the outside: where travellers have used AI to plan, they report saving between one and three hours per trip, and a meaningful share say the tools leave them more confident in their decisions and help them discover places they would not otherwise have found.
But the deepest reason planning earns trust is that planning is a safe place to give it. The stakes are low, the actions are reversible, no money changes hands, and the traveller stays in control, free to verify a suggestion before acting on it and to discard anything that does not fit. One study restricted its scope to pre-trip planning precisely because that stage carries high emotional investment and a dual character, at once a technical information task and a social interaction, and found that both the efficient handling of information and a warm, socially responsive manner raised satisfaction, each through the traveller’s inferred sense that the AI was benevolent. There is even a trust advantage here that exists nowhere else in the journey: travellers value what they perceive as AI’s commercial neutrality relative to a human agent on commission, but only during comparison-shopping, when options are being weighed. That is the planning stage exactly. The one place where perceived neutrality is worth something is the one place the trust is being given.
Why that trust stops at the point of payment
If the tools are trusted this far, why do the same travellers stop short of letting them book? The answer is not that the machine becomes incapable at the checkout. It is that trust is withheld where the stakes turn real, and the surrounding machinery to earn it does not yet exist.
Trust is withheld nearest the transaction, and not for reasons of performance. The literature on AI aversion finds that travellers prefer human recommendations because they believe the machine lacks lived experience, emotional engagement, contextual understanding and the instinct to ask the right follow-up question, and, crucially, that this resistance persists even when the AI is shown to outperform the human. That is the decisive point. If aversion held only where AI was worse, it would be a capability problem waiting on a better model. Because it holds even where AI is better, it is a trust problem, and a better model does not touch it. The survey evidence bears this out at the transaction: by various measures only about two to eight per cent of travellers would let an AI agent complete a purchase on their behalf, and when OpenAI built a booking button into ChatGPT it quietly withdrew it in early 2026, not because the model could not assemble the booking but because travellers were content to ask for ideas and then left to buy somewhere they already trusted.
Capability is probably not the binding constraint. The research that comes closest to autonomous booking suggests the machine can already do most of the work. One benchmark plans a trip, sequences it, collects live pricing and produces a verified proposal, and then excludes the purchase itself as simply not relevant to what it set out to test, a design choice rather than a confession of inability. Another places autonomous booking only midway up its own maturity model, gated not by intelligence but by interoperable data standards, cybersecurity, and unresolved rules on liability and auditability. The machine can assemble the transaction. What it does not do is execute the payment and own the consequences, and that is a matter of rails and rules, not reasoning.
The ecosystem is not built. There is no well-established, agent-native payment flow, and today’s payment systems assume that a person authorised the purchase, an assumption agentic commerce breaks the moment an agent acts on a broad instruction rather than a fixed order. The pieces are being assembled, capped virtual cards, programmable authorisation, emerging protocols to verify a trusted agent in a transactional setting, but they are early, and until they are standard the question of who is the merchant of record and who indemnifies whom when an agent buys the wrong thing has no settled answer. Booking is not a harder modelling problem than planning. It is a trust-and-accountability problem sitting on top of an infrastructure problem, and neither is solved by making the model cleverer.
The mismatch runs both ways
There is an irony worth naming, because it sharpens the point. Trust and capability are misaligned at both ends of the journey. We extend trust to the plan rather freely even though the plan is imperfect, since travellers report outdated and inaccurate responses concentrated at exactly this stage, a survey of more than seven thousand users found nearly half had hit incorrect or stale information, and an audit of AI-generated itineraries found linguistic, representational and stereotype biases that quietly narrow the world toward a familiar, largely Western canon. And we withhold trust from the booking even though the machine could very likely carry it out. In other words, we may trust the planning slightly more than its accuracy warrants and the booking considerably less than its capability warrants. The line between the two is drawn by perceived stakes and by control, not by where the technology is actually strong or weak, which is precisely why it will not move simply because the next model is better.
What this means
The practical consequence is that anyone waiting for a capability breakthrough to unlock agentic booking is waiting for the wrong thing. What has to arrive first is not a smarter model but an infrastructure and a warrant of trust: a payment flow designed for agents, standards that let inventory and authorisation pass cleanly between systems, a clear allocation of liability, and the visible guardrails, the shown options, the human approval step, the recoverable mistake, that let a traveller hand over control at the one moment they are least willing to. The contest for the traveller has meanwhile moved upstream into the planning, where the trust already lives, which means being present in the set of options a co-created itinerary draws upon, synthesisable and machine-readable rather than merely visible, a theme I have pressed in Winning travel’s AI shift. And it means the margin pressure of an AI-mediated market lands on the planning layer first, not the checkout, which is the argument of the companion to this piece, What Is My Moat, and Will It Protect Me From Disintermediation?
