The software teams furthest ahead on AI in this region still have a person review almost everything it produces. Eighty-six per cent of them review or validate AI output always or most of the time.
They also save a great deal of time. Fifty-five per cent now save at least seven hours a week, up from 18% a year earlier.
Both figures come from the Agoda AI Developer Report 2026, and they belong in the same sentence. The time saving is real and has tripled in a year. It sits alongside near-universal checking, not instead of it.
Why look at developers at all
This is not a post about writing code. The report surveyed more than 800 developers and engineering leaders across Indonesia, Malaysia, Thailand, the Philippines, Singapore, Vietnam and India, online between June and August 2026.
Software teams are worth watching for a simple reason: they are further ahead on AI than any other function in the region. More than half of them (53%) now have AI agents in production or in broad use. If any group had been able to stop checking, it would be this one.
They have not. Seventy-nine per cent still require a person’s approval before anything goes into production. The most AI-fluent people in the region have not removed the checking step. They have moved their work into it.
Agoda’s chief technology officer, Idan Zalzberg, put the cost in two sentences: “The real cost is not simply the model call. It is the work required to make AI trustworthy in production.”
The same point from the other side of the table
A separate piece of research, published in MIT Sloan Management Review on 29 September, arrives at the same place from a different direction.
It opens with Alturas Capital Partners, a real estate investment firm whose in-house lawyers used to rely on outside counsel for much of their lease work. With generative AI they now do that work internally, “saving the firm hundreds of thousands of dollars in spending and compressing lease work that once stalled for weeks.”
The article is written for the outside firms losing that work. Its advice is that they cannot win it back by insisting on their expertise. They have to compete on cost, on being easier to buy from, and on “absorbing the quality-assurance burden clients would otherwise carry themselves.”
Read that last phrase from the client’s side. When Alturas brought the lease work in house, it did not only bring in the drafting. It also brought in the checking that the law firm used to do and to answer for. The saving is real, and so is the new job.
What this means for a business case
Most proposals for an AI tool are costed on two lines: the licence, and the hours saved. The checking appears on neither.
That is how a project can save seven hours a week on paper and still disappoint. The review time was always going to be there. It was just never written down, never assigned, and never protected when the team got busy.
When nobody owns the checking, it gets skipped. The first sign is usually a mistake that reaches a customer.
The developers in Agoda’s survey are not checking because the tools are poor. Sixty-two per cent say AI-generated code is usable often or almost always. They check because “usable most of the time” is not the same as “safe to send without looking”, and they have built the looking into how they work.
Checking is a skill, not a formality
Some of the hours saved come straight back as review, and review done badly is worse than none. It produces a tick in a box and a false sense that someone looked.
Three things make review work in practice. The reviewer knows the subject well enough to spot an error that looks plausible. They have the source to compare against, not just the output. And they have enough time that reading carefully is the expected thing rather than the heroic one.
That is why the checking cannot simply go to whoever is most junior. A junior colleague can catch a typo. Spotting a figure that is plausible but wrong usually takes someone who knows what the right figure looks like.
The Agoda survey points the same way from another angle. Only 38% of respondents consider their codebase ready for full autonomy. The people closest to the work are the least willing to stop looking at it.
Three things to write down before you start
How long the checking takes. Take one task your team already does with AI, or plans to. Time the review honestly for a week. That number belongs in the business case, beside the licence fee.
Whose job it is. Not “the team”. A name, with time set aside. A check that belongs to everyone is done by whoever has a spare moment, which in a busy week is nobody.
What the check is for. Accuracy, tone, confidentiality, or a figure that must match its source. A reviewer who knows what they are looking for takes minutes. One who is asked to “have a look” either takes an hour or waves it through.
Being straight about the sources
Agoda is a travel booking platform. It uses AI rather than selling it, so it has no direct commercial interest in these findings, although the report also promotes its own engineering team. It was produced with Macramé Consulting.
The figures are percentages of developers and engineering leaders in seven Asian markets, not of companies or of all workers. The sample size and fieldwork dates come from coverage of the report rather than from Agoda’s release itself.
The MIT Sloan Management Review article is by José Parra-Moyano, Karl Schmedders, Olivier Laplace and Benjamin Torben-Nielsen. Part of it sits behind a paywall, and the Alturas saving is the authors’ description rather than an audited figure.
The teams furthest ahead on AI have not stopped checking. They have made it part of the job. A business case that leaves the checking out is not cheaper, only incomplete.