Language was never the problem: Inside SEA’s real AI adoption gap

Ask most people what is holding back AI adoption in Southeast Asia, and the answer usually circles back to language. Bahasa Indonesia, Vietnamese, Thai and Malay are still treated as the great unsolved frontier for global models, the assumption being that once AI speaks the region fluently, enterprises will follow.

Kai Yong Kang, Partner at GenAI Fund, thinks that framing is out of date.
Global models, he argues, can already hold a conversation, translate, summarise and handle basic customer service across the region’s major languages. The real gap has moved elsewhere — into whether an enterprise can trust an AI system to run inside its business, securely and at scale.

Also Read: The AI revolution in emerging markets: Local models, global impact

“The gap is no longer simply whether AI can speak a Southeast Asian language,” he says. “It is whether an enterprise can trust it to execute a business process accurately, securely and at scale.”

That distinction sits at the heart of a conversation with Kang, whose firm has spent the past year running some of the region’s most active AI builder programmes, from a 3,000-strong buildathon with KFC Vietnam and Tasco, to a venture-building sprint with a toll-collection operator, to advisory work inside Vietnam’s National Assembly.

Where localisation quietly breaks down

Kang’s central argument is that most companies stop localising too early. They translate the interface and assume the job is done. Genuine localisation, he says, runs across three layers: language, culture, and operations, a layer which most companies skip. An AI system might understand exactly what a customer wants and still fail if it cannot pull the right information, apply company policy, or complete the action itself.

GenAI Fund got a close look at this problem through Agentic AI Build Week, a five-day buildathon it created that drew more than 3,000 registered AI builders producing over 400 solutions for enterprises including KFC Vietnam, Tasco and Guardian. A companion report the firm authored, The State of AI Builders in Southeast Asia 2026, drew on 2,719 approved builder registrations across 55 countries. Agentic systems made up roughly 18 per cent of project themes, followed by automation and workflows at 15 per cent, conversational AI at 10 per cent, and retrieval-augmented generation at 7 per cent.

The team that won KFC Vietnam’s F&B track, Twohearts, is Kang’s favourite illustration of the point. Long before the buildathon, the group had quietly been running a chunk of The Joi Factory’s delivery orders through Messenger and Zalo since 2020, giving them an intimate feel for how Vietnamese customers actually order and where a human needs to step in. They turned that muscle memory into an agentic ordering system that pulls live menu data, applies vouchers and loyalty points, confirms orders, pushes them to the point-of-sale system, and hands off anything complicated to a person.

“Teams with direct experience of a workflow often localise more effectively than teams approaching the problem as a purely technical or translation exercise,” Kang notes.

Who should be paying for language data, and who actually is

Southeast Asia’s language-data gap is often framed as a funding problem waiting for a government cheque. Kang’s view is messier: responsibility should be shared across governments, universities, technology companies and enterprises, because data alone was never going to be enough. What’s missing is less about corpora, more about repeatable mechanisms connecting datasets to real institutional problems.

Also Read: Featherless.ai wants to make AI model switching as easy as streaming Netflix

GenAI Fund’s engagement with Vietnam’s National Assembly is the case study he returns to. It began modestly, with an AI and Digital Parliament workshop in March 2025 where local AI startups demonstrated their tools to lawmakers. Within six months, that had turned into an actual deployment, with portfolio company Arcanic AI supplying the technology and telco Viettel backing the wider digital transformation.

By July 2026, a delegation led by Secretary General Lê Quang Mạnh was in Hong Kong studying AI governance and digital government at a programme GenAI Fund helped advise.

On the private-sector side, Kang points to Wash3000, a venture-building sprint GenAI Fund ran with VETC, part of Tasco Group and one of Vietnam’s largest electronic toll operators. Builders got access to live car-wash sites, mapping infrastructure from GoongIO, and VETC’s user base of more than four million people, proprietary context no public dataset could replicate.

The “wrapper” question, and why it misses the point

Southeast Asian founders building on foundation models are routinely dismissed as “just a wrapper.” Kang rejects the premise. Almost every modern software company sits on infrastructure someone else built — cloud, payments, maps — and using someone else’s model doesn’t make a product commercially thin by default. The GenAI Fund report backs this up: 81.7 per cent of participating builders use more than one AI platform, which tells Kang that model access itself has stopped being a moat. The value has migrated to the application layer above it.

Revve AI, a portfolio company building an AI customer-operations platform for contact centres across voice, email, Zalo and Facebook Messenger, is his working example of a defensible application layer. Its edge comes from omnichannel coverage of locally important platforms, deep integration with enterprise systems, a shared AI-and-human workspace for handoffs, and a no-code workflow builder with audit trails and version control.

Crucially, contact-centre managers can rewrite scripts and escalation rules themselves — the enterprise stays in the driver’s seat rather than depending on the vendor for every change. Vietnamese banks including VIB, VPBank and Sacombank are already running it.

“Enterprise customers do not pay for architectural purity; they pay for outcomes,” Kang says. “The real question is not whether a product is a wrapper, but what valuable layer it owns, and whether that value will remain as the underlying models continue to improve.”

If models stopped being the problem tomorrow

Kang doesn’t think foundation-model capability is where the fight will be won or lost much longer. Open-weight releases such as Moonshot AI’s Kimi K3 (a multimodal model with a one-million-token context window) are pushing capability once locked inside a handful of labs out into the open. Southeast Asian builders are already behaving accordingly: the GenAI Fund report found 27.5 per cent use at least one Chinese AI model, rising to roughly 39 per cent among those with three to five years of experience. They are picking models by task, not brand loyalty.

The real bottleneck, in his telling, is adoption — getting an enterprise to trust a system, integrate it properly, and rebuild a workflow around it. The report identified 661 builders already working inside large enterprises, even as 74.7 per cent of the broader community has fewer than two years of AI/ML experience. The talent is there; what’s missing is structured exposure to real problems.

Also Read: AI’s biggest bottleneck isn’t intelligence but fragmentation: i10X co-founder

GenAI Fund has built its business around plugging that gap in three ways: a matchmaking platform pairing over 3,000 AI startups with more than 150 enterprise use cases from companies including Coca-Cola and Shinhan Bank, which the firm says has helped drive over 500 proofs of concept; venture-building sprints like Wash3000 that hand builders real operating environments; and buildathons like Agentic AI Build Week that compress the distance between an enterprise problem and a working prototype into days.

For Kang, that’s the real competitive terrain going forward. “The next competitive advantage will not come from having access to the best model,” he says. “It will come from building the fastest and most reliable path from model capability to enterprise adoption.” In a region still being pitched language fluency as the finish line, that’s a considerably harder and more useful target to aim for.

The post Language was never the problem: Inside SEA’s real AI adoption gap appeared first on e27.