The conversation around AI voice has changed dramatically over the past year.

Not long ago, businesses wanted to know whether AI could hold a natural conversation. Today, that question has largely been answered. Modern voice agents can qualify leads, schedule appointments, resolve customer enquiries, and handle a growing range of routine interactions with remarkable fluency.

At Agora, we’ve noticed a corresponding shift in customer conversations. Businesses are no longer asking whether voice AI works. Instead, they want to know how well it performs when deployed across different markets, languages, and customer segments.

Industry data points in the same direction. Gartner found that 85 per cent of customer service leaders plan to explore or pilot customer-facing conversational AI in 2025, with 44 per cent specifically evaluating voice AI as part of their customer experience strategy.

As adoption accelerates, we’ve found that two assumptions frequently shape deployment decisions. Both deserve a closer look.

Misconception #one: Strong English performance means AI is ready for APAC

Many of today’s leading voice models achieve impressive performance in English. Demonstrations often showcase smooth, natural conversations that make the technology feel ready for immediate deployment.

Real customer conversations, however, are rarely that predictable.

Across Asia Pacific, people naturally switch between languages depending on the context of the conversation. A customer may discuss payment details in Bahasa Indonesia before mentioning a product feature in English. A caller in Singapore may move between English and Mandarin without thinking twice. Across Thailand, Vietnam, Malaysia, and the Philippines, regional accents, local vocabulary, and conversational habits add further variation.

These aren’t edge cases. They are everyday interactions.

For AI systems, however, these communication patterns introduce additional complexity. Speech recognition must accurately identify different languages, preserve context as conversations shift, and correctly interpret customer intent despite changes in pronunciation, vocabulary, or sentence structure.

This is one reason speech recognition continues to evolve. While recent advances have dramatically improved accuracy, real-world deployments still need to account for multilingual conversations, regional accents, background noise, and inconsistent network conditions that rarely appear in benchmark evaluations.

Performance in English, therefore, should be viewed as the starting point rather than proof that a voice agent is ready for every APAC market.

Also Read: You’re waiting for everyone to agree: The hourglass doesn’t care

Misconception #two: Supporting multiple languages is the same as localisation

Once businesses recognise the diversity of APAC, the next instinct is often to prioritise multilingual support.

Supporting more languages is certainly important. But localisation involves much more than expanding a language menu.

Customers who speak the same language do not necessarily communicate in the same way. Regional expressions, industry terminology, pronunciation, and code-switching all influence how conversations unfold. A system that performs well in one market may require further adaptation before delivering the same experience in another.

Research from Microsoft reinforces this point. The company found that multilingual users naturally switch between languages during conversations and respond more positively to conversational AI that adapts to those shifts instead of remaining rigidly monolingual.

Customer expectations reinforce the need for localisation. According to CSA Research, 76 per cent of consumers prefer buying products with information in their own language, while 88 per cent of Indonesian consumers prefer content presented in Bahasa Indonesia. Although the study focused on digital content, the same principle applies to voice interactions. Customers expect communication to feel natural, not translated.

Ultimately, customers don’t evaluate AI based on the number of languages it supports. They evaluate whether the conversation feels effortless. If they have to repeat themselves, avoid certain phrases, or adjust the way they naturally speak, the interaction becomes less effective regardless of how sophisticated the underlying model may be.

What businesses are prioritising now

One of the most noticeable changes we’ve seen is how conversations with enterprise customers have evolved.

A year ago, many discussions centred on whether voice AI could realistically replace traditional IVR systems or automate routine enquiries. Today, those questions have become far more operational.

Businesses want to understand how quickly voice agents can be adapted for new markets, how they perform across multilingual contact centres, how they integrate with existing customer workflows, and how consistently they serve customers who communicate differently from one another.

That shift reflects the broader maturity of the market. Organisations are moving beyond experimentation and focusing on deployment quality. Success is no longer measured by whether a voice agent can complete a demonstration. It is measured by whether it can deliver a consistently positive customer experience across thousands of real conversations.

Also Read: What AI safety researchers actually worry about

The next competitive advantage won’t be better voices, it will be better understanding

As foundation models continue to improve, the gap in conversational quality between voice AI platforms is likely to narrow. Natural-sounding speech will increasingly become an expected capability rather than a differentiator.

The next competitive advantage will come from understanding customers more effectively.

For businesses operating across Asia Pacific, that means recognising that localisation is not simply another feature to enable before launch. It is becoming a core deployment strategy that determines whether AI creates friction or removes it.

The organisations that succeed with voice AI will not necessarily be those that automate the greatest number of calls. They will be those that build voice experiences around the realities of how their customers communicate, market by market, language by language, and conversation by conversation.

As AI phone calls become a standard part of customer engagement, understanding people may prove just as important as understanding speech.

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post The localisation gap: Why multilingual AI isn’t enough for APAC markets appeared first on e27.

There’s a number that should reframe every stablecoin policy debate in the region, and it isn’t flattering. As of early 2026, the global stablecoin market is worth north of US$300 billion, and about 99.76 per cent of it is backed by the US dollar. Non-dollar coins, every euro, yen, ringgit and Singapore dollar experiment combined, split the remaining quarter of one per cent.

For two years, the region’s answer to that has been the same: write a better rulebook. Singapore built an open, multi-currency licensing regime under the Monetary Authority of Singapore and drew in issuers like StraitsX, Paxos and Circle. Malaysia is running the region’s most-watched ringgit stablecoin pilot, bank-anchored and Shariah-inclusive.

These are thoughtful pieces of policy, and the instinct behind them is right: digital money is becoming a question of sovereignty, and sovereignty is worth defending.

But a rulebook and a market are different things, and we keep confusing the two. That 99.76 per cent isn’t a gap in anyone’s licensing regime, but a verdict of sorts. Users, exchanges and treasurers have already chosen, and they chose the dollar, for its liquidity, its ubiquity across every wallet and venue, and the plain fact that it’s what everyone else is already holding.

You don’t legislate your way out of a network effect, and you can’t simply license a default into existence.

