A passport is not a clever piece of paper. Anyone with a good printer and bad intentions can fake one. What makes it work is the government standing behind it, an authority that vouched for the holder and can withdraw that vouching the moment something goes wrong.

Evin McMullen has spent a decade thinking about that distinction. “Humans rarely trust each other directly; we trust the institutions that will hold someone to account if they lie,” says the co-founder and CEO of AmanTrue, who studied the anthropology of information systems at Yale before building digital identity products at Disco and Billions. AI agents, she argues, are being asked to earn trust the hard way, on their own merit. “That is like building in the void.”

Also Read: Southeast Asian tech leaders are learning to trust AI agents, but not with production

The timing of that argument is no accident. Singapore’s Infocomm Media Development Authority (IMDA) has rolled out a starter kit for testing AI agents, an accreditation scheme with five approved third-party testers and a frontier AI security taskforce. Its guidance examines what an agent does, which tools it touches and what sensitive data it can reach, not merely what it outputs.

Yet, as e27 has reported, nothing yet compels anyone to use accredited testing. And companies are already stuck. An Adobe survey of 461 Southeast Asian executives found that 18 per cent of AI proposals die at security review, while only 12 per cent of firms have an AI owner with budget authority.

AmanTrue, which launched a couple of weeks ago, is betting that the missing piece is not a better exam but a better passport.

Trust needs a chain of accountability

McMullen’s starting point is that agents are stochastic. “An agent cannot reliably earn trust behaving well in a demo,” she says. “The past is not necessarily prologue.” What agents need instead is what people already have: “someone accountable standing behind it, a clearly defined mandate, and a record of data and action that can be checked.”

That is the logic behind AmanTrue’s Agent Passport, a credential derived from the verifiable identity of whoever deploys the agent, and Border Control, a gateway that sits on the receiving system’s side and decides what an agent may ask and do.

An identity vendor arguing that testing is not enough is, of course, exactly what you would expect. McMullen concedes the point more readily than most. Pre-deployment testing, which she says Singapore does “unusually well”, catches whether a model reasons badly, misuses a tool or falls for a prompt injection. “A passport does none of that. It will happily identify a badly behaved agent.”

What identity covers, she argues, is the stretch after launch: who is accountable for a live agent, what it was authorised to do and proof of what it did, across systems the tester never saw. “Testing is about the agent’s competence before launch. Identity is about its authority and accountability after launch. You need both, and anyone who tells you one replaces the other is selling you half a solution.”

Whose border, whose rules

The harder test is cross-border, which in Southeast Asia is the default rather than the edge case. Picture an agent from a Singapore bank acting on a Malaysian insurer’s system, a scenario the region’s insurers are only beginning to confront.

In McMullen’s model, the receiving side is sovereign. “Border Control sits on the insurer’s side and enforces the insurer’s rules, not the bank’s.” The insurer chooses which issuers it trusts. When two registries disagree, she points to trust frameworks “built the way visa and correspondent-banking arrangements were built between countries over time”.

That analogy cuts both ways. Visa regimes and correspondent banking took decades of treaties, scandals and fines to mature. Agents are unlikely to wait that long.

Also Read: AI agents broke into 100 firms for pocket change. SEA’s online shops fit the target profile

On a subtler threat, an agent quietly firing hundreds of narrow yes-or-no questions to rebuild a record it was never meant to see, her answer is more concrete. Each field carries a rule, most probing questions are refused before they are answered, and every query is written to a ledger recording who asked, what was checked and which rule decided. “A flood of oddly specific questions from one agent is exactly the signal an audit or an anomaly check is there to catch.”

The answers that stayed vague

Not every question drew a straight reply. Asked why AmanTrue needed to be a company separate from Billions Network, McMullen offered a single line: “Different products, capabilities, users, and strategies.”

AmanTrue’s website names TikTok, Microsoft and HSBC as running technology the team built, while noting they are not AmanTrue customers. Asked how many paying institutional customers it has, and whether any are in Southeast Asia, the company said it “has not yet publicly disclosed our working relationships”.

On the Billions token, the answer was a firm no. AmanTrue does not depend on it and can run on a sovereign cloud, a national data centre or a client’s own servers. “We make independence a deployment fact, not a promise,” she said, adding that “any implication otherwise is a misunderstanding of our competence”. Banks and ministries will still ask, given the long-running debate over whether AI agents need crypto rails.

On the 2am scenario, a rogue agent and a revoked passport, no figure was offered. Revocation is enforced at the point of use, so the next check fails; how quickly depends on “the process of the issuer, the setup of the system, and the nature of the revocation protocol”. For a CISO who has read about the agent that lied, that latency is precisely the number they will want in writing.

AmanTrue also walks into a crowding field. Microsoft and Okta have extended their identity platforms to cover AI agents, card networks such as Visa and Mastercard are building protocols to authenticate agents that shop and pay, and “Know Your Agent” is fast becoming payments’ next compliance phrase. A newly launched specialist will have to show why a bank should not simply wait for its existing identity provider to add the feature.

Who signs the cheque

The Adobe numbers raise an awkward question for any vendor: if nobody owns the AI budget, who buys? McMullen calls that critique “half right”. Governance is a genuine bottleneck, she agrees, but the budget does have an owner: “whoever is accountable when an agent gets something wrong.” In a bank, that is the CISO and the chief risk or compliance officer; in government, the ministry that owns the service.

These are the people rejecting proposals at security review today. Her pitch is that provable actions give them a reason to say yes, and in doing so create the accountable owner the org chart lacks. “Rather than replace governance, the technology gives governance something tangible to enforce.” It echoes a wider regional shift, where AI governance is moving from promises to proof even as enterprises remain unsure who referees their agents.

Do not wait for the breach

Should Singapore make agent identity mandatory for regulated sectors? McMullen would not mandate a product, “mine included”. She would mandate an outcome: anyone deploying an autonomous agent against regulated systems must be able to prove which agent acted, on whose authority and what it did. The case for not waiting is blunt. Leaked personal or health data is something “you cannot un-leak”.

Also Read: Banks want AI agents to move money, as long as someone is accountable

She is candid about the conflict. “I have a commercial interest in that answer, so weigh it accordingly.”

That is fair advice. But her central observation holds regardless of who is selling the fix. Singapore has built a rigorous exam hall for AI agents. What it has not yet built is the border post, where every time an agent tries to act, someone checks whose name is on its passport.

The post AI agents need someone standing behind them, says AmanTrue CEO Evin McMullen appeared first on e27.

On paper, Malaysia’s trade figures read like a victory lap. Technology-related exports jumped about 55 per cent year-on-year in September, and the monthly trade surplus hit its second record high of MYR 34 billion (US$8.38 billion).

When looking closer, though, the boom appears to be largely borrowed.

In a new note, CGS International argued that the surge has been powered mainly by a global appetite for memory-related products, not by Malaysia winning a bigger share of the market. The gains have also yet to show up where they matter most to ordinary Malaysians: in wages, jobs and fresh investment.