The honest summary is this. AI is trusted to plan because planning is safe, reversible and keeps the traveller in control, and because the conversation genuinely understands them in a way the old ranked list never did. It is not yet trusted to buy because buying is irreversible and accountable, because the resistance at the point of payment holds even where the machine performs well, and because the rails that would carry an autonomous purchase, and settle who is responsible when it goes wrong, are still being laid. The line between planning and booking is a line of trust and infrastructure, not of capability. Build the rails and earn the warrant and the line will move. Wait for a cleverer model and you will be waiting for the wrong thing.
Sources
- The stage-by-stage split in traveller use, and the Phocuswright and Simon-Kucher figures: “AI Travel Planning in 2026: What It Gets Right and Wrong,” Food Drink Destinations https://fooddrinkdestinations.com/ai-travel-planning/
- The Skift and McKinsey finding on extensive planning use, and OpenAI’s withdrawal of its in-chat booking button: Gimmonix, “Agentic AI Is Coming to Cut OTAs Out. Or Is It?” https://gimmonix.com/news/agentic-ai-is-coming-to-cut-otas-out-or-is-it
- The share of travellers willing to let an AI agent complete a purchase (roughly two to eight per cent): AltexSoft, “Will Agentic AI Remove OTAs From Hotel Distribution?” https://www.altexsoft.com/blog/agentic-ai-vs-ota-hospitality/
- Time saved, decision confidence and discovery among AI users: HFTP, reporting TakeUp’s “The Rise of AI-Planned Travel in 2026” https://www.hftp.org/news/4130509/new-data-shows-how-travelers-are-using-ai-to-plan-and-book-trips-and-its-not-what-you-think
- Travellers verifying AI plans before booking, and the incidence of incorrect or outdated AI information: “60% of Travelers Still Prefer Human Trip Planning Over AI,” Travel Professional News https://travelprofessionalnews.com/60-of-travelers-still-prefer-human-trip-planning-over-ai-new-civitatis-survey-finds/
- Payment systems assuming a human authorised the purchase, and the human-in-the-loop expectation for agentic commerce: PYMNTS, “When AI Becomes the Travel Agent, Who Owns the Journey?” https://www.pymnts.com/transportation/travel-payments/2026/when-ai-becomes-the-travel-agent-who-owns-the-journey
- Emerging agent-payment and trusted-agent protocols as a still-forming ecosystem: SourceTrail, “The Rise of Agentic Commerce” https://www.sourcetrail.com/software/the-rise-of-agentic-commerce-how-new-protocols-are-automating-travel-bookings/
- The shift from retrieval to co-creation, the role-taking function and the three-phase journey: I. A. Wong, Q. L. Lian and D. Sun, “Autonomous travel decision-making: An early glimpse into ChatGPT and generative AI,” Journal of Hospitality and Tourism Management 56, 2023, doi:10.1016/j.jhtm.2023.06.022
- The pre-generative recommender-systems tradition the shift departs from: S. Renjith, A. Sreekumar and M. Jathavedan, 2020, doi:10.1016/j.ipm.2019.102078; and M. Orabi, I. Afyouni and Z. Al Aghbari, 2025, doi:10.1016/j.ipm.2024.103970
- The aha-moment mechanism in AI-assisted planning: H. Zhao, B. Yuan, Y. Liu and Y. Liao, “Unlocking co-creation in travel,” Journal of Hospitality and Tourism Management 66, 2026, doi:10.1016/j.jhtm.2026.101396
- The dual technical-social character of planning, inferred benevolence and perceived commercial neutrality: G. G. Liu, L. Lv, L. L. Meng and J. Tao, “Beyond algorithms,” Journal of Retailing and Consumer Services 89, 2026, doi:10.1016/j.jretconser.2025.104607
- Trip planning as the most valued function, alongside inaccuracy at the same stage: H. T. Bui, V. Filimonau and H. Sezerel, 2025, doi:10.1016/j.tmp.2025.101392
- The AI-aversion model, and the finding that resistance persists even where AI outperforms: H. Kim, H. H. Shin, H. Yoon and H. Shin, “Why travelers prefer humans over artificial intelligence (AI),” Tourism Management 118, 2027, doi:10.1016/j.tourman.2026.105494
- Bias in AI-generated itineraries: S. Jia, C. Ma, O. H. Chi and A. Fan, “A socio-technical exploration of bias in generative AI travel planners,” Tourism Management 116, 2026, doi:10.1016/j.tourman.2026.105442
- Agentic booking benchmarks that exclude the purchase act by design, and the maturity-and-governance barriers to autonomous booking: S. Chemli, A. Vitale, Shekhar and M. Valeri, 2026, doi:10.1016/j.jengtecman.2026.101959; and J. Yu, “Preparing for an agentic era of human-machine transportation systems,” Transport Policy 171, 2025, doi:10.1016/j.tranpol.2025.05.030
Related on this blog: its companion piece, What Is My Moat, and Will It Protect Me From Disintermediation?, on why the margin pressure of an agent-led market lands on the planning layer before the checkout (add the link once published); and Winning travel’s AI shift on being found and represented inside the models.
Analysis and interpretation are the author’s own. Figures are drawn from the cited peer-reviewed literature and public 2026 reporting and should be verified against the primary sources before republication.