It’s worth looking at who set that default, because it wasn’t an accident. When the United States passed the GENIUS Act in July 2025, it did something the region should study closely. The law requires payment stablecoins to be fully backed one-to-one by dollar assets, which sounds like consumer protection and works like statecraft.

Every compliant token routes fresh demand into US Treasuries and widens foreign access to dollars. Washington understood that this contest isn’t won in the rulebook, it’s won in the reach. Whoever owns the default token exports their currency with it. China clearly agrees; it’s now drafting a yuan-stablecoin roadmap of its own. The big players are treating this as a distribution war. We’re treating it as a compliance exercise.

That’s the mismatch. A licence authorises a product; it doesn’t give anyone a reason to hold it. The best-regulated ringgit stablecoin in the world still has to compete against a dollar token that’s already in every wallet, already trusted, already the path of least resistance. In fintech circles, you hear the market’s indifference described, very politely, as “user preference for usability,” which is just a nice way of saying nobody cares where a coin was issued. Sovereignty on paper isn’t sovereignty in wallets, and no amount of regulatory craft closes that gap on its own.

The fair objection is that retail adoption may be beside the point. Maybe local-currency stablecoins aren’t built for consumers at all, but for business, cross-border settlement, remittances, corridor flows where banking relationships and regulation matter more than what sits in a shopper’s phone.

Also Read: Japan shows how non-USD stablecoins complement USDC and USDT

It’s a reasonable argument, and it’s partly true. But it doesn’t rescue the licensing-first strategy; it just moves the same problem upstream. The dollar’s incumbency in settlement is exactly the thing a regional coin has to dislodge, and incumbents don’t fall to frameworks. Retail or wholesale, it’s a battle for the default, and defaults are won on reasons to switch, not rules to comply with.

Here’s where the region’s real advantage has been hiding in plain sight, and where I think a marketer reads this problem differently than a regulator does. Southeast Asia doesn’t lack rails or rulebooks.

What it has, that almost no one else does, is distribution it already owns, the wallets and QR systems hundreds of millions of people open every day without thinking, from QRIS to PromptPay to the super apps that have quietly become default infrastructure. That’s the asset.

A local-currency stablecoin embedded as the native rail inside systems people already trust isn’t asking anyone to make a patriotic choice; it’s making the local option the easy one. Add corridors where a regional coin is genuinely cheaper and faster than a dollar round-trip, and you start giving people, retail and treasury alike, a concrete reason to switch that a licence never could.

None of this means the frameworks were wasted. They’re the floor. But a floor isn’t a strategy, and we’ve been mistaking one for the other, polishing the rules for money that keeps flowing in someone else’s currency. The awkward part is that the licensing is the easy bit. The hard part, the part that actually decides sovereignty, is distribution and trust. And that’s the part nobody’s resourcing.

Also Read: Taiwan’s stablecoin moment: Why the NTD could outshine the dollar

So the question isn’t whether the region can regulate stablecoins well. It plainly can. The question is whether it intends to contest the default itself, to fight for the thing people reach for first, or keep drafting careful rules while the digital dollar wins by simply being everywhere it already is.

On the present course, we’re reacting to a standard the dollar set and the GENIUS Act is now actively defending. Being part of rewriting the rules means fighting a battle that a rulebook, on its own, was never going to win.

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post Southeast Asia can’t simply license its way to stablecoin sovereignty appeared first on e27.

For years, artificial intelligence in games has mostly meant enemies that chase, dodge or shoot with varying degrees of believability. Razer and the National University of Singapore now want to push that idea into more demanding territory: AI systems that understand context, adapt in real time and eventually behave less like scripted software and more like responsive companions inside digital worlds.

The gaming hardware and software company has partnered with NUS Computing to establish the Razer-NUS Joint AI Research Lab, a Singapore-based research initiative focused on AI for gaming across hardware, software and services. The lab will be hosted on the NUS campus and will bring academic research closer to Razer’s product and engineering teams.

Also Read: AI in gaming: How Southeast Asia became the testing ground for virtual companions

The collaboration comes as gaming companies globally are trying to work out what generative AI and adaptive models can usefully do beyond producing concept art, dialogue or non-playable character chatter. For Southeast Asia, where Singapore has been positioning itself as a regional hub for AI research, game development and digital talent, the lab is also a sign that gaming is being treated less as a niche entertainment category and more as a serious testbed for real-time AI systems.

Razer, which is dual-headquartered in Irvine, California and Singapore, already operates AI Centres of Excellence in Singapore and France. The company says its AI lab network now includes close to 100 AI researchers, data scientists and engineers. The new NUS partnership is designed to add a more research-heavy layer to that ecosystem.

Why gaming is a hard AI problem

At the centre of the partnership is a research area Razer and NUS are calling Gaming Artificial Narrow Intelligence, or GANI. In plain terms, this refers to AI models built specifically for interactive digital environments rather than broad, general-purpose systems.

That distinction matters. Games are not static documents or one-way media. A model deployed inside a game must respond instantly to a player’s movement, strategy, skill level and decisions. It must also work within the technical constraints of live gameplay, where lag or awkward behaviour can quickly ruin immersion.

Razer and NUS say GANI could eventually power AI teammates that adjust tactics to a player’s style, generate objectives and dialogue in real time, and change difficulty during live matches. If done well, such systems could make games feel less repetitive and more personal. If done poorly, they could create unfair, unpredictable or intrusive experiences.

“Gaming presents some of the most demanding environments for AI, requiring systems to respond in real time, adapt to changing contexts, and enhance the player experience,” said Li-Meng Lee, Chief Strategy Officer at Razer. He added that breakthroughs in this area could also be relevant to education, simulation and other real-time digital experiences.

That wider applicability is an important part of the story. AI that can operate smoothly in a fast-paced multiplayer game may also be useful in training simulations, virtual classrooms, industrial interfaces or digital assistants that need to react to changing human behaviour.

From lab work to products

The Razer-NUS Joint AI Research Lab will focus on three areas: core model innovation, real-time content systems and advanced personalisation. The two partners plan to test research in both simulated and live gameplay environments, with successful outcomes potentially integrated into Razer’s proprietary systems.

