Enterprise Adoption · · 9 minute read
Adoption is up almost everywhere in Asia. Budgets are up too. Look at what the money is attached to and a quieter story appears: most organisations are still buying tools, and very few are buying redesigned work.
Almost every board I sit in front of in this region has already approved an AI budget.
That is new. Two years ago the question in the room was whether to spend. Now the money is committed, the line item exists, and somebody senior has been made accountable for it. Which means the interesting question has moved. It is no longer whether Asian enterprises are buying AI. It is what, precisely, they think they are buying.
The answer, read across the region’s own survey data, is uncomfortable and useful in equal measure. Asia is buying a great deal of AI. Most of it is not attached to any work that has changed.
Start with the number everyone quotes. In a ServiceNow survey of 200 senior leaders in Singapore published in August 2026, agentic AI adoption reached 51 per cent, up from 22 per cent a year earlier. Read on its own, that looks like a transformation happening at speed.
The same survey says three other things. Thirty-three per cent use AI to help employees with day-to-day tasks. Ten per cent have reworked their processes so that AI can complete a multi-step business task from end to end. Eighteen per cent report no progress at all on advanced adoption.
So roughly half of Singapore’s large organisations can honestly say they have adopted agentic AI. One in ten can say they changed how work moves. That gap is the entire story of enterprise AI in Asia right now, and it is a purchasing story before it is a technology story.
Singapore is not a laggard here, which is what makes the figure worth taking seriously. Its AI maturity score in that survey rose to 53 out of 100 against a global average of 51. Twenty-three per cent have replaced legacy systems, against 16 per cent globally. Twenty-eight per cent have formal AI governance, against 20 per cent globally. AI accounts for 15.4 per cent of IT spend, and 83 per cent expect to spend more next year. This is a market doing the work properly by international standards, and still only one in ten has bought a redesigned process.
Widen the lens past Singapore and the pattern holds. In a survey of 1,000 IT decision-makers commissioned by Alibaba Cloud across eight Asian markets, Hong Kong, Indonesia, Japan, Malaysia, the Philippines, Singapore, South Korea and Thailand, 95 per cent said they plan to increase AI investment and 75 per cent described AI as indispensable to operations.
The revealing detail is not the appetite. It is the layer the appetite is pointed at. Sixty-nine per cent expected infrastructure spend to rise by more than 20 per cent. Sixty-one per cent said the same of platform services. Fifty-eight per cent said it of model services. Capacity first, platforms second, models third.
Notice what is not in that ranking, because no survey asks about it: the fourth layer, which is the work itself. Compute can be procured in a quarter. Platforms can be procured in two. Redesigning how an underwriting decision, a claims process or a client onboarding actually runs takes a year and requires someone with authority to tell a department that its job has changed. One of those is a purchase order. The other is a political act.
IDC expects AI and generative AI spending across Asia Pacific to grow from 73 billion US dollars in 2024 to 370 billion by 2029, a 38.4 per cent compound growth rate, with generative AI alone reaching roughly 175 billion. That money is going to be spent. The open question is how much of it lands in the fourth layer, and on current evidence the answer is: not much, yet.
Ask Asian enterprises what is holding them back and the ranking is consistent across surveys. Data privacy and security, 48 per cent. High implementation cost, 42 per cent. Not enough internal expertise, 37 per cent. Regulatory and ethical concerns, 31 per cent. Integration difficulty, 28 per cent. In the Singapore survey, 68 per cent named data quality and access, and 58 per cent named privacy and security.
Not one of those is a model-quality problem. No serious AI programme in this region is stalling because the model is not clever enough. They stall on data, cost, people and law.
If you sell AI into Asia, that should reorganise your entire go-to-market. Your demo is not competing against a rival demo. It is competing against the security review queue, the data team’s backlog, a legal function reading three regulators at once, and a CFO who wants a number. The company that wins the deal is usually not the one with the better benchmark. It is the one that arrives with the security answer, the integration path and the business case already assembled.
The buying preferences in the same survey say exactly this. Forty-five per cent want integrated AI and cloud offerings. Thirty-four per cent want models that work across clouds. Sixteen per cent want standalone models. Raw model access is now the least wanted way to buy AI in Asia. Enterprises are buying assembled things.
The most sophisticated buyers in this region have already stopped shopping for a single AI supplier. OCBC has rolled out more than 30 internal tools built on open-weight models, using Google’s Gemma for document summarisation, Alibaba’s Qwen to assist with coding and DeepSeek for market trend analysis, across a bank that operates in Singapore, Hong Kong, Malaysia, Indonesia, Thailand and Vietnam simultaneously.
