Acrab Inc., a Singapore-headquartered frontier tech company, has announced it has secured more than US$350 million in cumulative financing since its founding in 2024, as it looks to establish itself as a foundational infrastructure provider for the next generation of agentic AI.

The company, which had previously operated in stealth, said the financing was raised across multiple rounds and includes participation from global venture capital firms and strategic industry investors. Vertex, the global venture platform backed by Singapore state investor Temasek, was among the earliest backers through its Vertex Ventures SEA & India and Vertex Growth funds, and has continued to increase its commitment as Acrab reached key technological milestones.

The announcement also marks the commercial debut of GΞLIX, Acrab’s first-generation compute platform designed to support local LLM inference and agentic AI workloads at the edge. The company said GΞLIX has been validated in demanding real-world deployment environments and is progressing toward first industry adoption and mass production.

Acrab’s full-stack architecture spans AI silicon, local LLM inference, operating systems, multi-modal human-machine interface (HMI), and agent orchestration technologies. The company positions its approach around heterogeneous computing — specifically, the coordinated interaction between central processing units (CPUs) and neural processing units (NPUs) — which it describes as central to the performance demands of agentic AI systems.

Also Read: Top 4 Best ERP for Large Enterprise in Malaysia

“CPUs are becoming increasingly important as AI systems evolve into heterogeneous computing environments, where execution depends not only on NPU performance, but on the seamless coordination between CPUs and NPUs,” said Dr. Ken Phua, CEO of Acrab. “Delivering this new generation of agentic experiences calls for a fundamentally new compute foundation.”

Phua brings extensive experience from the semiconductor sector, including a career at Arm UK, where he led Asia Applications Engineering and co-led global IP strategies, before serving as co-CEO of Arm China.

Vertex’s Kee Lock Chua, CEO of Vertex Holdings, said the firm’s conviction in Acrab rests on a belief that the next wave of AI will run at the edge rather than in the cloud. “Our confidence has only deepened as the team has translated that thesis into a validated platform, and we’ve increased our commitment at every step,” he said.

Acrab’s target markets span personal AI PCs, home hub devices, in-vehicle intelligence, industrial operations, and general robotics — environments the company believes will increasingly require private-by-design, context-aware computing that can carry out tasks in real time without relying on cloud infrastructure.

Also Read: Vietnam isn’t just inviting private capital in. It is structurally dependent on it

The broader thesis underpinning Acrab’s strategy is a distinction between generative AI, which responds to prompts, and agentic AI, which the company characterises as systems capable of inferring intent, co-ordinating tools, and executing tasks on behalf of users — a shift that demands a new class of compute infrastructure.

Acrab said it will use the new capital to accelerate platform development, expand partnerships with global technology firms, and strengthen its presence in key international markets.

The post Acrab raises US$350M to advance Agentic AI compute infrastructure appeared first on e27.

Carro, the Singapore-headquartered online automotive marketplace that has built a significant presence across Southeast Asia, has made its first major move into Australia with the acquisition of CarPlace, a used-car platform operated by Australian automotive group Autoleague.

The deal gives Carro a physical presence in three of Australia’s four largest markets in one go. It signals a strategic effort to export its technology-rich approach beyond Southeast Asia. For an industry accustomed to fragmented dealer networks and opaque pricing, Carro is pitching transparency, end-to-end workflows and a stronger wholesale pipeline, including vehicles from Japan, as its differentiators.

Also Read: Profitable, AI-driven, and IPO-ready: Inside Carro’s next chapter

A foothold Down Under

CarPlace, positioned among Australia’s more prominent used-car platforms, will remain linked to Autoleague, which becomes a strategic shareholder in the combined business. Autoleague’s continued involvement provides Carro with an immediate network of dealerships and operational scale that would take years to build organically.