Also Read: AI and the rise of gaming entrepreneurs

NUS Computing will lead research and talent development, while Razer will provide industry use cases, engineering access and commercialisation pathways. Associate Professor Ooi Wei Tsang from NUS Computing’s Department of Computer Science will serve as Director of the joint lab.

“Universities play a crucial role in advancing scientific discovery and shaping the evolution of emerging disciplines,” said Ooi. He said the partnership gives GANI “a dynamic testing ground” by combining NUS’s AI research strengths with access to Razer’s live engineering environment.

For Razer, the lab also supports ongoing AI projects such as Razer AVA and Project Motoko. AVA is the company’s concept for a digital human companion that can offer personalised interactions through memory, personality, contextual awareness and adaptive behaviour. Project Motoko is a wearable AI headset intended to bring multimodal generative AI into an everyday form factor.

Both ideas remain ambitious. Digital companions have often struggled with trust, usefulness and emotional authenticity, while wearable AI devices have faced questions around privacy, battery life and whether consumers truly want another always-on device. The research lab could help Razer test these assumptions before they become mass-market products.

The Southeast Asian angle

Singapore has been steadily building its position as Southeast Asia’s AI nerve centre, backed by state funding, university research, enterprise adoption and multinational technology partnerships. NUS, in particular, has become a key node in this strategy, with its School of Computing working across AI, cybersecurity, data science and digital trust.

The Razer-NUS lab fits neatly into that broader national agenda. Unlike pure software AI startups, gaming sits at the intersection of chips, devices, cloud infrastructure, creative production and consumer behaviour. That makes it a useful sector for training technical talent and spinning out applied research.

Southeast Asia also has a large and young gaming population, even if much of the region’s game development activity remains fragmented. Mobile gaming dominates markets such as Indonesia, the Philippines, Vietnam and Thailand, while Singapore has served as a regional base for publishers, esports organisers and technology firms. If AI can reduce content production costs or help smaller studios build richer game worlds, the impact could extend beyond Razer’s own ecosystem.

Still, the commercial upside will depend on whether research can be translated into tools that developers actually use. Game studios are cautious about technology that adds complexity to production pipelines or creates unpredictable player experiences. AI features must improve gameplay, not simply signal that a company is keeping up with the latest trend.

A crowded field

Razer is not the only company trying to define the future of AI in gaming. NVIDIA has been pushing AI-powered game characters through its ACE technology, while Microsoft has explored AI tools across Xbox, cloud gaming and developer workflows. Sony, Tencent and NetEase all have deep gaming interests and the resources to experiment with AI at scale.

On the software side, Unity and Epic Games are building AI capabilities for developers, while startups such as Inworld AI focus on AI characters and interactive storytelling.

Razer’s traditional rivals in gaming hardware, including Logitech G, Corsair, SteelSeries and ASUS ROG, are also exploring ways to make peripherals and gaming systems more intelligent. The difference is that Razer is trying to connect hardware, software, services and AI research under one umbrella. Whether that becomes a defensible advantage will depend on execution rather than lab announcements.

The market backdrop is attractive. The global AI in gaming market is projected to grow from US$4.2 billion in 2025 to US$66.8 billion by 2035, representing a 32 per cent compound annual growth rate. A recent survey cited by Razer found that 79 per cent of gamers are receptive to at least one AI-enabled feature they believe would improve their experience.

Those numbers explain why companies are moving quickly. But gaming audiences can be unforgiving. Players tend to welcome AI when it makes games more immersive, responsive or fair. They reject it when it feels like a shortcut, a gimmick or a way to replace human creativity.

Also Read: Gaming as the next social network: How Gen Z and Gen Alpha are redefining digital belonging

For Razer and NUS, the opportunity is to show that AI in gaming can be more than automated content generation. If GANI can produce systems that understand players without overwhelming them, Singapore could become an important base for a new category of interactive AI research.

The harder task now is proving that intelligence built in the lab can survive contact with real players.

The post Razer and NUS launch Singapore AI lab to rethink how games respond to players appeared first on e27.

In the second week of August 2009, four months into my first job as an R&D engineer at a pulp and paper plant in Riau, I told my supervisor I was leaving to join a bank. I was 23. I had spent four years studying chemical engineering, graduated at the top of my department, and worked my way into the bleaching process team at one of the country’s largest paper producers. I had also concluded that what I wanted to do for the next four decades was sit on the other side of how risk moves through systems, and that finance was where that work lived.

My supervisor told me, more politely than I deserved, that I was making a mistake.

I am now 15 years into a risk career that has run through three banks, an insurer, and eventually my own company. In every one of those seats, the skills I have leaned on hardest were not the ones I picked up in business school or in risk certification cycles. They were the ones I learned in a chemical engineering lab in North Sumatra.

This is not a coincidence. The principles that govern how chemical plants work happen to be the same principles that govern how financial systems work. The finance industry, by and large, has not noticed.

What I did not know I was learning

The chemical engineering curriculum I went through at Universitas Sumatera Utara was, on the surface, about reactors, separation processes, mass transfer, and process control. What it actually trained was a way of seeing systems.

The discipline forces you to think in terms of inputs, outputs, transformations, and constraints. Nothing inside a chemical process is allowed to be vague. If you do not know where a stream goes, you do not understand the plant. If you cannot predict how the system will behave when one variable changes, you cannot operate it safely. The mental habit this builds, the refusal to leave any part of a system unaccounted for, turns out to be the same habit you need to run a risk function inside a bank.

Also Read: Southeast Asia’s fintech apps don’t have a literacy problem, they have a fear problem​

Three principles that transferred

Three things in particular have stayed with me, and have shaped every risk decision I have made since.

Mass balance. In a chemical process, every kilogram of material in must equal every kilogram out, less what is accumulated or transformed inside the system. There are no unexplained gaps. The reflex this builds, that an unbalanced ledger is a mis-measurement, not a mystery, is the same reflex that catches operational losses, fraud patterns, and capital gaps inside a bank. A risk officer who instinctively believes that what comes out of a system must equal what went in, less what was retained, asks the right questions almost without thinking.