Also Read: AI boom pushes chipmakers to expand across Asia Pacific, FedEx report finds

That distinction is not academic. A country that rides a demand wave gets carried back down when the wave breaks. A country that gains competitiveness keeps its position after the tide goes out.

A boom built on memory

The growth has been narrow. It is concentrated in electrical and electronics (E&E) products, particularly memory, solid-state devices and integrated circuit (IC) parts, all of which have benefited from demand tied to AI and the data centre build-out reshaping the region’s chip supply chain.

Electrical machinery accounted for roughly 40 to 50 per cent of Malaysia’s export growth so far this year. The US was the single largest contributor to the increase in shipments, adding about 10 to 15 per cent to export growth.

The trade map tells its own story. Malaysia’s widening surplus with the US is offset by deficits with China and Taiwan. In plain terms, components flow in from East Asia, are assembled, tested and packaged in Malaysia, and then shipped out to American customers. CGS International describes this as Malaysia acting as a conduit within the global technology supply chain. Conduits are useful. They are rarely where the money pools.

The scoreboard that hasn’t moved

To test whether Malaysia is genuinely outperforming, CGS International used Revealed Comparative Advantage (RCA), a measure that compares how much a product weighs in a country’s exports against its weight in global trade. If Malaysia were winning, its RCA index would climb.

It has not. Malaysia already holds a competitive position in IC parts and solid-state devices, but the latest data shows no significant improvement in its RCA score despite the sharp rise in exports. The research house’s reading: Malaysia is benefiting from a bigger pie, not grabbing a larger slice.

The memory numbers make the point neatly. Malaysia’s memory exports rose 243 per cent year-on-year to date. The World Semiconductor Trade Statistics (WSTS) forecasts global memory sales growth of 249 per cent. Malaysia, in other words, is swimming slightly slower than the current.

The gains have also come from volume rather than pricing power. Memory shipment volumes in mid-2026 reached about three times their 2023 levels. That matters because volume growth depends on factory capacity, while soaring memory prices tend to reward front-end producers, the companies that design chips and fabricate wafers. Malaysia’s strength lies in the back end: assembly, testing and packaging, where margins are thinner.

There is one encouraging signal. OECD Trade in Value Added data suggests Malaysia is capturing a larger share of domestic value added, meaning more of the economic value from each export is staying in the country.

Where are the pay cheques?

The broader economy has yet to feel the boom.

Employee compensation in the E&E sector has stayed broadly flat since 2022-2023, according to CGS International. Employment in the industry did not grow significantly in the first eight months of 2026.

Also Read: AI demand lifts Malaysia’s chip sector, but not every player wins

Investment is the starker gap. Quarterly approvals in the sector have hovered between RM5 billion and RM10 billion (roughly US$1.2 billion to US$2.5 billion), far below the RM40 billion to RM80 billion (about US$9.9 billion to US$19.7 billion) peaks of the 2021-2023 investment boom.

The research house offers a caveat: wages, hiring and investment typically lag exports, so the benefits may still arrive. There is some evidence for that. Manufacturing capacity utilisation rose to 87.7 per cent in June 2026, its highest since 2015. Factories running that hot usually have two choices: expand or turn away orders.

Still, “it’s coming” is an argument that needs the cycle to cooperate. And the cycle may not wait.

The cycle is already cooling

CGS International expects E&E export momentum to moderate. WSTS forecasts global semiconductor sales growth to slow from about 90 per cent in 2026 to 27 per cent in 2027.

Malaysia’s exports are likely to follow. Because it sits at the back end of the supply chain, it may also feel the slowdown later than others, a delayed hangover after a party it mostly catered.

Climbing the value chain, again

The policy challenge, the research house said, is to sustain growth beyond this demand cycle by moving up the value chain, particularly into advanced packaging and semiconductor services. It views the National Semiconductor Strategy as a more pragmatic plan, with realistic targets for developing the industry and capturing more value.

Malaysia is not alone in this ambition. Thailand is targeting a US$80 billion semiconductor push to move beyond assembly, and neighbours from Singapore to the Philippines are chasing the same higher-value work. At home, startups such as GreatAsic, which is betting on indigenous chip design, and FusionAP hint at what an upgraded industry might look like, though they remain small against the scale of the task.

The constraints are real. The US is weighing tariffs under Section 232, a trade law that lets Washington impose duties on national security grounds, with the second phase of the measures still pending. Malaysia also faces shortages of skilled workers, electricity and other resources, pressures sharpened by the data centres crowding into Johor and competing for the same power grid.

These, CGS International warned, could limit Malaysia’s ability to fully capitalise on the semiconductor upcycle and attract further investment.

Also Read: 15 Southeast Asian semiconductor startups moving beyond assembly

The record surpluses are worth celebrating. But a boom that lifts exports without lifting wages, jobs or competitiveness is less a transformation than a good season. Malaysia’s task is to make sure the next downturn finds it higher up the ladder than this upturn found it.

The post Malaysia is selling more chips, but it isn’t winning more of the market appeared first on e27.

At about 4.20am Singapore time on October 30, while most of Southeast Asia’s crypto traders were asleep, money began leaving Binance. By the time the region’s markets opened, the world’s largest crypto exchange had suffered the biggest theft in crypto history.

Blockchain analytics firm Chainalysis has now published its reconstruction of the attack. It puts the loss at about US$6.4 billion across multiple blockchains, slightly above Binance’s own estimate of US$6.3 billion. It attributes the theft to TraderTraitor, a North Korean state-backed hacking group that has spent years targeting exchanges, DeFi protocols and investors to fund the regime.

Also Read: North Korea’s hiring scam just cost Binance US$6.3B. SEA’s engineers are on the shortlist

The attribution will get the headlines. The more revealing part of the report is how quickly everything happened. This hack was a race between a thief and the small number of companies that can freeze stablecoins, and neither side fully won.

How the attacker got in

According to Chainalysis, the attacker’s operating wallet on Ethereum was funded through Tornado Cash, a mixing service that hides where money came from. That wallet submitted valid signed authorisations that allowed a Binance hot wallet to run code the attacker had deployed. A hot wallet is an internet-connected wallet that exchanges use to process everyday withdrawals. The attacker then used that code to sweep the wallet’s assets into addresses it controlled. Other affected Binance wallets appear to have sent funds through ordinary signed transfers.

The word “valid” matters. No cryptography was broken. Binance’s systems signed exactly what they were told to sign. The exchange has traced the breach to a compromised laptop belonging to a single DevOps engineer, a point e27 examined in detail last week. Whatever the entry point, the outcome is the same: once an attacker can make an exchange authorise transactions, key security stops mattering.

A 54-minute sprint against the freeze button

Once the funds were out, the attacker faced a split problem. Billions were held in Bitcoin and Ether, which no one can freeze. Billions more were held in stablecoins such as USDT and USDC, which their issuers can blacklist.