Aaron Tan, Carro’s co-founder and chief executive, framed the move in blunt commercial terms: Australia, he said, is “one of the largest used car markets in Asia Pacific, with a consistent annual sales volume of 2.3 million used cars and fast-growing EV penetration.” He argued the market is “ripe for a platform like Carro to transform the used car landscape and deliver a better, convenient, more transparent customer experience that is powered by technology and AI.”

Tan’s pitch rests on a two-pronged strategy:

1. Strengthen retail pre-owned operations using Carro’s proprietary tools, such as vehicle inspections, inventory tracking, asset and lead management.
2. Grow wholesale activity, leveraging Carro’s Japan presence, to serve the Japan–Australia import corridor.

The latter is notable because Japanese models remain popular in Australia, and a reliable cross-border wholesale channel could be a lucrative niche.

What Carro brings: technology and dealer relationships

Carro’s operating model in Southeast Asia blends marketplace listings with back‑end services for dealers, combined with data and AI to reduce information asymmetry. In its markets, Carro has emphasised standardised inspections and clearer pricing, both of which can shorten transaction times and reduce post-sale disputes.

Also Read: Carro acquires Beyond Cars, bets big on Hong Kong’s strong EV growth

“Building and maintaining a strong dealer network has always been key for all the markets we’re in,” Tan said. “Carro has a track record of partnering with local dealers to support their growth, and we welcome ‘win-win’ partnerships in Australia. We’re confident in our strong Wholesales capabilities, thanks to our presence in Japan.”

For Australian dealers, access to Carro’s inspection protocols, lifecycle management tools, and potential new supply from Japan may help both sourcing and remarketing. For consumers, the promise is greater transparency on vehicle condition and pricing, a recurring sore point in used-car markets globally.

Autoleague’s backing reduces execution risk. Dan Kawai, Autoleague’s managing director and CEO, welcomed the partnership: “We’ve seen Carro’s technology infrastructure, streamlined operations, and unwavering commitment to transparency within the industry, and we’re confident in their goal to become a leading player in Australia.”

Southeast Asian angle and regional implications

Carro’s expansion is relevant to Southeast Asia’s increasingly competitive used-vehicle ecosystem. Over the last decade, regional players have experimented with a range of models: pure marketplaces, trade-only exchanges, and vertically integrated services that include financing, repairs and logistics. Carro’s Australia move demonstrates a reverse flow of scale: a Southeast Asian-grown company exporting its playbook to a developed market.

That matters for several reasons. First, it underscores the maturing capabilities of Southeast Asian tech startups, many of which have moved beyond consumer-facing apps into complex logistics and asset-heavy categories. Second, if Carro successfully integrates Australian dealers and the Japanese supply chain, it could create a blueprint for other regional expansions, including more formalised import/export routes between Japan, Southeast Asia, and Australia.

Finally, the move could stoke competition in Australia from both local incumbents and other regional challengers. Sellers and dealers in Southeast Asia may face new competition for Japanese used-vehicle stock if Carro scales its wholesale operations across borders.

Challenges ahead

Carro’s promise of technology-driven clarity is persuasive on paper, but execution risks are real. Australia’s market is dispersed and regulated at the state level for vehicle inspections, registrations, and consumer protections. Winning consumer trust will require consistent inspection standards and reliable post-sale support, areas in which established local players have credibility.

Moreover, scaling a Japan-Australia import corridor means handling customs, compliance, logistics, and model homologation issues, tasks that can be capital-intensive and operationally complex. The economic case depends on margins after shipping and compliance costs, especially for lower-priced models.

Also Read: Carro invests in digital content, marketing services agency Driven Communications

There is also the question of brand recognition. While Carro is well known in Southeast Asia, Australian consumers and dealers may initially view the company as an outsider. Close cooperation with Autoleague and local dealers could mitigate that, but it will take time to convert partnerships into sustained market share.