Process control and feedback. In a well-designed chemical process, the system measures itself continuously and adjusts in real time. The temperature drifts above the setpoint, a valve closes. The pressure spikes, a relief opens. There is no quarterly review committee. The system corrects, or it ruptures. This way of thinking about feedback, small adjustments made continuously against a known boundary, is exactly what most financial risk frameworks still do not do. They review periodically. They escalate sporadically. They operate, in engineering terms, like a reactor with no instrumentation.

Failure mode and effects analysis. Before any chemical plant goes live, the engineering team systematically maps every way the process can fail, ranks the failure modes by likelihood and consequence, and designs the controls before the failure happens. The discipline of asking “what would have to be true for this to go badly?”, not as anxiety, but as method, is the single most useful habit I brought into risk management. Most of the loss events I have seen across two decades were predictable inside a competent failure analysis. Most of them did not have one.

Where the finance training falls short

Finance has its own analytical apparatus. Modigliani-Miller, Black-Scholes, value-at-risk, the architecture of modern asset pricing. These are powerful tools inside the assumptions they were built for. Outside those assumptions, they are quieter than their reputations suggest.

The gap I have noticed, across many years of working with both engineers who entered finance and finance professionals who stayed in finance, is this. The engineer asks first: what could break this system, and what would be true if it did? The finance professional asks first: what is the expected outcome, and what is the variance around it? Both questions matter. But in a crisis, and risk management is the discipline of crises, the engineer’s question is the one that saves the institution.

Also Read: How do you finance a first nuclear reactor for a data centre? The deal structure is finally coming together

What this means for anyone choosing a non-traditional path

I get asked, perhaps once a month, by an engineering student whether they should leave technical work for finance, strategy, or consulting. My answer has become consistent.

The pivot is not the question. The skills you have built in engineering are some of the most transferable skills any discipline produces. You can move into finance, and your engineering training will quietly do half the work of your new role. The question is whether you are leaving for the right reason, because the systems you want to understand are now financial systems, not chemical ones, or because you think the new field will be more glamorous than the one you trained in.

If it is the first reason, the move is sound. If it is the second, no field will deliver what you are hoping for.

I do not draw flowsheets on whiteboards. I do not solve heat transfer equations. But the way I think about risk, about systems, balances, feedbacks, failures, was shaped by four years at USU and four months in Riau before I ever opened a banking textbook. 15 years later, that training is still doing more of the work than anything I learned afterwards.

The most useful risk management education I have ever received was a chemical engineering degree. It just took me a decade to fully realise it.

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post The principles that govern both chemical plants and financial systems appeared first on e27.

A great deal of bad strategy begins with a comforting number.

Market share is one of the most over-trusted measures in business because it looks like proof of strength while often revealing very little about actual control. It tells you how much of the market you currently touch. It does not tell you who sets the terms, who shapes behaviour, who captures the best economics, who sees demand first, who becomes hard to route around, or who gets stronger when everyone else grows.

That distinction matters more than most leadership teams admit.

A company can have impressive market share and still be structurally weak. It may be large but replaceable. It may serve many customers without controlling any decisive part of the system. It may be visible in the market but absent from the points where pricing power, dependency, switching cost, regulatory comfort, workflow design, or standards actually get determined. In that situation, share creates exposure more than power. The company has more revenue to defend, more cost to carry, and more surface area to lose.

Power comes from something else.

Power comes from control points.

By control points, I mean the parts of a market that others must pass through, design around, conform to, or receive permission from. These are the places where choice narrows, dependence increases, economics concentrate, and leverage becomes durable. Control points are not always the biggest part of the value chain. In many markets they are the smallest visible layer and the most important strategic position.

Market share measures presence, control points determine terms

This is the first distinction serious strategists need to make.

Market share answers the question, how much of the market do we currently serve? Control points answer a much more important question: under what conditions does the market operate, and how much influence do we have over those conditions?

That is a harder question because it forces leaders to examine where actual leverage sits. Does the company control distribution? Does it control customer identity? Does it control switching friction? Does it control access to demand? Does it control compliance interpretation? Does it control the data that trains the system, validates performance, or proves value? Does it control the workflow where alternatives become painful? Does it control the commercial mechanism through which everyone else gets paid?

These are very different positions from simply being widely used.

Also Read: AI-powered business automation: How SMEs are transforming operations in Southeast Asia

The real contest in markets is usually over choke points, not customers

We often describe competition as a fight for customers, but that is usually only the surface-level view. Underneath that visible contest is another one. Companies are competing to own the choke points that shape how customers are acquired, how products are integrated, how risk is managed, how spending is justified, and how alternatives are compared.

This is where strategic thinking gets more interesting.

A control point may sit in onboarding, where identity and trust are established. It may sit in the workflow, where staff do not want to relearn behaviour. It may sit in the reporting layer, where leadership sees value and performance. It may sit in compliance, where approval becomes easier for one route than another. It may sit in the commercial structure, where procurement can buy one thing cleanly but struggles to buy the alternative. It may sit in data custody, where the history required for tuning, insight, and continuity quietly accumulates in one place.

None of these is glamorous in the way market share is glamorous. But they are often far more consequential.

Control points are often hidden inside boring functions

One reason leaders miss control points is that they expect power to sit in obvious places. They look for power in brand visibility, revenue scale, installed base, or category leadership. They do not always notice that durable influence is often buried in functions that appear mundane.

Billing can be a control point. Identity can be a control point. Audit records can be a control point. Procurement approval paths can be a control point. Data lineage can be a control point. Technical certification can be a control point. Distribution rights can be a control point. Default settings can be a control point. Even complaint handling can become a control point if it determines who the institution trusts when something goes wrong.

These positions rarely get celebrated in market narratives because they are not as exciting as product innovation or growth curves. Yet they are often where strategic reality lives.

A company that owns a boring control point can quietly become impossible to displace. Everyone else may appear more dynamic, more loved, or more talked about. But when the market has to choose under pressure, the firm sitting inside the operational necessity tends to win.