So the attacker moved fast. US$500 million in stablecoins went to a wallet set up only to swap them. Over nearly 20 minutes, it converted US$325 million into US$285 million of ETH, giving up roughly US$40 million to get the money beyond the issuers’ reach. Ten minutes later, the remaining US$175 million went to stopover wallets, apparently in an attempt to stay ahead of a blacklist.

It was not fast enough. Freezes began at 5.14pm Eastern Time, 54 minutes after the first outflows. Chainalysis says about US$2.8 billion has been frozen so far. Binance has put its verified frozen total at US$3 billion, of which Tether and Circle together account for about US$2.4 billion.

The other side of the ledger is less reassuring. The attackers still control roughly US$2.4 billion in assets that cannot be frozen.

The awkward truth about who saved the day

The rescue relied on centralisation, the thing crypto was supposed to remove. Stablecoins could be frozen because a company sits behind each one and can switch individual tokens off. In this case, that is a feature. In most other cases, it is a dependency.

For Southeast Asia, this is not an abstract point. Stablecoins are fast becoming everyday financial infrastructure in the region. Nearly half of Asia-Pacific consumers say they are open to using them, and in markets such as Vietnam they have moved from workaround to regulated tool. Yet the power to freeze sits with a handful of issuers, mostly outside the region. That is part of the dollarisation problem the region has yet to address, and why it cannot simply license its way to stablecoin sovereignty.

Also Read: If Binance says user funds are safe, why did Bitcoin still fall to US$77,037.20?

Chainalysis also notes that much of the freezing happened independently, with organisations acting in parallel. That worked this time. But it is improvisation, not a system, and the next response may be slower.

Bitget’s fingerprints

The attribution rests on a specific piece of evidence. Chainalysis investigators watched in real time as funds stolen from Binance commingled on-chain with funds stolen in last month’s Bitget exploit, which had already been attributed to North Korea. That earlier heist cost Bitget US$387.5 million. Chainalysis announced its attribution at 12.30am Eastern Time on October 30, about eight hours after the first outflows. TRM Labs has separately said it has “high confidence” that TraderTraitor was behind the attack.

The pattern now covers the top tier of the industry. Bybit lost US$1.5 billion in February 2025 in a hack the FBI attributed to North Korea and TraderTraitor. Bitget followed last month, and now Binance. Size did not protect any of them. If anything, it made them more attractive targets, and it raises hard questions about the security standards exchanges have promised since 2022.

Laundering on autopilot

The laundering itself may be the most significant new detail. Chainalysis says it is seeing patterns it has not seen before. Thousands of transactions have run in a short period, with funds circulating between addresses and often returning to wallets used earlier in the chain, rather than moving in a straight line.

Many addresses follow the same sequence. They receive exactly 1 TRX, the native token of the Tron network, then receive larger amounts, then forward funds onward. On Tron, a new address has to receive a little TRX before it can be used. The identical first deposit points to a script creating and activating wallets in bulk, not a person moving money by hand.

Also Read: Crypto’s new threat is not a hack, but a knock at the door

That changes the economics of investigation. Laundering that runs on software can generate more noise than human analysts can work through, which pushes investigators towards automated tools of their own. It is worth noting that Chainalysis sells exactly those tools. Its report ends with a pitch for its KYT and Address Screening products. That does not make the analysis wrong, but readers should keep it in mind.

What comes next

Chainalysis expects the stolen funds to keep moving in the coming days. Binance has pledged to cover all losses from its SAFU reserve and company funds. The wider industry, meanwhile, is reconsidering whether so much money should sit with centralised exchanges at all. Standard Chartered has suggested the hack could accelerate a shift towards decentralised exchanges.

That debate will take years. The theft took less than an hour, and the laundering is now running automatically.

The post Binance’s US$6.4B hack was decided in 54 minutes, and the chase isn’t over appeared first on e27.

For years, Southeast Asian founders were told to win their home market before expanding abroad. Prove the product, build the team, establish revenue, then move into neighbouring countries.

That advice still makes sense. But for startups seeking institutional venture capital, it is becoming less sufficient.

Investors are paying closer attention to the limits of domestic markets, the cost of entering new countries and the size of the eventual exit opportunity. A startup can build a strong national business and still be too small to produce the returns expected by a venture fund.

This does not mean founders should open offices across ASEAN immediately. Premature expansion can drain cash and distract management. It means companies increasingly need to be designed so they can cross borders when the time is right. Regional ambition must be separated from regional sprawl.

The domestic ceiling

The point at which a startup outgrows its home market depends on the country and sector.

Indonesia offers enough domestic scale for some consumer companies to grow for years without expanding. Singapore provides capital and high-value customers but has a limited population. Thailand and Malaysia can support substantial businesses, although specialised startups may work through their largest potential customers relatively quickly.

The constraint is often clearest in business-to-business markets. A startup selling software to Thai insurers, Malaysian hospital groups or Singapore retailers may secure prestigious clients but still face a limited number of similar buyers. The business may be profitable, but its next phase of growth requires another product category, customer segment or country.

Capital creates further pressure. Larger funding rounds often involve investors with regional or international mandates. These funds need to see an opportunity large enough to absorb their capital and generate a meaningful portfolio return.

The same applies to exits. A company serving one narrow national market may have only a limited number of strategic buyers. Regional revenue can widen that pool, but only when foreign operations strengthen the core business. A collection of loosely connected subsidiaries does not create regional scale. It creates complexity.

Also Read: Half the market will not take your money: The access problem for regulated startups

One region, many markets

Southeast Asia looks highly attractive in aggregate. ASEAN has about 684 million people, while the region’s digital economy exceeded US$300 billion in gross merchandise value in 2025, according to Google, Temasek and Bain.

But these totals can be misleading. Southeast Asia is not one unified market. The region contains different languages, legal systems, income levels, payment habits, distribution networks and rules governing data and foreign investment.

A fintech product approved in Singapore may require extensive regulatory work in Indonesia. Software sold in English to Malaysian companies may need a different interface and sales process in Vietnam. A consumer product promoted through social commerce in Thailand may require different pricing and distribution in the Philippines.

ASEAN has made progress towards deeper digital integration, including concluding negotiations on its Digital Economy Framework Agreement in 2026. But implementation still lies ahead.

For startups, the implication is simple: regional expansion is rarely a copy-and-paste exercise. The common core must travel. The edges must adapt.

Build for the second market

A regional strategy begins before a startup registers its first foreign subsidiary.

Technology should be configurable across currencies, languages, tax systems, payment providers and data requirements. If assumptions from the home market are embedded throughout the product, entering another country may require an expensive rebuild.

Contracts and partnerships also matter. An exclusive distributor may provide immediate revenue but restrict future options. A corporate investor may open doors while making competitors reluctant to become customers. A large client may demand so much customisation that the product becomes harder to sell elsewhere.

A useful partnership makes the next sale easier. A poor one leaves the startup supporting another bespoke product and another set of restrictions.

Branding should also travel, but founders should not confuse an English website with an exportable business. Consumer marketing, onboarding and support usually require meaningful localisation.

