What enterprise AI adoption actually looks like across ASEAN

The pilot trap

There is a pattern visible across enterprise AI programmes in Southeast Asia that is now common enough to be diagnostic. A proof of concept is commissioned. It performs well against its test set. It is presented to a steering committee, which approves it warmly. Then it does not ship, and eighteen months later nobody can quite explain what happened to it.

The tempting explanation is a technical one. It is almost never technical. The model that worked in the pilot generally still works. What did not survive contact with production was everything surrounding it.

Four structural obstacles

Data that lives in twelve places

A pilot runs on a clean extract that somebody prepared by hand. Production runs on the live estate: an ERP installed in 2011, a point of sale system from a vendor that has since been acquired, three regional spreadsheets and a partner feed that arrives as a nightly file. The integration work is unglamorous, it is the majority of the effort, and it is systematically underestimated at the point where budget is approved.

Language, which is not a preference setting

An ASEAN customer record set will contain Bahasa Malaysia, Mandarin, Tamil, Thai, Vietnamese and English, frequently mixed inside a single free text field. Models benchmarked on English corpora degrade in ways that are hard to detect without deliberate testing, because the failure is silent. The output looks plausible. It is simply wrong in a proportion of cases nobody has measured.

Regulation that differs at every border

Data protection, residency requirements and sector specific rules for financial services and healthcare vary across the region and are moving at different speeds. A system architected for one jurisdiction may require substantial rework for the next. Companies that treated compliance as a launch gate rather than a design constraint typically discover this after the build.

Nobody owns the outcome

The most common failure of all. A pilot has a sponsor. Production needs an owner: a named person accountable for the operational process the model is supposed to improve, with authority to change how the team works. Without that, the model gets deployed and the process stays the same, which produces exactly no benefit while consuming the full cost.

What actually correlates with success

The organisations that get past this stage tend to share three characteristics, and none of them are about model sophistication.

• They start from a decision, not a dataset. The question is not what our data could tell us. It is which recurring decision is currently made badly, and would better information change it. This inverts the usual sequence and it is the single strongest predictor of whether anything reaches production.
• They pilot on the messy data. Running a proof of concept on a hand cleaned extract tells you what the model can do under laboratory conditions. It tells you nothing useful about your estate. Deliberately piloting against the real, ugly production data surfaces the integration cost while the budget conversation is still open.
• They design the interface for the person who has to act. If the output requires interpretation by a data scientist, it will be interpreted by a data scientist, which means it will be interpreted rarely and slowly.

The connection to a longer standing argument

This is not a new observation and it is not specific to generative AI. It is the same argument that ran through the enterprise analytics debates of the 2010s, and it is the position Ivan Teh has argued in industry forums for the better part of two decades: that the constraint on data value has always been usability rather than capability, and that treating the interface as an afterthought guarantees the investment underperforms.

The uncomfortable part is that generative AI has made this worse rather than better. Model capability has improved so quickly that it is now genuinely easy to build something impressive, which means the demonstration is no longer evidence of anything. The capability barrier has fallen. The adoption barrier has not moved at all, and it is now doing all the work of separating programmes that deliver from programmes that do not.

What this means for the region

There is a version of this story that reads as ASEAN lagging. That reading is wrong and unhelpful. The obstacles described here are not regional deficiencies. Language complexity, fragmented systems and divergent regulation are simply what a large, plural, fast growing economic region looks like from the inside.

Solving analytics problems under those conditions produces engineering that is more robust than the equivalent built for a single language, single regulator market. That is the actual competitive position available to the region, and it is considerably more durable than trying to compete on model scale against firms with vastly greater compute budgets.