Also Read: Why AI-empowered teams are getting smaller, and why that is harder than it sounds

The most valuable control point is often the one that feels legitimate

Not every choke point becomes durable power. Some create resistance, regulatory backlash, or market workarounds. The most defensible control points are usually the ones that feel justified by the system rather than artificially imposed on it.

This matters a great deal.

A control point lasts when participants accept that it serves a real function. It reduces uncertainty. It simplifies coordination. It lowers risk. It improves trust. It makes the system easier to govern. It creates a common language for decision-making. It becomes part of how the market keeps itself stable.

This is why legitimacy matters more than mere friction.

An artificial barrier can generate temporary leverage, but a legitimate control point generates embedded authority. Participants may not love it, but they recognise that the market works better with it than without it. Once that happens, the control point stops feeling like an advantage and starts feeling like infrastructure.

That is when strategy becomes hard to attack.

Strategy is not only about getting chosen, it is about becoming hard to route around

That is the deeper idea underneath this whole argument.

Many strategies are built around being selected again and again. That is fine in open competition, but it is exhausting and fragile if every decision resets the contest from the beginning. Truly strong strategic positions do something else. They reduce the frequency with which choice is genuinely reopened.

This does not always mean lock-in in the crude sense. It can mean being embedded in the reporting layer where value is measured. It can mean being the trusted source of operational truth. It can mean becoming the easiest path through governance. It can mean owning the transition cost. It can mean sitting where multiple parties coordinate. It can mean controlling the evidence required to compare alternatives fairly. It can mean becoming the familiar answer in moments of uncertainty.

These are all forms of route control.

Once a firm occupies that place, competitors may still exist, customers may still express dissatisfaction, and market share may still move at the edges. But the company remains difficult to route around because it has become part of the operating logic of the system itself.

That is much closer to power than popularity ever is.

—

The post Market share is not power, control points are appeared first on e27.

Southeast Asia sits at an inflection point. The region commands 10 per cent of global GDP, controls critical supply chains, and possesses over 500 million people, a workforce larger than the European Union. Governments from Singapore to Japan have committed tens of billions to AI development. Training programmes have created millions of “AI-skilled” professionals. Investment is flowing. Ambition is high.

Yet the region remains a consumer of AI technology rather than a builder of AI leadership.

The gap isn’t ambition. It’s infrastructure. Specifically, the infrastructure to identify, verify, and rapidly deploy genuine AI capability at scale. Southeast Asia has solved the training problem. It hasn’t solved the capability verification problem. And that distinction is costing the region billions in unrealised opportunity.

The paradox: Trained workforce, unverified capability

Consider what Southeast Asia has accomplished in workforce development. Singapore alone has trained over 555,000 workers through SkillsFuture programmes. Indonesia, Vietnam, and other regional economies have launched comparable initiatives. Japan’s government has committed to ¥10 trillion (US$63.4 billion) in Trustworthy AI investment by 2030, explicitly supporting workforce development. South Korea, Taiwan, and the broader APAC region are following similar trajectories.

The scale is impressive. The investment is real. The problem is also real: credentials don’t predict capability.

Research is unambiguous on this point. Cloud Range’s 2025 analysis of technical workforce readiness found: “Knowledge is what you learn. Readiness is what you can perform, and those are not the same. In a live incident, the difference between knowing what to do and being able to execute in real time under uncertainty is dramatic.”

Google discovered this through hiring at scale. After years of screening candidates using transcripts, GPAs, and certifications, the company concluded these credentials were essentially “worthless” for predicting actual job performance. Only 43 per cent of people in STEM roles even have STEM degrees, yet these roles are filled regardless.

In Southeast Asia, the signal integrity problem is regional. Hiring challenges across the region are driven by “skills specificity rather than qualification mismatches,” meaning organisations aren’t struggling to find credentialed people; they’re struggling to find people with demonstrated expertise in the specific capability required. This distinction, credential versus demonstrable capability, is the invisible ceiling on regional AI leadership.

Also Read: The localisation gap: Why multilingual AI isn’t enough for APAC markets

The infrastructure gap: Why Southeast Asia can’t deploy capability

Three infrastructure deficits are blocking Southeast Asia’s path to AI regional leadership.

First, no reliable capability signals. When Singapore employers report that 24.3 per cent experience skills gaps, and 49.9 per cent say this causes increased workload for other staff, they’re describing a system where trained people can’t actually perform.[6] The problem isn’t training quality. The problem is that the system has no way to verify whether trained people possess the capability they’re supposed to have. Certifications and credentials are issued; capability remains unverified.

Second, assessment infrastructure doesn’t exist at regional scale. Enterprise hiring relies on either credentials (which don’t predict capability) or expensive, time-consuming assessments (which only large enterprises can afford). SMEs, which represent 98 per cent of Southeast Asian businesses, have no middle ground. They can’t afford individual assessments. They can’t trust credentials alone. Result: hiring becomes a guessing game.

Third, deployment pathways are unclear. Even when organisations identify capable talent, they lack systematic frameworks for rapid deployment. In Japan, 16 per cent of specialised professional, manager, executive, and technician (PMET) roles remain unfilled for six months or longer, not because capable people don’t exist, but because organisations can’t verify who is actually capable and move them into roles quickly. This deployment friction slows regional AI adoption.

These three gaps – signal integrity, regional assessment infrastructure, and deployment speed – are the hidden constraints on Southeast Asia’s AI leadership ambitions.

The framework: Capability-by-doing vs capability-by-credential

Building regional AI leadership requires distinguishing between two fundamentally different types of assessment.

Capability-by-credential is what currently exists. A person completes training, passes a test, receives a certificate. The credential signals that they completed the training program. It does not signal that they can actually perform the work under real conditions.

Capability-by-doing is what the region needs. This means assessing whether someone can actually do the work: solve novel problems, integrate with existing systems, perform under pressure, teach others. Capability-by-doing requires more than training completion; it requires evidence of actual performance ability.

The distinction transforms everything. Organisations moving from credential-based to capability-based assessment make fundamentally different hiring decisions. They identify hidden capability that credentials miss. They avoid hiring people whose credentials exceed their actual ability. They deploy talent faster because they know what people can actually do.

