Before AI Touches the Guest Experience, One Travel Platform Is Learning to Say No

On 24 September, Agoda released its second AI Developer Report, a survey of more than 800 software engineers across India, Indonesia, Malaysia, the Philippines, Singapore, Thailand and Vietnam, commissioned with the research firm Macramé Consulting. The headline finding sounds like a technology story: 53% of these developers now run AI agents in production or across their teams, up sharply from a year ago, while only 38% consider their own codebase ready to hand a task to one of those agents unsupervised. Read differently, it is a story about restraint. The same organisation that books more than six million holiday properties and 300,000 activities is teaching its own engineers exactly how much of that trust an AI has earned, and how much it has not. That distinction, long before it ever reaches a guest, is where the next chapter of the travel industry's AI story is actually being written.

It matters because the travel industry tends to talk about AI in terms of what a guest will notice: a faster booking flow, a chat window that answers at 2am, a recommendation that feels uncannily well-timed. Agoda's own report, drawn from the engineers building underneath all of that, describes something quieter and more consequential. Developers there are not asking whether AI can do the work. They are asking, task by task, how much of it they are willing to let it do without a person checking first. Documentation, low stakes, gets handed over freely. Anything touching production, security or a real business outcome does not: 79% of respondents still require a human to sign off before an AI agent's work goes live, against 21% for the least consequential tasks. That gradient, allowing autonomy exactly in proportion to how easily a mistake can be caught, is the shape every hospitality brand experimenting with AI will eventually have to draw for itself, whether the task is a code change or a message sent to a guest at midnight.

How Agoda's engineers are learning what to trust

Agoda's own Chief Technology Officer, Idan Zalzberg, frames 2026 as the year AI stopped being an assistant for individual tasks and became something closer to a participant, capable of planning and finishing meaningful pieces of work on its own. The company's engineers describe what that transition actually costs. Royee Goldberg, Agoda's VP of Engineering, put it simply: a coding agent is trivial to start and genuinely hard to host, since it needs the real repository, the real dependencies, a database it can reach, and the freedom to fail for twenty minutes before it gets something right. Agoda's answer was to build its own hosting layer, CodeMaster, rather than build its own agent, on the reasoning that agents are improving faster than any team can track and the platform beneath them should outlast whichever one is fashionable this year. The goal for 2026 is modest on paper: 10% of Agoda's own merge requests written inside it.

A companion piece on BriefAsia this week looks at the same report from the cost and governance side. Here, in a magazine written for the traveller, what matters is what this discipline is quietly protecting: the experience at the other end of the booking. A guest does not see a merge request. A guest feels it when a recommendation is wrong, when a rate looks arbitrary, or when a message from a property reads as though nobody actually read it. Agoda's engineers hold a line: an agent should not act unless a mistake is cheap and easy to undo. That is the same line the whole industry will eventually need to draw for a chatbot deciding a rate, or an assistant drafting a guest's itinerary unattended.

The same restraint, well beyond one company

Three other companies profiled in Agoda's report show this same carefulness at work well beyond one engineering team. At the payments platform Omise, Director of Engineering Sylvain Dormieu describes AI as pushing the frontier of what can be automated, while accountability stays firmly with people. At the Indonesian media-intelligence firm dataxet, Head of Engineering M. Ridwan Agustiawan puts the risk in a single sentence: "The real risk isn't AI being expensive. It is AI being used carelessly." One small shortcut compounds into another until nobody can trace where the spend, or the judgment, went missing. And at SCB 10X, the innovation arm of Siam Commercial Bank now building agents that complete entire production services, Tech Intelligence and Insights Manager Oravee Smithiphol offers the line that could sit over this whole report: "Speed has genuinely surprised us. Trustworthiness in production hasn't." Her advice, plainly stated: do not scale an agent faster than the organisation's own ability to check its work.