What to watch next

In the near term, Carro will likely focus on plugging CarPlace’s dealer network into its tech stack, standardising inspections, and piloting Japan-sourced inventory flows. Observers should monitor three metrics: dealer uptake of Carro’s platform features, wholesale volumes moving from Japan to Australia, and post-purchase consumer satisfaction scores.

If Carro succeeds, the acquisition will mark a pivotal moment for Southeast Asian mobility tech firms: proof that their operational models can scale into developed markets and that regional cross-border automotive supply chains are commercially viable.

For Australia’s used-car market, the arrival of a technology-led, cross-border operator could accelerate transparency and digitisation to the benefit of buyers and dealers alike, but only if the promises match the execution. The partnership with Autoleague gives Carro a bridge; the rest will depend on how quickly it turns that bridge into reliable, everyday commerce.

The post Singapore’s Carro expands regional footprint by acquiring CarPlace: what it means for SEA, Australia markets appeared first on e27.

A Jakarta courtroom delivered a verdict on Thursday that is already reverberating far beyond Indonesia’s borders.

Four senior executives from two of the country’s most prominent state-backed venture capital firms, MDI Ventures and BRI Ventures, have been sentenced to prison over investments made in the now-defunct agritech startup TaniHub.

Also Read: Indonesia detains 3 more suspects in TaniHub investment fraud case

The ruling has sent a chill through the regional startup ecosystem, prompting an immediate and forceful pushback from the defence and igniting a broader debate about the criminalisation of business failure.

The sentences

The Jakarta Corruption Court sentenced former MDI Ventures CEO Donald Wihardja to five years in prison and fined him US$28,000, with a substitute sentence of six months in detention should the fine go unpaid. His former vice president of investment, Aldi Adrian Hartanto, received a two-year prison term and a US$14,000 fine.

On the BRI Ventures side, former CEO Nicko Widjaja was handed a three-year sentence and fined US$19,600, with 110 days of additional imprisonment as a final substitute if asset seizure proves insufficient. Former vice president of investment William Gozali received two years and a US$14,000 fine.

Prosecutors alleged that all four approved investments in TaniHub without performing adequate due diligence, leaning too heavily on data supplied by the startup itself. Authorities put MDI Ventures’s losses to the state at approximately US$20 million and BRI’s at around US$5 million.

How a farming unicorn hopeful unravelled

Founded in 2016, TaniHub was once among Indonesia’s most celebrated agritech bets. The company raised US$92.5 million in total disclosed funding over its lifetime, including a high-profile US$65.5 million round in 2021 in which both MDI Ventures and BRI Ventures participated alongside other investors. It positioned itself as a transformative platform connecting smallholder farmers to buyers and, eventually, consumers, scritping compelling story in a country where agriculture employs nearly a third of the workforce.

The story, however, did not end well. TaniHub ran into severe financial difficulties, carried out sweeping layoffs, and eventually wound down much of its operations. By September 2023, the value of TaniHub shares held by BRI Ventures had fallen to approximately US$380, a near-total write-off on what had been a US$5 million position.

Also Read: “Special Projects” and shady metrics: TaniHub whistleblower speaks as top execs detained

It was that collapse in value that prosecutors and, ultimately, the court treated as evidence of state financial loss.

What the court found

In the case against Widjaja, the Panel of Judges found that BRI Ventures’s investment process, split between a US$2 million Series A+ round and a US$3 million convertible note round, violated the prudential principle on multiple counts.

The court found that the deep feasibility study relied too heavily on TaniHub-supplied data, without sufficient independent verification, and that the analysis was based in part on unaudited financial statements. The outstanding receivable concerns were not adequately interrogated.

The panel also noted that the investment committee at the time consisted only of the president director, and that the oversight function of the board of commissioners had not operated optimally.

On the question of personal enrichment, a standard element in Indonesian corruption cases, the court acknowledged that Widjaja received no personal benefit whatsoever. Nevertheless, it held the element fulfilled because the US$5 million flowed to TaniHub as a third-party corporation. The court further held that cooperation in the investment decision-making process was sufficient to establish joint participation without requiring proof of an explicit agreement.