For Southeast Asia specifically, capability-by-doing assessment is the infrastructure gap that, once closed, unlocks regional leadership.

Also Read: You’re waiting for everyone to agree: The hourglass doesn’t care

How AI-powered capability assessment changes the game

Closing this gap requires infrastructure that didn’t previously exist. This infrastructure has three layers.

• Layer one: Deterministic signals. Traditional hiring captures keywords and semantic understanding, does someone know the vocabulary and concepts required? This is necessary but insufficient.
• Layer two: Demonstrated understanding. Can the person explain concepts in their own words and apply frameworks to new situations? This shows deeper mastery than credential completion.
• Layer three: Capability reasoning. This is where AI transforms the game. Using advanced language models specifically trained for capability assessment, organisations can now ask: Given this person’s demonstrated knowledge, work history, and reasoning, can they actually perform this role? Can they solve novel problems? Will they perform under pressure?

Layer three is what currently doesn’t exist at regional scale. It’s the infrastructure gap that prevents Southeast Asia from rapidly identifying and deploying genuine AI capability. It’s also the infrastructure that, once deployed, enables organisations to stop hiring based on credentials and start hiring based on demonstrated capability.

Building regional tech leadership: The path forward

Southeast Asia’s path to AI regional leadership requires three simultaneous moves.

First, governments must shift measurement metrics. Current training programmes measure completion rates and credential issuance. Governments should measure capability verification: Of people certified in AI skills, what percentage can actually perform AI work? This single metric shift transforms incentives across the entire training ecosystem.

Second, organisations must adopt capability-based hiring. Rather than filtering for credentials, organisations should assess demonstrated capability before hiring. This applies to SMEs as much as enterprises; capability assessment infrastructure must be affordable and accessible at regional scale.

Third, the region needs shared assessment infrastructure. Just as SkillsFuture created shared training infrastructure, Southeast Asia needs shared capability assessment infrastructure. This infrastructure should:

• Verify demonstrated AI capability (not just credential ownership)
• Be accessible to SMEs and enterprises alike
• Operate across national boundaries (Singapore, Japan, Korea, Taiwan, Vietnam, Indonesia)
• Enable rapid talent deployment across the region
• Create regional, verifiable signals about who can actually do AI work

This infrastructure is the missing piece. Without it, Southeast Asia remains talent-rich but capability-constrained. With it, the region becomes a global AI leadership centre.

Also Read: What AI safety researchers actually worry about

The competitive imperative

The timeline matters. China is rapidly building AI manufacturing and infrastructure capabilities. The US dominates AI research and development. India has established AI services and outsourcing leadership. Europe is building AI regulation and compliance infrastructure.

Southeast Asia’s opportunity is different: become the global leader in identifying, verifying, and rapidly deploying AI talent at scale. This isn’t about competing on research (the US leads) or manufacturing (China leads). It’s about building the human infrastructure that turns global AI innovation into global AI value creation.

But this opportunity has a window. Early movers in regional capability assessment infrastructure will define regional AI leadership for the next decade. Organisations that adopt capability-by-doing assessment now will have access to the best talent, fastest deployment, and highest competitive advantage as regional AI adoption accelerates.

Building the foundation

Southeast Asia has the talent pool. The region has the investment. The region has the government commitment and regulatory tailwinds. What the region lacks is the infrastructure to verify capability and deploy it effectively.

Building this infrastructure is the single highest-impact investment Southeast Asia can make in regional AI leadership. It’s not about training more people. It’s about finally being able to tell who is actually capable and moving them into roles where they can create value.

Regional AI leadership isn’t a function of how many people you train. It’s a function of how effectively you identify the truly capable and deploy them at scale. That’s the infrastructure gap Southeast Asia must close. That’s the foundation regional tech leadership requires.

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post Southeast Asia’s AI talent and infrastructure: Building the foundation for regional tech leadership appeared first on e27.

For much of the last decade, “going global” was almost a rite of passage for startups. The formula seemed straightforward: build a product that solves a real problem, raise capital, expand into new markets, and let globalisation create the scale. Technology was increasingly borderless, capital flowed relatively freely, and businesses assumed that the world would become progressively more interconnected.

Over the past two years, however, that assumption has quietly changed. Trade tensions, semiconductor restrictions, sovereign AI strategies, and shifting investment priorities have begun reshaping the environment in which businesses operate. These developments often make headlines as geopolitical events, but for founders, they translate into very practical questions about where to build, who to partner with, and how to grow. I have not had to abandon international ambitions, but I have had to redefine what “going global” actually means.

The most significant business decision I made in the past year was not entering a new market. Instead, it was deciding to invest more deeply in building intelligence into my product before pursuing expansion. I founded a platform that tracks corporate events and strategic actions across listed companies, helping investors and business professionals understand developments that may influence investment decisions. Initially, the objective was relatively simple: make corporate information more searchable and easier to access. However, as I continued analysing thousands of announcements, I began noticing that many corporate decisions were no longer driven solely by commercial considerations.

Increasingly, companies were relocating manufacturing facilities because supply chains had become politically sensitive rather than simply cost inefficient. Strategic partnerships were being formed to satisfy local ownership requirements. Investments in semiconductor manufacturing, artificial intelligence infrastructure, renewable energy, and critical minerals were often influenced as much by national industrial policies as by market demand. Even financing decisions reflected changing global capital flows, with companies seeking investors whose strategic priorities aligned with shifting geopolitical realities.

This changed the direction of my own product. Rather than becoming another searchable database of announcements, it became increasingly important to explain why each corporate event mattered. Every announcement now carries additional context through impact assessments and strategic analysis, helping users understand not only what happened, but also what broader forces may have influenced the decision. As geopolitical fragmentation accelerates, I believe context has become significantly more valuable than information alone. Data is increasingly abundant; interpretation is becoming the real differentiator.