The deeper test is whether the company’s success can be repeated. A startup may appear scalable because its product is digital, while its sales still depend on the founder’s personal relationships. Overseas pilots can create impressive logos without proving that customers can be acquired consistently or profitably.

The second market reveals whether the company has built a transferable product or merely benefited from local conditions.

Also Read: Investors do not just fund startups. They fund predictability

Why expanding early can fail

Investors’ preference for regional scale should not be accepted blindly. Venture funds need large outcomes. Their economics can push founders towards expansion before the company is ready.

Some businesses are also inherently national. Financial services, healthcare, education and government technology often depend on local licences, regulations and institutional relationships.

Local network effects may not travel either. A marketplace with strong liquidity in Bangkok gains little from building a thin network in Jakarta. A logistics company may create more value by increasing route density at home than by launching scattered operations abroad.

Expansion can also hide weakness. A startup with poor retention may compensate by adding customers in new countries. A series of foreign pilots may generate announcements but little recurring revenue.

The strongest domestic company is often more valuable than the fifth-largest competitor in several countries.

Choose the second market carefully

A credible regional strategy does not begin with a map covering all of ASEAN. It begins with a reasoned choice of the next country.

The best second market usually shares customers, regulations, supply chains, language or commercial partners with the first. A Thai enterprise software company may choose Malaysia because procurement can be conducted in English and multinational customers operate in both countries. A Singapore fintech may retain its regulatory base at home while targeting larger customer pools elsewhere. A Vietnamese consumer company may accept the greater cost of entering Indonesia because the potential scale justifies it.

The best second market does more than add revenue. It lowers the cost or risk of entering the third. A customer in Malaysia may become a reference in Singapore. Approval from one regulator may support an application elsewhere. A regional distributor may open several markets through one relationship.

The number of countries is therefore a poor measure of success. A better question is whether each new market creates assets that can be reused.

Also Read: Startups keep scaling ops before they scale data — Here’s why it backfires

What investors should test

Cross-border revenue is useful evidence, but its quality matters. A recurring contract secured at full price is more meaningful than a discounted pilot arranged through the founder’s network.

Investors should ask:

Does the same customer problem exist in the next market? Will buyers pay a similar price? How much of the product must change? Can the existing team support the launch? Will localisation damage margins?

They should also examine what remains centralised. Engineering, finance and data may support the whole group. Sales, compliance and partnerships may require local control.

A company has achieved regional scale when additional markets reinforce a common system. It has merely expanded geographically when every country adds roughly as much complexity as revenue.

Southeast Asia will continue to produce successful domestic startups. Not every company needs venture capital, and regional expansion should not be treated as a measure of ambition.

But founders seeking large institutional rounds increasingly need a credible route beyond the first country. That does not require immediate expansion. It requires early decisions that preserve the possibility.

The question is no longer simply whether a startup can win at home. It is whether entering the next market will make the company more valuable, or merely more complicated.

—

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 In Southeast Asia, winning at home may no longer be enough appeared first on e27.

Ask a Hong Kong engineer why their product deserves a place abroad and you will often hear the same answer: the city is the hardest test bed they know. Towers scramble satellite signals, space is scarce, labour is expensive and safety rules leave little room for error. Survive that, the argument goes, and most other cities feel easy.

Three companies travelling with the Hong Kong Science and Technology Parks Corporation (HKSTP) delegation are taking that pitch to Singapore Urban Mobility Week, taking place this week from 4 to 6 November. Trunk Tech builds driverless trucks for ports. FJDynamics delivers 3D scanners and autonomous robots Aerovision Technology Limited (Aerovision) makes thermal detection systems, driving simulators and satellite-jamming detectors.

One already runs a live deployment at Tuas Port. The other two are still knocking on doors. All three see Singapore as Hong Kong’s closest twin: dense, wealthy, short of manpower and obsessive about safety. Whether that kinship turns into contracts is what the week will test.

Also Read: Waymo’s Singapore entry raises the stakes for autonomous mobility in Asia

Trunk Tech: Taking the driver out of the port

Founded in 2017, Trunk Tech builds Level 4 autonomous trucks and the software that runs them as a fleet. Dr An Lifeng, Partner and Head of Product Center, frames it as a two-step plan. “Our first step is to create an AI driver, one that can replace the human driver, moving cargo and handling loading and unloading more safely and for longer,” he says. “The second step is to build an AI fleet manager.”

Each truck reads its surroundings and makes its own decisions, but it also shares what it sees with the fleet, which computes the best move for every vehicle. At a container terminal, the operating system assigns hauling jobs, the fleet plans routes and time windows, and the trucks handle exceptions themselves.

The headline number comes from China. Trunk Tech helped design a new terminal with no drivers from day one, supplying the vehicles, roadside units and cloud platform. “The smart zero-carbon terminal reduces staffing levels by 60 per cent versus traditional terminals.” An says. 

Singapore is familiar ground. Trunk Tech hosted Economic Development Board officials two years ago, then met more than 30 representatives from government, port groups, Changi Airport Group and YCH Group last year. It also became a licensed supplier to PSA and began volume deliveries for Tuas Port, which An calls “an extremely large-scale project”. The system is now in commercial operation.

The company now wants a local partner for truck integration, project delivery and on-site support. Its next frontier lies beyond the terminal gates: hauling between terminals and industrial parks, then intra-city and point-to-point freight.

Also Read: PSA, Granite Asia launch US$50M fund to scale supply-chain innovation

FJDynamics: Digitising the city, one scan at a time

FJDynamics was founded in Hong Kong. It now has about 1,000 employees, more than 60 per cent of them in R&D, and serves customers in more than 120 countries. Its technology portfolio includes high-precision positioning, 3D LiDAR scanning, digital mapping and intelligent equipment solutions for construction, agriculture, landscaping and facility management.

Group CFO Kiki Zhang keeps the philosophy plain. “Innovation should not be complicated or expensive,” she says. “It should solve real problems, improve productivity and make people’s jobs easier and safer.”

In Singapore, FJDynamics is exploring opportunities with the Building and Construction Authority and other relevant government agencies, industry stakeholders and local partners on construction digitalisation, virtual inspections and practical technology adoption. It has also undertaken a tree-scanning and digital-mapping pilot involving NUS and is developing a local operational presence to support customers and projects. The company is introducing its crane and heavy-lifting solutions to help improve safety in challenging construction-site conditions. “Singapore is much more than a sales market,” Zhang says.

Her focus is on a particular challenge: the shortage of experienced, skilled workers willing and able to operate in demanding environments such as construction sites, alongside high operating costs and complex infrastructure maintenance needs. LiDAR-based tools can help teams capture reliable site information for inspections and asset management. Intelligent equipment and assisted-operation solutions can reduce repetitive manual tasks and exposure to hazardous working conditions.

FJDynamics has already demonstrated the practical value of many of these solutions in Hong Kong. In Singapore, Zhang hopes to work with government agencies, transport and infrastructure operators, contractors, industry specialists and local companies on more pilot projects and application-led development, backed by the company’s growing local support capabilities. The ambition is to adapt proven solutions to Singapore’s needs and build long-term partnerships that deliver measurable value on the ground.

