How enterprise AI could reshape Asia’s economies

Cohere announced in August that it is building its Asia-Pacific hub in Seoul. The Canadian AI company, which builds large language models tailored for enterprise use, is expanding in South Korea with plans to grow its regional workforce. Its offer is aimed at companies and governments that need secure, multilingual AI systems and more control over where their data and models sit.

Enterprise AI is becoming a business priority because it will shape how organisations operate and what they can offer customers. Its next phase will be shaped by whether a bank can serve clients more effectively, an insurer can spot fraud sooner, an engineer can build and test software faster, or a public agency can improve a service while keeping people accountable.

That is where the economic stakes sit. When AI enters the systems through which organisations work, it can alter costs, capacity, service quality and the roles people perform. Across Asia-Pacific, where large economies have deep pools of accountants, lawyers, engineers, analysts, developers and customer-service teams, even modest improvements across recurring tasks could add up.

Three tracks of AI adoption

AI is developing along several tracks at once.

Consumer AI is what most people meet first: chatbots, image generation, personal productivity tools, search and creative software. It changes how individuals write, find information and make things.

Government AI covers public-service delivery, national AI plans, data rules, digital infrastructure and the safeguards that shape how agencies use automated systems.

Enterprise AI is embedded in the work of an organisation. Its potential applications range from customer service, financial analysis, software development, legal work and knowledge management to supply chains, healthcare, manufacturing, government services, cybersecurity and internal decision-making. The evidence in this Analysis comes principally from finance, customer service, software and knowledge work.

The boundaries overlap. A chatbot may become a customer-service tool. A government’s cloud rule may decide whether a bank can use a model. Enterprise AI enters a workflow that already employs people, handles information and costs money. Its value comes from changing that workflow well enough for an organisation to keep using it.

The choice of deployment shapes the cost, control and integration work required to achieve that value. It also shapes which organisations are able to capture the productivity gain.

Four paths into enterprise AI

Cohere: private deployment for sensitive work

Cohere represents the enterprise, sovereign and private-deployment path. Its Seoul hub is directed at Korean as well as APAC companies and public institutions that want advanced models without placing sensitive work entirely in a public cloud.

This can suit a bank, manufacturer or government body whose data, security requirements or procurement rules limit where a model can operate. It also places more responsibility on the buyer. Private deployment requires computing capacity, technical staff, security controls and a plan for maintaining the model. Control has a cost.

LG CNS and Cohere have partnered to develop Korean-language models and offer them to clients in finance, manufacturing and retail. In an April interview, Cohere co-founder Ivan Zhang also referred to an AI collaboration memorandum with Hanwha Ocean and Hanwha Systems in shipbuilding and defence. Together, these show supplier and partner activity around private enterprise deployment in South Korea.

Microsoft and OpenAI: distribution through the software firms already use

Microsoft and OpenAI have a different advantage. Their models can reach organisations through the corporate software, cloud contracts and productivity tools many already use.

South Korea's KakaoBank shows how that route can work in a regulated market. Its August earnings release says more than 5 million people have used its conversational AI services, including AI Search, an AI financial calculator, AI transfer and a customer-service chatbot. Microsoft says the bank used Azure OpenAI through South Korea’s Innovative Financial Services sandbox.

The South Korean Financial Services Commission has since eased selected network-separation rules for specified cloud software on internal networks, subject to security conditions. For South Korea’s financial sector, the story is partly about technology and partly about permission. A bank can only place AI in a live service when its technical architecture and its regulatory route both hold.

KakaoBank’s services bring AI into the bank’s daily relationship with millions of customers. They are an early service layer, rather than evidence of a rebuilt core-banking system.

Google: models, cloud and enterprise infrastructure

Google brings models, cloud infrastructure and a large enterprise platform to the same conversation.

HSBC announced a multi-year partnership with Google Cloud on 17 June, covering Gemini models, Google’s enterprise agent platform and engineering support from Google DeepMind. The bank says it expects more than 200 AI use cases over two years, beginning with wealth management, financial crime and client-facing work. In July, it said its global AI centre of excellence would open in Singapore in the second half of 2026.

HSBC expects each high-value initiative to generate more than USD 100 million in revenue or efficiency gains. While those are the bank’s estimates, they indicate where HSBC believes the returns may lie: work repeated at scale and closely connected to revenue, risk or client service.

Alibaba, DeepSeek and regional models: more choice for Asian buyers

Alibaba, DeepSeek and other regional models offer buyers a fourth route: lower-cost or open-weight models connected to Asian and Chinese technology ecosystems.

Alibaba held its first international Qwen conference in Singapore in May, presenting Qwen 3.7-Max and an agentic product stack to regional developers and businesses. Open-weight models can give an organisation more flexibility over where a model runs and how it is adapted. They can also reduce dependence on a single supplier.

OCBC provided a useful picture of that approach in 2025. Donald MacDonald, then in the bank’s data office, wrote that OCBC maintained a stable of around 10 foundation models across text, coding, speech and visual language. He named Llama, Gemma, Qwen-coder, Qwen-VL and DeepSeek R1. A separate guardrails service allowed the bank to evaluate and change models without disrupting applications built on top of them.