That is a harder discipline than it sounds, and the report is honest about where the money actually disappears while a company tries to learn it. Erik Perttu, Head of Engineering at the education-review site Edu2Review, tracked one small change, a rename repeated across a dozen files, and found the AI's own work cost barely half a dollar. Close to nine-tenths of the total spend went into the design and review sitting around it. Generation is cheap, as Perttu puts it. The trust-building around it is where the spend actually lives. For a hospitality brand weighing its own AI-assisted guest tools, that is worth sitting with: the model call is the easy part, and it is not where the real investment, or the real risk, is going to live.

Why this caution is being written here first

Microsoft's Harley Young, quoted in the report, notes that the region added roughly 13 million new developers on GitHub in a single year, with India alone on track for 57.5 million by 2030, more than one in three new developers anywhere in the world. Southeast Asia and India are not catching up to a story that started somewhere else. For once, in software at least, they are writing it first, which means the caution this report describes is not an import. It is being built, tested and occasionally second-guessed inside a company whose product a great many of this magazine's readers have already used to book a room.

Where the comparison runs thin

It would be a stretch to read a survey of eight hundred software engineers as a verdict on how hospitality should treat guests. That stretch deserves naming plainly, here, instead of being quietly smoothed over. A developer's comfort handing documentation to an agent says little directly about whether a general manager should let one draft a guest's welcome message unattended. A codebase and a guest's stay are not measured the same way, and Agoda's own engineers were never asked the second question.

What still travels from the code to the guest

What travels between the two is the shape of the caution, not the specifics: autonomy that scales with consequence, a cost that hides in the checking more than in the doing, and a very human insistence that someone, somewhere, still has to answer for the result. The real test, put plainly, is whether hospitality extends that same insistence to the moments a guest actually feels, and not just to the code no guest ever sees. Nobody surveyed here has been asked that question yet, and the honest answer is that no one knows.

What this means for hotel operators

For a general manager or revenue manager weighing an AI concierge, a dynamic-pricing tool or an automated guest-messaging system, Agoda's own practice offers something more useful than a slogan: a working method. Decide in advance, system by system, which decisions an agent may make unattended and which ones still need a person to check before a guest sees the result. A confirmation email or a standard FAQ answer can likely sit closer to the 21% end of Agoda's own gradient. A rate change, a promise made to a guest, or a message sent after a complaint sits closer to the 79% end, where a human still has to sign off. Treating every guest-facing tool the same way, whatever the stakes, is the shortcut Agoda's own engineers have already ruled out for themselves.

The same discipline applies to who inside a hotel is meant to own the checking, and the tool alone is not what determines that. Oravee Smithiphol's caution at SCB 10X, to keep a system's autonomy from outrunning the organisation's own ability to check its work, does not require a hospitality group to build engineering infrastructure of its own like Agoda's CodeMaster. It does mean that whoever adopts a vendor's AI concierge or pricing engine should be able to say, in plain terms, who inside the property reviews its decisions and how often, starting with a pilot group of guests before switching it on across the property.

What to watch

Three things would settle this over the next year or two. Whether hotel and travel brands start publishing their own version of this risk gradient, naming in public what an AI concierge, pricing tool or messaging assistant may decide alone and where a person still has to sign off, instead of leaving guests to find out by accident. Whether the anxiety this report documents among junior developers about their standing in an AI-era workplace, felt by 49% against 17% of CTOs and VPs of Engineering, shows up as a parallel unease among junior reservations and guest-relations staff once agentic tools reach the desk and not just the codebase. And whether a guest is ever told, the way a pull request is reviewed line by line, that a recommendation or a rate came from an agent and was checked by a person, or whether that checking simply stays invisible, the way it mostly does today.

Sources

Agoda AI Developer Report 2026, commissioned by Agoda with Macramé Consulting

PR Newswire APAC, "Agoda Releases AI Developer Report 2026: Agentic AI Adoption Outpaces Enterprise Readiness Across Southeast Asia and India," 24 September 2026