The defence fires back

Widjaja’s legal team, led by Ditho Sitompoel, did not mince words in their response.

“The line between a failed investment decision and a criminal act must be carefully preserved, so that criminal law is not used to judge a business decision based solely on its outcome — that is, hindsight bias,” Sitompoel said in a statement.

The defence raised six pointed objections to the court’s reasoning. Chief among them was the argument that the panel’s prudential standard was simply wrong for the asset class. Venture capital, by its very nature and by the explicit mandate of POJK 35/2015 (Indonesia’s regulatory framework for the industry), requires firms to take measured risk on high-growth companies that are often pre-revenue, unaudited, and operating at a loss. Holding a VC firm to the verification standards of a bank extending credit to a mature business, the defence argued, would render nearly every venture investment in Indonesia legally suspect.

The team also highlighted that BRI Ventures had, upon identifying deteriorating conditions at TaniHub, halted Series B funding and pursued divestment, hardly the conduct of a reckless actor. The panel acknowledged these steps but held that they did not negate the original unlawful act.

Perhaps most significantly, Sitompoel pointed to the systemic implications. “If investment failures that went through proper approval, review, and governance can still be criminalised, this creates legal uncertainty with the potential to produce a chilling effect on investors, directors, commissioners, and professionals who are required every day to take the legitimate business risks needed to drive economic growth.”

A verdict with consequences beyond one courtroom

The concern is not abstract. Indonesia has spent years trying to build itself into a credible destination for venture capital, and state-backed funds like MDI Ventures and BRI Ventures have been central to that ambition, deploying government-linked capital into the startup ecosystem at a time when private capital was still finding its footing.

Also Read: Nicko Widjaja’s legal defence team on the prospect of winning: “We are confident enough”

The TaniHub verdict raises an uncomfortable question: if a VC fund manager at a state-linked institution can be imprisoned when a portfolio company fails, even absent any personal gain or proven misconduct, who in their right mind will take the job?

The four convicted executives are widely expected to appeal. How Indonesia’s higher courts handle those appeals may well determine the country’s investment climate for years to come.

The post Four VC executives. Zero personal gain. Three years in prison appeared first on e27.

Banks are still talking about AI governance as though it belongs mainly to policy teams, future regulation, and committee oversight. That view is already out of date. AI governance is now an operating model issue.

The real danger is not only that a model makes a poor recommendation or that an employee shares sensitive information with the wrong tool. Those risks matter, but they are also the easiest to recognise. The deeper problem is that AI adoption is spreading faster than many banks can govern judgment, accountability, customer treatment, and operational response.

That matters because banking is not judged only by the quality of its policies. It is judged by whether decisions can be explained, defended, and corrected when customers are affected. If AI use spreads faster than the control model around it, the consequences will not stay inside technology teams. They will surface in complaint handling, operations, fraud casework, collections, service recovery, and exceptions, which are all areas where judgment carries regulatory and reputational weight.

The wrong starting point

A great deal of discussion still begins with model risk, privacy, or compliance. These issues matter, but they are not the most useful place to begin. A better question is this: what happens to banking operations when employee judgement is partly assisted by systems that are fast, persuasive, difficult to challenge, and unevenly understood across the organisation?

That question changes the conversation. It moves the bank away from abstract talk about AI innovation and towards the daily reality of operational decisioning. It forces leaders to look at how a complaints handler uses AI to summarise a case, how an operations analyst drafts an exception rationale, how a collections agent shapes a customer conversation, or how a fraud investigator relies on AI driven prioritisation.

In each case, the issue is not simply whether the tool exists. The issue is whether the bank can still stand behind the quality, fairness, and traceability of the resulting action. That is the point where AI governance stops being a technology topic and becomes a banking discipline.