Also Read: How Hong Kong solves tech’s hardest problem: the final mile to market

Aerovision: From smouldering cargo to rehab clinics

Aerovision’s method, says General Manager Horace Cheung, is to find a pain point in logistics, transport or property management, build a core technology to fix it, then hunt for other industries that need the same fix.

Its thermal detection system began as a way to spot air cargo burning inside with no outward sign. It adjusts to ambient conditions and tells sunlight-heated cargo apart from internal heat, which cuts false alarms. The same system now screens passenger luggage at Hong Kong International Airport and flags water leaks and overheating equipment in buildings.

In Singapore, Aerovision is proposing it to the Land Transport Authority for the Bukit Panjang LRT. The target is “pitting”: arcing between a train’s carbon brushes and the power rail melts small dents into the metal, which raise resistance and generate heat. Abnormal heat patterns tell maintenance crews which rail sections need attention.

The simulators have taken a stranger route. Aerovision built one to train aircraft tow tractor drivers, since aircraft cannot reverse on their own, and a container truck simulator for a local university that won an award in Geneva. Then Singapore visitors saw something else in it: rehabilitation. Force-feedback steering, a motion platform and in-cabin sensors that track performance make it useful for patients recovering from trauma. A partner is now promoting it to hospitals and rehab centres. “It was something we never expected our system to be used on,” Cheung admits.

Singapore is Cheung’s preferred entry point among the region’s markets. “We don’t sell them a canned food solution,” he says. And he wants to see the market first-hand: “I always stick to my own eyes and ears when it comes to verifying and validating the information.”

Also Read: Hong Kong builds a test track for robots that keep failing outside the lab

The HKSTP badge and the Hong Kong stress test

Why travel in a delegation? For Cheung, it is credibility: buyers vet unknown vendors hard, but HKSTP’s screening works as a trusted gateway. An sees Hong Kong, where Trunk Tech opened an AI innovation centre in 2024, as the bridge between mainland firms and global markets.

All three point to the same edge: university talent, mainland manufacturing next door, and a city that punishes weak products. Building its jamming detector among Hong Kong’s towers, Aerovision had to beat signals bouncing off buildings. “If your solution works in Hong Kong, then it will work pretty much anywhere,” Cheung says.

Marking HKSTP’s first dedicated mobility mission in Singapore during Singapore Urban Mobility Week (4–6 November 2026), the HKSTP Pavilion will showcase 10 Park companies featuring breakthrough mobility solutions. Leveraging this government-led engagement platform, the mission facilitates direct engagement with key transport authorities, industry leaders, and potential partners across Singapore’s mobility ecosystem.

The post HKSTP drives Hong Kong mobility innovators to test and scale in Singapore appeared first on e27.

Most small business owners I speak to compare website quotes the same way: they line up the first invoice and pick the lowest. It feels sensible. It’s also how many end up paying several times more than they planned.

In October 2026 we looked at the published prices of popular website packages in Malaysia and added up what each costs over three years: the setup or package fee, plus every renewal or monthly fee in that time. The difference between the first invoice and the three-year total depends almost entirely on how the package is billed.

One-off packages stay close to their sticker price. A five-page package priced at RM1,980 (about US$475) with a RM462 (about US$111) renewal from year three totals about RM2,442 (about US$586) over three years: 1.2 times the first invoice.

Monthly plans aren’t bad value in themselves. They usually bundle hosting, maintenance and edits that a one-off package charges for separately or not at all. The problem is comparing a one-off price against a monthly plan’s setup fee, as if they were the same thing.

Three habits fix most of this:

• Turn every quote into a three-year number. Ask: what do I pay in year one, year two and year three, and what does each payment include? Put the totals side by side.
• Check what happens to the domain. Promotional domain prices usually apply to the first year only. .my domain offers advertised at a few US dollars typically renew at around US$25 to US$39 a year, and Malaysia’s domain registry raised .com.my prices by RM5 (US$1.20) in March 2024. Make sure the domain is registered in your company’s name, so you can leave without paying to get it back.
• Ask what you can’t do without paying. Text changes, new photos, an extra page, a second language. A plan that includes them may be cheaper over three years than one that bills each change separately.

Also Read: Malaysian pension fund KWAP moves to contain damage after eFishery fraud shock

There’s a second cost that rarely appears on a quote: a website nobody can find. A site that isn’t indexed by Google, or doesn’t track which enquiries it brings, can be cheap and still poor value. When you compare quotes, ask how the provider will check that your important pages appear in Google and how you’ll know what the site brings in.

None of this needs expertise, just a spreadsheet and three questions. The businesses that get burnt are rarely the ones that paid the most; they’re the ones that compared the wrong number.

Prices are taken from providers’ published price pages, checked on 19 October 2026. Three-year totals are shown before SST.

—

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 Malaysian SMEs are comparing the wrong number when they buy a website appeared first on e27.

Washington has spent the past year making it harder, pricier and more nerve-racking for foreign talent to work in America. Asian founders, if a new survey is to be believed, have mostly shrugged.

Singapore-founded immigration platform Beyond Border’s August 2026 Founder Mobility Pulse found that 82 per cent of respondents had not changed their US expansion plans despite immigration uncertainty. Another 11 per cent were either bringing their timelines forward or exploring other visa routes.

Also Read: Europe’s trade truce with America has not restored trust in American business

On its face, that is a show of founder grit. It also comes from a company whose business is helping founders get US visas, which is worth keeping in mind before anyone treats it as a verdict on the whole region.

Customers, not capital

The headline motivation is commercial. Some 85 per cent of the 307 founders surveyed, more than 200 of them based in Asia, said access to US customers was their main reason for expanding. Only 8 per cent pointed to fundraising or investor networks, with talent and partners (5 per cent) and compliance (2 per cent) trailing.

The timelines are aggressive. Sixty per cent plan to land in the US within 12 months and a further 25 per cent within two years.

That customer-first logic makes sense, but the sample explains part of it. Roughly 83 per cent of respondents are at seed (35 per cent), bootstrapped (28 per cent) or pre-seed (20 per cent). Only 5 per cent have reached Series B or beyond.

Companies that have never raised institutional money are unlikely to cross the Pacific for American VCs. For them, the US is a sales territory, not a cap table strategy. Singapore’s MetaOptics, which is eyeing the US with a potential US$10 million warrant facility, is closer to the later-stage exception than the survey’s typical respondent.

A crackdown that hits some doors harder than others

The policy backdrop is not subtle. Late in August, the Trump administration proposed a US$103,265 annual fee for new H-1B visas, following an earlier US$100,000 fee that a court struck down. The report also notes recent moves affecting green-card sponsorship at major tech companies and investigations into university visa programmes.

Beyond Border’s reading of US Citizenship and Immigration Services (USCIS) data suggests the squeeze is uneven. Between FY2022 and FY2024, approved L-1 petitions, the route companies use to transfer staff, fell 31 per cent, while O-1/O-2 approvals for people with “extraordinary ability” rose 1 per cent. Between FY2023 and FY2025, EB-1A green-card approvals climbed 20 per cent, while EB-2 National Interest Waiver (NIW) approvals fell 39 per cent.