Also Read: The new map of growth: Where Southeast Asia fits in a fragmented world

These changes also made me rethink international expansion. In the past, founders often assumed that once a product worked well domestically, it could simply be replicated across multiple countries. Today, that assumption is becoming increasingly difficult to sustain. Different jurisdictions are introducing their own approaches to data residency, AI governance, cloud infrastructure, and digital sovereignty. Products that appear technically identical may require entirely different operating models depending on where they are deployed.

For businesses building AI-powered products, these considerations are no longer technical details left for engineers to solve after expansion. Decisions about where customer data is stored, which AI models can be deployed, which cloud providers are used, and how information is processed increasingly become strategic business decisions. Governments across Asia are investing heavily in sovereign AI capabilities and placing greater emphasis on domestic digital infrastructure. As founders, we therefore need to think about regulatory compatibility from the beginning rather than treating localisation as a problem to solve after entering a new market.

At the same time, I do not believe geopolitical fragmentation is purely a constraint. It also presents opportunities, particularly for Southeast Asia. For decades, the region has benefited from maintaining economic relationships with both East and West. While major economies are increasingly competing for technological leadership, Southeast Asia continues to occupy an important middle ground. Businesses here have developed experience operating across multiple legal systems, languages, cultures, and regulatory environments. That adaptability has become an advantage in itself.

Rather than attempting to replicate Silicon Valley or compete directly with larger technology ecosystems, Southeast Asian companies have an opportunity to specialise in navigating complexity. The region’s strength lies in its ability to connect different markets, understand diverse customer needs, and operate within varying regulatory frameworks. Those capabilities are becoming increasingly valuable as the global economy becomes more fragmented.

The shift is equally visible in capital markets. Investors today appear less focused on how quickly a company can expand geographically and more interested in how resilient the business will remain if external conditions change. Questions about supply chain dependence, regulatory exposure, customer concentration, and operational flexibility are becoming more prominent during funding discussions. Growth remains important, but sustainable growth supported by resilient business models is attracting greater attention than expansion at any cost.

Also Read: The alliance economy: How founders and investors should position in a fragmented world

This has influenced how I think about building my own business. Rather than optimising solely for rapid market expansion, I am more interested in building a platform that continues creating value regardless of changing political or economic conditions. If the rules of international business continue evolving, then products capable of helping businesses interpret those changes become even more relevant. In many ways, geopolitical fragmentation has strengthened the long-term rationale behind what I am building.

As a result, my definition of “going global” has changed. It no longer means trying to establish a presence in as many countries as possible. Instead, it means building something that remains valuable across different markets despite their differences. It means understanding local regulations before entering a market, designing products that accommodate varying policy environments, and recognising that expansion strategies will increasingly differ from one jurisdiction to another.

If I were advising founders beginning their journey today, I would encourage them to study policy, industrial strategy, and capital flows with the same attention they devote to competitors and customer acquisition. Markets are no longer shaped purely by consumer demand or technological innovation. Government priorities, geopolitical relationships, and regulatory developments are becoming equally influential in determining where opportunities emerge and where risks accumulate.

Globalisation has not disappeared, nor do I believe it will. What has changed is the assumption that one strategy fits every market. The future belongs to founders who understand that technology may be global, but regulation, capital, and strategic priorities are becoming increasingly local. Success will belong not only to those who build the best products, but also to those who best understand the environment in which those products operate.

In a fragmented world, information remains valuable, but context has become indispensable. That realisation has shaped the most important business decision I have made over the past year, and I believe it will continue shaping how many founders approach growth in the years ahead.

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post Building through borders: A founder’s perspective on a fragmented world appeared first on e27.

The Generative AI revolution is here, but for many enterprises in Singapore and Southeast Asia, adoption has hit a hard wall. The barrier isn’t a lack of use cases; it is Data Security and Hallucination Control.

When dealing with highly sensitive domains (like Healthcare, Insurance, or Financial Data), passing raw payloads to external Large Language Models (LLMs) without strict guardrails is a compliance nightmare. Over the last few weeks, I set out to solve this exact problem by architecting a secure, multi-agent conversational platform on Google Cloud.

Today, I’m open-sourcing the reference architecture and the codebase.

The architecture: Defence-in-depth for LLMs

To build an enterprise-grade agent, you cannot simply connect a frontend directly to an LLM. You need a multi-layered security protocol. I leveraged GCP Sensitive Data Protection (SDP), Model Armour, and Vertex AI LLMOps Guardrails to create an impenetrable filtration layer before the data ever touches the Gemini 2.5 Flash agent.

• The PII redaction layer (SDP)

When a user inputs a prompt or uploads a medical document, the payload is intercepted by the GCP SDP engine. This engine uses custom inspection templates to hunt for specific Southeast Asian PII patterns (such as Singapore NRICs/FINs, local phone numbers, and names).

Before the LLM even sees the prompt, it is tokenised. For example, “Hi, my name is Syam, and my NRIC is S1234567A” becomes “Hi, my name is [PERSON_NAME], and my NRIC is [SINGAPORE_NRIC_FIN]”.

• The content filtration layer (Model Armour)

Even with PII redacted, the prompt must be scanned for malicious intent. I integrated GCP Model Armour as a firewall to detect prompt injections, jailbreaks, and toxicity. It scans both the inbound prompt and the outbound agent response to ensure the system cannot be manipulated into leaking internal system instructions.

• LLMOps guardrails: Fact-based constraints

In highly regulated sectors like insurance and healthcare, an AI system cannot give medical or financial advice. It is strictly restricted to providing fact-based details and retrieving policy information.

To enforce this, I implemented strict System Instructions and Guardrails managed through the Vertex AI LLMOps process. This acts as the final perimeter. If a user asks the agent for a medical diagnosis, the guardrails force the agent to politely decline and redirect the user to a human specialist. This virtually eliminates dangerous hallucinations.

• Agentic routing and Apigee integration

Once the sanitised prompt clears these three security layers, it hits the Agent Router. Using Google’s Agent SDK, the router dictates which specialised agent (e.g., Clinical Inquiry vs. Customer Support) should handle the request. These agents are wrapped in an Apigee API Gateway, allowing them to securely pull real-time enterprise data from internal databases.