That is not a procurement decision. It is a routing decision, made per job, and it is driven by two things that matter far more in Asia than in the United States.
The first is price. A RAND analysis published in early 2026 put Chinese models at roughly a sixth to a quarter of the cost of comparable American systems. In markets where margins are thin and the alternative to automation is cheap labour rather than expensive labour, that differential is not a nice-to-have. It decides whether the use case has a business case at all.
The second is language. Models trained overwhelmingly on English do not handle Bahasa Indonesia, Thai, Vietnamese or Tagalog with the same fluency, and every regulator and customer base in the region notices. Hence the sovereign build-out: Malaysia’s ILMU, Indonesia’s Sahabat AI, India’s Sarvam covering 22 languages. These are not vanity projects. They are procurement options that a bank in Kuala Lumpur or an insurer in Jakarta can now put in a shortlist.
There is a counter-example to all of this, and it is worth studying precisely because it is rare. DBS has generated approximately one billion Singapore dollars in economic value from AI, machine learning and data analytics, running more than 1,500 models across 370-plus use cases as of May 2025, since grown past 2,000 models and 430 use cases.
The model count is not the interesting part. Any bank with a budget can accumulate models. The interesting part is that the value is tracked per use case, which means DBS can answer the question most enterprises cannot: not what did this cost, but what did this earn.
Set that against the failure data. MIT’s NANDA initiative reviewed more than 300 publicly disclosed deployments and found that 95 per cent of enterprise generative AI pilots delivered no measurable return. The figure has been contested, and should be, because “no measurable return” often means nobody set up a measurement. A separate 2026 analysis puts the share of pilots that never reach production at 88 per cent. The most cited root cause is not technical: it is the absence of a business objective attached to the pilot on day one, which means nothing ever forces the journey from experiment to deployment.
The distance between one billion dollars of tracked value and a portfolio of pilots that returned nothing is almost never the technology. It is whether a number was named before the work started, and whether somebody owned it.
One more shift, and it is the one I think is most under-discussed in this region. The way AI is priced is coming apart. Seat-based pricing fell from 21 per cent to 15 per cent of software companies in twelve months, while hybrid models rose from 27 per cent to 41 per cent. Outcome pricing has moved from theory to practice: HighRadius abandoned per-seat entirely in February 2026, charging nothing up front and taking a share of measured savings instead.
This matters more in Asia than anywhere, for an unglamorous reason. Per-seat AI pricing is implicitly a bet on the cost of the person sitting in the seat. Thirty US dollars per user per month is trivially justified against a Singapore or Tokyo salary. The identical tool, sold against the identical job in Jakarta, Manila or Ho Chi Minh City, has to clear a labour cost that is a fraction of that. The product does not change. The arithmetic does.
Which is why so many AI products that win in Singapore stall three markets later, and why their founders usually misdiagnose it as a localisation or channel problem. It is a pricing-architecture problem. Consumption and outcome pricing travel across income levels. Seats do not.
If you sell AI in Asia. Sell into a process, not to a person. Arrive assembled, with the integration and the security answer, because standalone model access is the least wanted shape of the purchase. Price to consumption or outcome if you intend to cross more than one market. And put the business case in the first meeting, not the fourth, because cost and data are beating you far more often than a competitor is.
If you buy AI. Your adoption number is already fine and it is telling you almost nothing. The question is where the next dollar goes. One redesigned process with a named number attached will do more for you than ten more tools distributed to willing individuals. And if nobody can state what a pilot is supposed to earn, you have not commissioned a pilot. You have commissioned a demo.
If you lead. Change the question you ask in the management meeting. “Are we using AI” will always return a comfortable yes. “Which process did we rebuild this quarter, and what did it earn” will return the truth, and it is the only version of the question that puts you in the ten per cent.
Half of Singapore’s enterprises can now say they use agentic AI. One in ten can say they changed how work moves. Every dollar of the 370 billion heading into this region over the next four years will land on one side of that gap or the other. It will not be closed by buying more tools.
If you are deciding what your organisation should actually be buying, that is the work I do with leadership teams through Intelligence & Advisory.
The Briefing
Companies, business models and ideas shaping the AI economy across Asia. No hype. No urgency. Unsubscribe in one click.
More writing
The real leap is not better prompting. It is building an operating environment where context, tools, team memory, and boundaries can compound.
Why first-principles thinking outlasts every tool that comes through.