Banks are governing two different problems

One reason many governance frameworks remain vague is that banks often blend together two different forms of AI use.

The first is employee AI use. This includes drafting, summarising, searching, translating, analysing, and preparing work. On the surface, that can look like a productivity issue. In practice, it is also a judgment and control issue. Once employees rely on AI to shape internal outputs, customer responses, escalation summaries, or operational recommendations, the work is no longer purely human in the old sense. It becomes a blended output. That means the bank needs clear rules on what can be delegated, what must be checked, and what remains fully owned by the employee.

Also Read: The shadow ledger: Why AI governance is the new architecture of brand trust and enterprise revenue

The second is AI-assisted business decisioning. This sits much closer to customer outcomes, prioritisation, approvals, exceptions, operational triage, and vulnerability handling. Here, the stakes are higher because AI is not only improving speed or language. It is influencing actions that can shape fairness, access, treatment, and trust.

These two categories require different controls. Too many governance models treat them as one broad AI issue and end up being both too loose and too blunt.

The first failures will be quiet

Banks often expect AI risk to arrive as a major breach, a headline incident, or an obvious model failure. That expectation is misleading. The first operational failures are more likely to be quiet and cumulative.

Sensitive data starts appearing in prompts because employees are under pressure, and approved workflows are slower. AI-generated summaries begin to flatten nuance in complaint files, which weakens later review. Decision support outputs become trusted because they sound coherent, even when the reasoning beneath them is thin. Frontline teams rely on wording that is technically acceptable but badly judged for vulnerable customer situations. Managers approve outputs without really knowing how much human challenge has been applied. Control teams discover that the audit trail feels sufficient for internal reassurance but is not strong enough for real scrutiny.

None of these issues looks catastrophic in isolation. Together, they create a more serious pattern. The bank begins to lose clarity about where human judgment ends, and AI-shaped judgment begins.

Why traditional model governance is not enough

Banks already know how to govern models in a traditional sense. They have approval forums, validation standards, documentation requirements, monitoring routines, and control owners. Those disciplines still matter, but they are not sufficient for the current AI wave.

What is needed now is governance of assisted judgment.

That means understanding not only whether a model is technically sound, but also how AI-shaped outputs enter workflows, how much challenge is realistically applied by employees, where reliance becomes routine, how explanations are captured, how errors are detected, and how customer harm is identified when the decision was not fully automated yet was no longer fully human either.

Also Read: Why emerging markets need AI governance infrastructure before AI scale

This is where current governance language can become misleading. Many organisations take comfort from calling a tool assistive rather than determinative. In practice, assistive systems can heavily shape outcomes because they influence what employees see first, how they frame the case, what options appear reasonable, and how quickly they move.

Complaint handling is the real test

Many banks will discover the quality of their AI governance not through a model review, but through complaints.

Complaint handling is where customer harm, operational judgement, fairness, explanation, and record quality all come together. If AI is being used poorly elsewhere in the operation, complaint handling is often where the weaknesses become visible. Cases become harder to assess. Summaries lose context. Root cause becomes harder to establish. Responses sound polished but unconvincing. Reviewers struggle to tell whether the original decision reflected genuine consideration of the customer’s circumstances or simply repeated an AI-shaped conclusion.

This is why complaints should be treated as a primary governance lens. If a bank cannot defend AI-influenced operational decisions inside its complaints process, it is not governing them properly. If complaint handlers do not know where AI was used earlier in the customer journey, they cannot properly assess fairness. If complaint responses themselves are AI-assisted, the bank needs very clear boundaries on what can and cannot be delegated.

A better model

A stronger governance model starts with use case classification. Banks need to separate low-consequence productivity support from high-consequence decision support, while also recognising the large middle ground where AI is not making the final decision but is materially shaping human judgment.

From there, governance needs to move into workflow design. Where is AI allowed? What data can be used? What outputs are acceptable? What level of verification is required? What must be recorded? What escalation is needed if the AI output conflicts with case evidence or customer context?