Also Read: Singapore firms embrace agentic AI, but audit trails remain thin

Validating the architecture: GPT-as-a-judge

Building an agent is one thing; validating it at scale is another.

To prove this architecture works across the diverse linguistic landscape of Southeast Asia, I built a Batch Evaluation Platform. We ran simulated prompts through the pipeline in English, Cantonese, Malay, and Bahasa. Instead of manual review, I engineered an automated LLM-as-a-judge pipeline to score the responses based on relevance, harmfulness, and regional localisation accuracy.

Open source and next steps

As Singapore continues to push its Smart Nation agenda, securing AI workflows will be the defining challenge for our tech ecosystem. We cannot sacrifice privacy for innovation.

I have open-sourced the Terraform infrastructure and the backend codebase for this architecture on my GitHub. I encourage local developers and enterprise architects to fork it and adapt it for their own secure AI deployments.

• View the Infrastructure Repo here.
• View the Agent Security Framework here.

This article was originally published here. 

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post Securing Agentic AI for Singapore enterprises: A reference architecture appeared first on e27.

The Philippines is embracing generative AI at remarkable speed. But unless we learn to question it, govern it, and retain ownership of our decisions, the technology meant to empower us may quietly weaken our ability to think and act for ourselves.

EVER asked ChatGPT to write an email, settle an argument, interpret a medical symptom, choose a career path, or tell you whether you were right?

For many Filipinos, artificial intelligence is no longer merely a search tool. It is becoming an adviser, tutor, editor, programmer, confidant, and sometimes even an authority figure. We turn to it because it is fast, articulate, endlessly patient, and available at any hour.

The benefits are real. Generative AI can help workers draft reports, analyse documents, generate code, summarise meetings, translate ideas, and overcome the anxiety of a blank page. It can widen access to information and improve productivity, especially for people who lack immediate access to specialists.

But there is a question we are not asking often enough: What happens when AI does not simply assist our thinking, but begins to replace it?

This is the emerging risk of situational disempowerment—the gradual erosion of a person’s ability to form accurate beliefs, make authentic value judgments, and act according to independently considered reasons.

The danger is not that AI suddenly takes control. It is subtler. It happens when a confident answer feels more credible than it deserves, when an agreeable chatbot validates a mistaken belief, or when a user follows a recommendation without understanding how it was reached.

Consider the workplace. A junior employee uses AI to finish a task in half the time. The output appears professional, and the deadline is met. Yet when a client raises a difficult question, the employee cannot explain, defend, or revise the work. Productivity improved on the surface, but capability weakened underneath.

This creates a dangerous productivity inversion: AI accelerates routine tasks while potentially making people less prepared for complex, ambiguous, and high-value work.

The same risk extends beyond the office.

A person may ask AI whether a partner is manipulative, whether a family member should be cut off, or whether a major financial decision is sensible. The model may provide a coherent answer based only on one side of the story. Because the response sounds calm and authoritative, the user may treat it as objective judgment rather than probabilistic text generated from incomplete information.

Also Read: The localisation gap: Why multilingual AI isn’t enough for APAC markets

There are three ways this can distort human autonomy.

• Reality distortion occurs when AI reinforces false or unsupported beliefs. Large language models can fabricate facts, express false confidence, or agree with a user simply because agreement produces a satisfying conversation.
• Value-judgment distortion happens when users allow AI-generated moral positions to override their own principles. AI can help clarify choices, but it should not silently become the source of our values.
• Action distortion occurs when people execute AI recommendations with little independent review—sending a message, filing a document, making an investment, ending a relationship, or taking some other consequential step they later regret.

At the centre of these risks is what may be called belief offloading: allowing AI not only to organise information, but to form and maintain beliefs on our behalf.

Humans have always delegated cognitive work. We use calculators, maps, search engines, advisers, and institutions. Delegation is not inherently harmful. The problem begins when we stop evaluating evidence because the system appears more informed, more articulate, or more certain than we feel.

This matters greatly in the Philippines.

Filipinos are among the world’s most enthusiastic users of generative AI. At the same time, the country continues to struggle with functional literacy, misinformation, digital scams, and unequal access to high-quality education. In such an environment, fluency of language can easily be mistaken for accuracy, while confidence can be mistaken for competence.

Cultural tendencies may also intensify the issue. Filipinos often place substantial weight on expertise, hierarchy, social harmony, and trusted relationships. When an AI system is framed as an expert, mentor, or companion, challenging its output may feel less natural—even when challenge is precisely what responsible use requires.

The answer is not to reject AI.

Also Read: What AI safety researchers actually worry about

Businesses, schools, and government agencies should instead develop a more disciplined model of human–AI collaboration.

Users must know what to delegate and what to retain under human judgment. They must learn to give clear instructions, examine sources, identify unsupported claims, disclose AI use when appropriate, and remain accountable for final decisions.

Organisations also need acceptable-use policies that identify approved tools, restricted information, required review procedures, and decisions that must never be fully delegated to an AI system. Without such rules, employees will continue using multiple platforms invisibly, creating not only cognitive risks but also privacy, security, and governance exposure.

Most importantly, AI fluency must not be reduced to prompt engineering. Knowing how to obtain a polished answer is not the same as knowing whether that answer is correct, ethical, appropriate, or aligned with one’s objectives.

The decisive skill of the AI era will not be generating more output. It will be preserving human discernment amid an abundance of plausible output.

AI should expand our capacity to think, not become an excuse to stop thinking. It should sharpen judgment, not substitute for it. And it should remain a tool whose recommendations we examine—not an authority whose conclusions we simply obey.

The future of AI in the Philippines will not be determined only by the sophistication of the models we use. It will depend on whether Filipinos remain active authors of their beliefs, values, and actions.

Let our AI assist us. But let us never surrender the responsibility to decide.

This editorial is adapted from the dissertation proposal “Situational Disempowerment in the Age of Generative AI: Examining Reality Distortion, Value Judgment Distortion, and Action Distortion Among Filipino Large Language Model Users.”

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post When AI starts thinking for us appeared first on e27.

The post Bitcoin holds US$64,341 while miners bleed US$1.26B: What is really happening? appeared first on e27.