There are two caveats. The comparison windows differ by pathway, and none of the series fully captures the latest round of tightening. Approval counts also measure what the government granted, not what founders wanted.

Also Read: Taiwan’s startup talent problem is a matching problem, not a shortage

Then there is the inquiry data. Beyond Border analysed 1,140 immigration inquiries from Asia-based individuals between August 2025 and September 2026 and found that about 45 per cent (510) concerned the O-1 visa. That is a useful signal, but O-1, EB-1A and EB-2 NIW are also the pathways Beyond Border says it specialises in. A platform built around a visa will attract inquiries about that visa. e27 has previously explored why Beyond Border sees visas as part of the founder playbook; this survey is, in part, an extension of that pitch.

The founder as a one-person beachhead

The most telling finding concerns who actually makes the move. Thirty-five per cent expect the founder to relocate alone, and 29 per cent expect the founder to go with a small launch team. Only 16 per cent plan to put a US-based hire on the ground instead.

That is lean, and potentially punishing. A founder who sells in New York by day while managing a team in Jakarta or Bengaluru by night is running two companies across a 12-hour time difference. The report raises the right question but leaves it unanswered: when does a founder-led foothold need to become a real local operation? Anyone who has faced cross-border due diligence questions will know that American buyers and investors eventually expect more than one person and a co-working desk.

Asia keeps the engine room

For Southeast Asia, the more consequential number may be 72 per cent: the share of founders who intend to keep engineering and product in Asia. Many also plan to keep operations (40 per cent), sales and marketing (30 per cent) and senior leadership (20 per cent) in the region.

That supports the region’s engineering employers, but it also highlights an old pattern. When customers and revenue are in the US while code is written in Asia, cost arbitrage often explains as much as talent strategy. It also adds pressure to a pipeline already under strain from AI, as the junior developer dilemma makes clear.

The returnee numbers are real, but modest

The report’s final section looks at talent flowing the other way. Recruitment firm Xpheno estimates that 12,700 white-collar tech professionals returned to India from the US in the first nine months of 2026, compared with 15,100 in all of 2025, 9,800 in 2024 and 7,200 in 2023. At the current pace, 2026 would come in at around 17,000, which is a rise but not an exodus.

For scale, the US National Science Board says India and China accounted for 29 per cent and 12 per cent respectively of foreign-born workers in US science and engineering occupations in 2024. Beyond Border acknowledges that this says nothing about how many might actually leave.

The report names Singapore as a potential beneficiary. That is plausible, but returning talent has a habit of leaving again. Vietnam has already discovered that keeping returning engineers is harder than attracting them.

Also Read: Malaysia wants 300,000 AI jobs by 2030. Talent will decide if it gets there

Beyond Border’s own conclusion is the most defensible line in the report: tighter US immigration policy alone will not bring talent back. Pay, career progression and serious work in fields such as AI will. On that front, the US still competes better than most of Asia, whatever happens to the visa fee.

The post Visa fees are soaring, but founders want US revenue more than US investors appeared first on e27.

I run OTSATE, a small lemongrass sauce brand in Malaysia. I’m the founder, the cook and the person who does the numbers, so everything below is one operator’s experience, not a benchmark.

Most of my admin now runs through Claude Code, an AI agent that works inside a terminal and can read my files and write to my spreadsheets. I’m going to describe one task it does well, one moment it got badly wrong, and the one column it will never fill in.

The task it now owns: Monday numbers

Every Monday I give it four things. Sales last week. What I spent on ingredients and packaging. Cash in the bank today. The decision I’m making this week.

It writes the row into a Google Sheet and works out the rest: revenue per day, food cost as a percentage of sales, and how many weeks of fixed costs the week’s sales would cover. A small script it wrote reads my bank statement PDFs and sorts each transaction into a category. Every category it guesses gets flagged until I confirm it, because it cannot tell a lemongrass supplier from a water bill without being told once.

The honest reason this works: I avoided my own trackers for months. Typing a bad week into a cell felt like a verdict on me. Saying “sales were low” out loud to a terminal and watching the row app still get recorded. That is the whole point.

What I stopped doing is opening the spreadsheet to type. What I do with that time is read it. Ten minutes, every Monday, and I actually look.

Also Read: I used AI to rebuild my websites without knowing how to code

The moment it was confidently wrong

In July I asked it to cost a new product: Kam Heong sauce in 100g retort pouches, five for RM45 (about US$10.80). I gave it the recipe and asked for the contribution margin.

It came back with a clean, well-formatted answer. Contribution of 12 to 19 per cent. Reprice to RM60 to RM65 (about US$14.40 to US$15.60) for five, or don’t launch.

I nearly believed it. The table looked like something a consultant would charge for. Then I read the cost lines.

It had priced every ingredient at what you’d see on a Malaysian supermarket shelf or an online grocer. Dried shrimp at RM90 (about US$21.60) a kilo. I pay RM37.50 (about US$9) at the wet market. Oyster sauce at RM21 (about US$5.04) a kilo. I pay RM4 (about US$0.96) in bulk. Fresh chilli at RM14 (about US$3.36). I pay RM5.63 (about US$1.35). Every single line was high, most by double or more, and the errors compounded through the recipe. It also added a per-pouch fee for outsourced retort processing. I own the retort machine.

With my real prices the contribution was 61 per cent. That’s all. The model wasn’t lying. It never asked where I shop, and I never told it, and it filled the gap with the most visible number on the internet.

In late September it did a smaller version of the same thing. It suggested I fry my own garlic instead of buying it fried, to bring food cost down. On paper, correct. But I don’t draw a salary. My labour appears nowhere in the P&L. Buying fried garlic at RM5 (about US$1.20) per 100g is how I force my own time into the ingredient cost so the pricing stays honest. The tool optimised the number it could see and ignored the one it couldn’t.

What changed about how I check its work

I gave it a standing rule, saved in the memory file it reads at the start of every session: never estimate my costs. Every line in a costing is either a price I supplied or marked OPEN. No recommendation on top of an OPEN line. If the answer depends on something it doesn’t know, it has to say so and ask.

The second rule is on me. I read the inputs before the conclusion now. A margin figure is only as true as the ugliest line underneath it, and the tool will format an ugly line just as nicely as a true one.

Also Read: How AI became my sparring partner

The column it can’t fill

The Monday sheet has a column called “the decision this week”. It is always blank until I write in it.

In August I cancelled a December launch for a nyonya achar I had been testing in five-kilo batches. The costing was fine. The product was good. I cancelled it because every jar needed days of sun and hours of my hands, and when I pictured making a thousand of them I did not want that life. No spreadsheet carries that information.

If I let the tool decide, it would pick whichever option lines it had. The costing story above is exactly why that’s dangerous: the lines it has are not always the real lines. And it has no stake. It will not be the one at the stove in December.