After that comes monitoring. Not only whether a tool is secure and available, but how it is actually being used, where teams are becoming over-reliant, where outcomes are drifting, and which workflows are creating defensibility problems.

Then comes response capability. If something goes wrong, the bank needs to know how to pause usage, investigate the impact, identify affected customers, explain what happened, and remediate quickly. Governance without response capability is only paperwork.

—

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 AI governance in banking operations and decisioning appeared first on e27.

Most voice AI demos are built for perfect conditions.

The room is quiet. The Wi-Fi is stable. The user speaks clearly, waits for the response, and follows the script. In that setting, many modern voice agents can sound impressive.

The real test begins outside the demo room.

A user in Jakarta speaks over mobile data during a commute. A child in Manila interrupts an AI tutor halfway through an explanation. A customer in Ho Chi Minh City switches between English and Vietnamese. A smart device in Thailand moves from stable Wi-Fi to weak 4G.

Suddenly, the AI no longer feels intelligent. It feels delayed, stiff, and difficult to trust.

For teams building voice AI in Southeast Asia, this is the uncomfortable lesson: the issue is not always intelligence. It is whether the full voice experience can survive real-world conditions.

Physical AI needs more than a voice

Physical AI is not just AI inside a device.

It is AI that lives in the user’s environment and becomes part of daily life. It may appear as an educational toy, a wearable translator, a companion robot, an in-car assistant, a smart camera, or a home device that can talk, listen, and respond.

That changes the standard for user experience.

When people type into a chatbot, waiting feels normal. The interaction is already asynchronous. A short delay may be acceptable because the user is looking at a screen and expecting software to process a request.

When people speak to a device in the room, the expectation is different. They expect rhythm. They expect responsiveness. They expect the device to understand when a conversation has started, shifted, or ended.

This is why physical AI has to be evaluated differently from text-based AI. It is not enough for the answer to be correct. The interaction has to feel natural in the moment.

The bot feeling usually starts with latency

When users say a voice AI “feels like a bot,” they are often reacting to timing before they are reacting to content.

A pause after every sentence makes the experience feel mechanical. A delayed answer makes the user wonder whether the device heard them. A voice agent that continues speaking after the user has moved on feels disconnected from the conversation.

This sensitivity is not new. In real-time voice traffic, Cisco cites the ITU G.114 recommendation of less than 150 milliseconds of one-way end-to-end delay for high-quality voice. Voice AI adds more layers on top of that, including speech recognition, model response, speech output, and routing between services.

Also Read: To Voice AI or not – The changing face of customer experience

Human conversation depends on small timing cues. We pause, overlap, correct ourselves, interrupt politely or impatiently, and change direction mid-thought. These are not edge cases. They are the normal shape of how people speak.

Voice AI breaks when it treats conversation as a clean sequence: user speaks, machine processes, machine replies.

A more natural system needs a fluid loop. It has to listen, process, generate, and speak with minimal delay. It also has to adapt when the user changes direction. That requires real-time audio transport, streaming speech recognition, response generation, speech output, and interruption handling to work together.

For builders, latency is not just an engineering metric. It is part of the product’s personality.

Southeast Asia turns weakness into a market problem

Southeast Asia is an important region for voice-first AI because the use cases are practical.

The region’s digital economy is already large enough for these experiences to matter at scale. Google, Temasek, and Bain estimate that Southeast Asia’s digital economy is set to surpass US$300 billion in gross merchandise value by 2025. The same report frames the region’s next phase around AI adoption, after years of growth in digital services.

The opportunity is clear: mobile-first users, multilingual households, growing demand for education technology, rising adoption of connected devices, and many situations where voice can make technology easier to use. A screenless device, a voice tutor, or a smart assistant can be valuable when typing is inconvenient, when users move between languages, or when a product is used by children, older adults, or workers on the move.