In 2025, MacDonald also said that OCBC had more than 30 generative-AI applications in production, running on its own GPUs. He described a private-banking onboarding process that had moved from 10 days to under an hour. That account shows why open-weight models appeal to some enterprise buyers: they can be treated as components within a wider operating system.

Meanwhile, Manulife Hong Kong’s June agreement with Alibaba Cloud shows another move. The companies plan to work on fraud detection, personalisation and a joint AI hub.

Across these paths, the technology choice is tied to an operating choice. Private deployment can offer more control at the cost of more internal capability. A hyperscaler can speed integration for firms already using its software. Open-weight models can provide flexibility, while requiring the expertise to test, secure and maintain them. The path a company takes will influence both the cost of adoption and its ability to turn AI into productivity.

Productivity is where the economics begin

The significance of enterprise AI lies in what it does to the productivity of knowledge work.

A financial analyst who can interrogate a large document set faster, a lawyer who can find and compare clauses more quickly, a developer who can test code in less time, or a customer-service team that can resolve routine cases sooner may each produce more with the same hours. At the scale of Asia’s large workforces, those changes could affect company margins, the price and quality of services, hiring patterns and public-sector capacity.

That does not mean every productivity gain turns into the same economic result. Companies may retain it through higher margins. Workers may gain through better tools, higher output or more valuable roles. Consumers may see lower prices or faster, more tailored services. Governments may use it to process applications, deliver assistance or manage public resources more efficiently.

The division of that gain will be shaped by competition, labour markets, regulation and the way an organisation chooses to redesign work.

Singapore’s Ministry of Manpower provides an early measure of that process. Among firms using AI, 70.7 percent reported improved worker productivity. Only 6.2 percent reported lower headcount. More firms reported redesigning job functions, at 18.9 percent, and 13.9 percent said they had created AI-related jobs.

Those figures suggest that the first change is often inside the job. Teams reorganise work around a new tool. Some routine tasks move more quickly; human attention moves toward judgment, client relationships, exceptions and oversight. The long-term outcome for workers will depend on whether organisations use the extra capacity to improve services and build skills, or simply to reduce the number of entry-level tasks and roles.

Singapore also shows how far there is to go. In MOM’s Artificial Intelligence Survey 2026, for which fieldwork ran from January to March 2026, most private-sector firms with at least 10 employees had not started using AI. Only 3.8 percent were integrating it into core processes. MOM found that implementation costs and a lack of in-house expertise were the most common constraints, while larger firms also cited integration complexity and data-security concerns. That gap could become an economic dividing line.

From return on investment to economic contribution

While the consumer AI boom is highly visible, enterprise AI moves more quietly, through procurement decisions, software integration, security reviews, workflow redesign and staff training. However, its consequences may be more significant because it can affect investment, infrastructure, productivity, trade, skills and industrial competitiveness at the same time.

For businesses, the first test is return on investment. Did a service team resolve cases faster? Did an analyst produce better work with the same time? Did a retailer reduce waste, or did an insurer identify risk earlier? These are useful measures, but they capture only the first layer of value.

The larger return comes when a company builds new ways of operating around the technology. That requires structured programmes rather than scattered experiments: clear priorities, data that can be used safely, redesigned workflows, people who know when to rely on the system and when to override it, and leaders willing to keep investing after the first pilot.

There is a useful parallel with the spread of electricity through industry. Early factories could replace a steam engine with an electric motor and gain some efficiency. The larger gains followed when they redesigned the factory itself: production lines changed, equipment moved, jobs were reorganised and workers learned different skills. Electricity became economically powerful because businesses rebuilt their operations around it.

Enterprise AI will demand a similar period of adaptation. A model added to an existing process can save time. A company that rethinks how information moves, how decisions are made and where people add judgment can create new products, stronger services and more resilient operations. That takes planning, capital and patience.

Over time, a mature view of AI return will reach beyond short-term cost savings. It will include whether an organisation can innovate faster, absorb disruption, develop more capable teams and create better value for customers. When those gains spread through large employers and public institutions, they become an economic contribution: higher-quality services, more competitive industries, new skills and, potentially, broader productivity growth across APAC.

What to watch

• Enterprise-AI investment in APAC. Watch for spending that moves beyond pilots and generic software licences into integration, data preparation, security, training and the systems needed to run AI at scale.

• How major companies mix their AI structures. As models improve and prices change, firms will decide how much work belongs in a private deployment, an existing Microsoft or Google environment, or an open-weight model. Those decisions will reveal what they value most: cost, capability, control or speed.

• How companies measure effectiveness. Headline user numbers are a starting point. Stronger evidence will show whether a workflow became faster, more accurate, less expensive or more useful to customers and staff — and who receives the gain.

• Regulation that affects deployment. South Korea’s cloud changes, Singapore’s agentic-AI guidance and sector-specific rules across the region will shape which services move from pilots into everyday use.

---