What shifted with people

My husband sits as an informal advisor on the business. The Monday sheet has a column for the one question I’m bringing to him that week. Our conversations got shorter and more specific, because the numbers arrive before the feelings do. The friction is that he can now see a bad week in a cell, same as me. We are both still learning to talk about that.

If you run something small

Three things from doing this.

Give the tool your real inputs or it will invent plausible ones. Plausible is worse than blank.

Let it do the typing and keep the reading for yourself. The reading is where the business is.

And write down the decisions it is not allowed to make. Mine fit in one column. Yours probably do too.

—

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 AI told me to raise my prices by 40 percent but it had never been to a wet market appeared first on e27.

Bitcoin has climbed back above the critical US$81,000 baseline. This 24-hour rebound does not erase a difficult month. The digital asset remains in its weakest October performance in six years, down roughly 4.2 per cent, while conflicting forces keep price action in a sideways range. The market has spent a long time without a decisive break. That pattern reflects a fierce tug-of-war between strong rebound catalysts and firm upward caps, all against severe macro headwinds and a looming inflation report.

The first support comes from the Federal Reserve. Risk appetite expanded after the central bank voted 10-2 to leave the federal funds target range steady at 3.75 to 4.00 per cent. Avoiding a pre-election liquidity shock triggered an immediate relief rally. Resilient institutional inflows add another layer of stability. Spot US Bitcoin ETFs pulled in US$110.34 million in net inflows on October 26, followed by another US$176 million on October 27. That steady backing provides a firm price floor.

Several forces actively suppress upward movement. Options traders brace for dramatic swings as a massive US$11 billion options expiry closes this Friday. With 70 per cent of dealer hedging sensitivity, known as gamma, concentrated heavily between US$75,000 and US$82,000, market makers actively suppress large expansions. US demand has turned noticeably soft.

The Coinbase Premium Index, which measures the gap between US and offshore spot buying, has not posted one positive day all of October and hit its lowest levels since June. Technical resistance failure compounds the issue. On-chain analysts note that Bitcoin recently suffered a clear fakeout near the US$83,000 mark. Failure to reclaim and hold this baseline keeps technical day traders on the defensive.

Also Read: The Fed decides today: Here are the 2 levels every Bitcoin trader should watch

Bitcoin is defying significant macro headwinds. While the Fed paused, Chair Kevin Warsh shifted language to suggest the economy is expanding strongly rather than solidly. At the same time, two policymakers dissented in favour of an immediate hike, pushing market odds for a December rate increase up to nearly 75 per cent.

Crude oil and yield pressures loom large. Brent crude remains stubbornly elevated at US$100 to US$111 per barrel amid ongoing Middle East war headlines. This energy shock keeps the 10-year US Treasury yield near a multi-decade high of 5.1 to 5.3 per cent, intensifying competition for non-yielding assets.

Midterm election de-risking adds another layer. Capital is aggressively de-risking ahead of the November 3 US midterm elections. A highly polarised outlook dictates vast price variances. Institutional consensus projects year-end targets of US$90,800 under a unified Republican Congress, versus US$76,450 if Democrats sweep both chambers.

The immediate catalyst is the September 2026 Personal Consumption Expenditures price index report, the Federal Reserve’s preferred inflation gauge. Economists project a sharp acceleration in price pressures. The US Bureau of Economic Analysis releases the data at 8:30 AM Eastern Time (8:30 PM Singapore Time).

Consensus models anticipate a broad-based monthly and annual tick-up. The September 2026 forecast for the headline PCE price index shows a +0.5 per cent month-over-month increase, up from the +0.3 per cent actual in August. The year-over-year forecast stands at +3.7 per cent, or roughly 3.6 per cent, up from the 3.4 per cent actual in August.

Analysts forecast the core PCE price index, which excludes food and energy, to rise +0.3 to 0.32 per cent month-over-month, up from the +0.2 or 0.25 per cent actual in August. Annually, core PCE will likely rise +3.1 per cent, up from the 3.0 per cent actual in August.

Also Read: 3 scenarios for October 28: Which one breaks Bitcoin’s US$80,000 resistance

Several key drivers explain the expected jump. Higher crude oil and natural gas prices, driven by supply disruptions in the Middle East, pushed energy costs up. Oxford Economics estimates PCE energy prices rose roughly 5.4 per cent in September alone. Persistent core services and goods remain sticky.

Following hot Consumer Price Index and Producer Price Index releases earlier in the month, core categories have proven stubborn. Bank of America analysts noted that price pressures span both core goods and core services. Higher input and component costs tied to ongoing corporate artificial intelligence buildouts continue to filter into the computer software, hardware, and accessory sub-indices.

This report drops less than 24 hours after the Federal Open Market Committee voted 10-2 to hold interest rates steady at 3.75 to 4.00 per cent. A hot PCE print confirming headline inflation running at 3.7 per cent and core at 3.1 per cent, well above the Fed’s stubborn 2 per cent target, will validate the two dissenting votes for an immediate hike. Interest rate futures have already priced in roughly an 85 to 91 per cent chance of a rate hike by December.

An upside surprise in today’s PCE figures could cement expectations that the central bank will resume tightening sooner than risk assets prefer. A hotter-than-expected print will likely lift US Treasury yields, keeping the 10-year near 5.0 per cent or higher, pressure non-yielding assets like Bitcoin and gold, and spark immediate turbulence across equity index futures and the USD/EUR exchange rate.

Also Read: 92% of Bitcoin liquidations were shorts: What happens when the squeeze ends?

Global financial markets closed slightly higher after the Federal Reserve split on keeping the benchmark interest rate steady. The S&P 500 rose 0.28 per cent to 7,743.98. The Nasdaq Composite gained 0.27 per cent to 26,549.85. The Dow Jones Industrial Average added 0.19 per cent to 52,475.10. Major US indexes posted their third straight daily gain during regular hours but stayed mostly range-bound. The initial relief rally has given way to sharp divergence and tech-sector volatility in early-morning extended trading as investors closely scrutinise mega-cap AI spending.

In early Asian trading and overnight futures, Nasdaq 100 futures have dropped over 1 per cent. The Federal Open Market Committee voted 10-2 to keep interest rates unchanged, as expected. The tone was subtly hawkish. Governor Lisa Cook and Cleveland Fed President Beth Hammack dissented, actively voting for an immediate 25-basis-point hike.

The Fed also upgraded its economic assessment to “expanding at a strong pace,” up from “solid,” while noting that inflation remains elevated. Chair Kevin Warsh avoided forward guidance but emphasised readiness to act. The CME FedWatch tool now shows an 85 per cent probability of a December rate hike. Warsh announced an adjusted 2027 calendar, reducing rate-setting votes to seven of eight annual meetings.

In my view, Bitcoin sits at a precarious intersection. Resilient institutional inflows and aggressive macro headwinds define the current sideways range. The US$81,000 level is a temporary resting point, not a sign of strength. The impending PCE report and the massive options expiry will likely dictate the next leg of volatility. The token has gone too long without a decisive break, and until the Fed pivots or Bitcoin breaks decisively above US$83,000, it remains hostage to macro data and derivative positioning.