But the same conditions also make the region difficult.

Indonesia shows the scale and complexity. DataReportal counted 212 million internet users in Indonesia at the start of 2025, along with 356 million cellular mobile connections, equivalent to 125 per cent of the population. Yet Ookla data cited by DataReportal showed a median mobile internet download speed of 29.06 Mbps. For voice AI, that gap between massive connectivity and uneven consistency is where user experience problems appear.

The fragmentation is regional, not just local. Data found that mobile-only smartphone users make up less than 10 per cent of users in markets such as Vietnam, Brunei, and the Philippines, but more than a quarter in East Timor, Laos, Thailand, and Malaysia. In Laos, Thailand, Malaysia, Cambodia, and Indonesia, more than 40 per cent of smartphone users have no or very limited Wi-Fi use.

This turns voice quality from a technical detail into a market expansion issue.

If a product only works in controlled conditions, it cannot scale confidently across the region. If it struggles with accents, unstable networks, or mixed-language behaviour, users will not wait for it to improve. They will stop using it.

Also Read: How voice AI is revolutionising the fintech scene

The stack needs to be built for interruption

A strong physical AI product needs more than a model and a synthetic voice.

• The device needs reliable audio capture: If the microphone hears too much background noise or misses the wake word, the experience fails before the model is involved.

• The voice pipeline needs low-latency transport: Audio has to move quickly between the device, the cloud, and the AI services without adding noticeable delay.

• The system needs interruption handling: Humans do not wait politely for a machine to finish talking. They correct it, interrupt it, and change direction. A natural voice agent must be able to stop, listen, and respond without making the user repeat everything.

• The AI needs memory and context: This is where physical AI starts to feel different from a basic voice bot. A companion device that remembers preferences, routines, or past interactions can create a sense of continuity.

• The product needs a persona: Not every device should sound friendly. Some should sound calm, professional, playful, or neutral. A toy, a healthcare assistant, and an in-car agent should not share the same personality.

The “soul” of physical AI comes from this full stack. The model matters, but it is only one part of the experience.

Builders should measure the conversation, not just the model

Many teams still evaluate voice AI by asking which model is the smartest.

That is too narrow.

A better question is: what does the full conversation feel like in the user’s actual environment?

For teams building in Southeast Asia, that means testing on mobile data, not just office Wi-Fi. It means testing noisy rooms, not just quiet meeting spaces. It means testing repeated use, mixed-language behaviour, unstable networks, and users who do not follow a script.

Product and procurement teams should ask practical questions before committing to a voice AI stack:

• What does the experience feel like on a real 4G connection?
• How quickly can the agent respond during natural turn-taking?
• Can it handle a user changing direction mid-conversation?
• What happens when the network becomes unstable?
• Can the stack support different models, speech providers, and deployment needs?
• Can the product preserve useful context across sessions?
• Can the voice persona be adapted for different markets and product categories?

Also Read: Never fear, AI is here: Helping midlife artists build their social media voice

The next AI device will be judged by how it feels

The next wave of AI will not stay inside chat windows.

It will live in toys, robots, wearables, cars, cameras, appliances, and industrial devices. In those environments, users will judge AI less like software and more like something sharing their space.

They will notice whether it listens at the right moment. They will notice whether it talks too much. They will notice whether it remembers. They will notice whether it is helpful or frustrating when the environment becomes noisy and unpredictable.

For Southeast Asia, this is both a challenge and an advantage. The region is difficult to build for because of its network complexity, language diversity, and mobile-first behaviour. But any physical AI product that performs well here will be stronger in many other markets.

The question for builders is no longer whether voice AI can speak. The question is whether it can stay useful when the real world gets messy.

If your physical AI device had to hold a natural conversation today on a crowded Southeast Asian mobile network, would it still feel alive?

—

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 Give physical AI a soul: Why your voice AI still feels like a bot appeared first on e27.