I believe, and I said that yesterday, that the path of least resistance remains downward in the near term, given hawkish Fed rhetoric and technical failures, while the long-term debt-devaluation narrative provides a formidable floor. The September PCE report will serve as the immediate arbiter, determining whether Bitcoin can sustain this rebound or succumb to mounting pressure.

—

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 Bitcoin rebounds to US$81,000 but faces weakest October in 6 years appeared first on e27.

I use AI almost every day. Sometimes, several times a day.

As someone managing a digital publication and overseeing marketing for an esports business, I’ve come to rely on ChatGPT (and sometimes Claude) for everything from writing and research to campaign planning and website troubleshooting.

It’s saved me countless hours, particularly when it comes to producing first drafts, organising information and working through tasks that would otherwise take considerably longer.

But there’s one thing I’ve learnt from using AI so extensively: just because it sounds confident doesn’t mean it’s right. And the more useful AI becomes, the easier it is to forget that distinction.

When AI gives you an answer that looks perfectly reasonable

One of my more memorable experiences involved something completely unrelated to work: checking Singapore’s TOTO lottery results. I’d asked ChatGPT to help me check the winning numbers against a particular set of numbers on a regular basis.

It seemed like a straightforward task. The results are publicly available, and comparing two sets of numbers hardly requires complicated reasoning. Especially when I had provided the URL to the results page. Except that there were several occasions when the results it provided were incorrect.

What made this frustrating wasn’t simply the mistake. It was the way the information was presented, as though it were factual and verified. I had to question the results and check them against the official source.

It might sound like a trivial example, but it illustrates a much bigger problem. If AI can confidently get something as straightforward as lottery results wrong, what happens when we’re using it to research business opportunities, interpret industry developments or prepare information that other people will rely on?

The issue isn’t necessarily that AI makes mistakes. People make mistakes too. It’s that AI can present inaccurate information in language that gives us very little reason to doubt it.

The more you use AI, the more tempting it becomes to trust it

When I first started using ChatGPT regularly, I approached its responses with a fair amount of curiosity and scepticism. I’d read through the output carefully, compare it with what I already knew and make adjustments.

But once you’ve used a tool successfully hundreds of times, it’s natural to become more comfortable with it. You start recognising how much time it saves. You learn which prompts work. You become accustomed to receiving useful answers almost immediately. And that’s where complacency can creep in.

For The Wellness Insider, the digital wellness publication I run, I frequently use AI to help turn press releases into articles, prepare SEO information and adapt content for social media. It’s an incredibly useful starting point, but I still need to ensure that names, dates, statistics and claims are accurate. A well-written article containing incorrect information is still an incorrect article.

As a publisher, I’m responsible for what appears on my platform. I can’t simply blame ChatGPT if something inaccurate gets published. That responsibility doesn’t disappear because the first draft was generated by AI.

Also Read: Muggles struggle, wizards juggle: What happens when everyone gets an AI familiar

Sometimes, the problem isn’t accuracy. It’s continuity

Another frustration I’ve encountered is AI losing track of instructions during longer projects. This has been particularly noticeable when working on website development and marketing campaigns.

It’s a strange experience because AI can be remarkably capable at solving a technical problem, yet occasionally struggle to maintain continuity across an extended project. The challenge is that its conversational nature creates an impression of shared understanding.

When you’ve spent hours discussing a project, you naturally expect the other party to remember what you’ve agreed upon. But AI doesn’t necessarily retain or apply every instruction consistently, particularly when a conversation becomes lengthy or complex.

I’ve learnt that important decisions sometimes need to be restated, and completed work needs to be verified rather than assumed. In practice, that means treating AI-generated work as something that requires review, not as a finished deliverable simply because the system says it’s done.

AI can help me make decisions, but it shouldn’t make them for me

One of the ways I find AI most valuable is as a sounding board. When I’m considering a sponsorship proposal, developing a marketing campaign or deciding how to position a service, I can use ChatGPT to explore different approaches. It can help me identify potential weaknesses, organise my thinking and consider alternatives I might have overlooked.

But it doesn’t know everything about my businesses. It doesn’t have the same understanding of my relationships with clients, the nuances of a particular negotiation or the practical constraints I’m dealing with at any given moment.

For example, AI might suggest an ambitious marketing campaign with multiple content formats, outreach activities and promotional channels. On paper, the strategy might look sensible. But if I’m the person responsible for executing most of it, I have to consider whether it’s realistic. A campaign that requires more time, money or manpower than I have available isn’t necessarily a good campaign, regardless of how impressive the plan looks.

AI can help me evaluate options. It can’t decide which trade-offs I’m willing to make. That’s still my job.

Also Read: How we use AI to triage inbound B2B leads without blasting everyone the same way

I’ve learnt to manage AI rather than simply use it

The longer I’ve worked with AI, the more I’ve realised that getting value from it requires a certain amount of management. I need to provide clear instructions, explain the context, establish expectations and review the results.

When something is inaccurate, I have to correct it. When the output doesn’t fit my brand, I have to refine it. And when it makes a recommendation that doesn’t make sense for my business, I need to be comfortable rejecting it.

In some ways, this resembles managing a junior colleague, although the comparison has its limits. A human colleague can develop contextual understanding through experience, take responsibility for decisions and recognise consequences in ways that an AI system cannot. AI also doesn’t necessarily know when it’s wrong.

That’s why I think one of the most important skills for anyone using AI isn’t prompt writing. It’s judgement. You need enough understanding of the task to recognise when the output doesn’t make sense.

For factual information, that means checking reliable sources. For technical work, it means testing the results. For business decisions, it means evaluating the recommendations against your actual circumstances.

The amount of checking should also depend on the consequences of getting something wrong. A mediocre social media caption is relatively easy to correct. An inaccurate health article, broken website function or misleading business proposal could have much more serious implications.

I still wouldn’t want to work without it

Despite these frustrations, I remain an enthusiastic AI user. It’s made a meaningful difference to how I work, particularly as someone handling responsibilities that would ordinarily be spread across several people.

I can produce first drafts faster, explore ideas more easily and tackle technical projects that I previously wouldn’t have attempted on my own. It has also really helped me gather my thoughts and opinions when I’m doing a product or service review that may span over a few days or weeks. Those benefits are real. But so are the limitations.

I don’t think the solution is to stop using AI or to distrust everything it produces. That would mean ignoring a genuinely useful tool. Instead, I’ve become more deliberate about deciding when to rely on it and when to slow down and verify the output.

For routine drafting and brainstorming, I’m comfortable allowing AI to do much of the initial work. For factual claims, technical changes and important business decisions, I take a more cautious approach.

Ultimately, AI has changed how I spend my time, but it hasn’t removed my responsibility for the quality of my work. I may have gained an incredibly useful assistant, but I’m still the one who has to make sure it gets things right.

And that’s probably how it should be.

—

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 saves me hours, but I don’t trust it blindly appeared first on e27.