Across Asia-Pacific, enterprises are discovering that building an impressive AI agent is considerably easier than making one work reliably inside a complex business. From Singapore’s financial services sector to Japan’s manufacturing industry, organisations are experimenting with autonomous systems that can make decisions and take action across multiple applications.

However, the distance between experimentation and enterprise-wide deployment remains substantial.

Research from in 2025 found that just 13% of surveyed respondents in Japan and India reported AI agents integrated into workflows, while 64% were still experimenting, piloting or using agents under supervision.

Singapore illustrates this contradiction particularly clearly. ServiceNow’s 2026 found that agentic AI adoption had increased from 22% to 51% in a year. Yet only 10% of enterprises were using agents to operate autonomous, end-to-end workflows.

The technology is advancing. The more difficult question is whether enterprise infrastructure and management practices can keep pace.

Why Successful Pilots Struggle in Production

The fundamental problem is that controlled demonstrations rarely reflect the complexity of everyday business operations.

“Most pilots run on clean, curated data and a friendly process owner. Production has neither,” says , Founder and CTO of Glokal AI. “The pilot proves the model can reason; it doesn’t prove the agent can survive the systems around it.”

For enterprises managing decades of technology investment, this distinction is critical. An agent might successfully process information in isolation but encounter inconsistent records and undocumented procedures when connected to production systems.

2026 Pulse of Agentic AI research found that 60% of APAC executives identified technical challenges as the principal obstacle to advancing agentic AI, the highest proportion among the regions surveyed.

Applied AI Researcher and Executive Advisor, argues that the obstacles extend beyond technology. “The gap between a pilot and a production system is almost never a capability problem. It’s a governance problem,” he says.

Pilots operate within carefully defined boundaries. Production systems must handle unexpected circumstances, incomplete information and decisions that could affect customers, finances or operational continuity.

Japan provides a revealing example. October 2026 Enterprise AI Adoption Impact Index found that only 38% of Japanese respondents considered their data sufficiently mature and governed for reliable AI, compared with 74% globally.

These limitations become particularly significant in manufacturing environments where agents may interact with enterprise resource planning, production scheduling and supply chain applications.

Pattanaik recalls encountering integration problems between SAP and store systems during his work at ALDI. Different interpretations of business data and repeated transactions required a redesigned approach that saved more than €1 million.

“Agents walk straight into the same traps, only faster,” he commented.

The implication is clear: connecting autonomous systems to established infrastructure requires more than application programming interfaces. Enterprises must understand the meaning, reliability and operational consequences of the information agents consume.

Redesigning Workflows Around Business Outcomes

Once enterprises address integration challenges, they face another decision: should agents automate existing processes or fundamentally change how work gets done?

The distinction could determine whether agentic AI delivers incremental efficiency or meaningful business transformation.

Ayhaam Alzaaim, Senior Vice-President of Asia Pacific and Middle East at , argues that organisations should begin with processes that are understood, repeatable and measurable. “The best place to start is with tasks that are well understood, repeatable, measurable and governed,” he says.

Understanding the process provides the foundation for defining what an agent can do independently, where exceptions occur and when human intervention remains necessary.

For financial services organisations, this might involve automating transaction investigations while retaining human approval for consequential decisions. In manufacturing, agents could coordinate routine scheduling adjustments while escalating changes that threaten production continuity.

Maurizio Garavello, s SVP for APAC and EMEA, highlights fraud detection as an example of why operational context matters. “Customer behaviour and payment patterns differ across markets, so the same signals may not mean the same thing locally,” he says.

He points to Singapore bank OCBC’s decision to extend an AI-based financial-crime detection proof of concept with additional test data before implementation.

However, improving an existing workflow is not necessarily the same as transforming it.

Jaggi suggests examining whether a process derives its value from the sequence of activities or the outcome it produces. Where established sequences are essential for regulatory compliance, careful adaptation may be appropriate. Where processes reflect historical limitations, redesign could deliver greater benefits.

The commercial case must also withstand scrutiny.

and HFS Research found that IT operations and customer support were among the functions reporting measurable agentic AI outcomes in 2026, at 50% and 42% respectively.

Nevertheless, activity is not profitability. Enterprises need evidence that autonomous systems improve resolution quality, reduce operating costs and deliver sustainable results after implementation and oversight expenses. “Usage isn’t adoption and deflection isn’t resolution,” Pattanaik told Silicon UK. “If verification eats the saving, you don’t have a saving.”

Why One APAC Deployment Cannot Fit Every Market

Scaling across Asia-Pacific introduces complexities that enterprises cannot resolve through technology standardisation alone.

Singapore, Japan, Australia and Southeast Asian markets operate within different regulatory frameworks, business cultures and customer expectations. Even apparently identical processes may depend on different definitions of data, approval procedures and communication practices. “The biggest mistake enterprises make is assuming that what worked in one market will work in another,” says Jaggi.

While governance principles can be shared, local configurations require independent validation. Language creates another challenge. An agent that can translate customer communications accurately may still misunderstand cultural expectations around formality, hierarchy, or indirect responses.

Jaggi, drawing on his experience with neural machine translation, distinguishes between linguistic accuracy and effective communication. “Translation accuracy and communication accuracy are two different things,” he says.

For customer-facing deployments, these differences can influence trust and whether an automated interaction successfully resolves a problem. Technology entrepreneur Pavlo Kharmanskyi similarly argues that enterprises need local expertise to assess how agents communicate, rather than relying exclusively on translation capabilities.

There are also important architectural implications. Data sovereignty requirements may determine where information can be processed, while differences in enterprise systems can affect which actions agents can execute. Garavello argues that enterprises must validate performance in individual markets before expanding. “Scaling means validating an agent’s decisions and actions in each market – and that evidence matters more than a successful demo.”

Successful regional deployment, therefore, requires a balance between centralised governance and local operational flexibility. Enterprises should standardise security principles, monitoring and accountability while allowing workflows, language and business rules to reflect local requirements.

Governance Becomes the Real Test of Autonomy

As AI agents move from recommending actions to executing them, the consequences of inadequate oversight become more serious.

and Oxford Economics’ 2026 research found that 40% of surveyed Japanese organisations lacked agent permissions or access-management controls, while 36% had no human review process within agent workflows.

These findings raise difficult questions about accountability when autonomous systems operate across multiple applications. Jaggi identifies scoped permissions, immutable audit logs and visibility into agent communications as essential controls. Without these safeguards, enterprises risk granting autonomous systems access they cannot adequately monitor.

Pattanaik proposes another principle: the degree of autonomy should depend on whether an action can be reversed. Where consequences cannot be reliably undone, human approval should remain mandatory.

Governance also requires clear organisational ownership. “The technology team owns the infrastructure. The business unit owns the outcomes. But somebody has to own the agent itself,” Jaggi says.

Alzaaim believes enterprise leaders must understand processes before delegating authority. “The objective should not be maximum autonomy. It should be the right level of autonomy for the process, supported by the right context and controls,” he commented.

Performance measurement must evolve accordingly. Beyond time saved, enterprises should monitor error rates, human escalations, reversed decisions, repeat customer contacts and the costs of verification.

These measures help distinguish genuine operational improvements from automation that merely transfers work elsewhere.

For APAC enterprises, the next phase of agentic AI will be defined less by the sophistication of individual models than by the ability to operate them safely, consistently and profitably.

The organisations that succeed will connect technology investment to measurable business outcomes, redesign processes where necessary and establish governance that can accommodate regional differences.

The decisive question is no longer whether an AI agent can perform a task. It is whether the enterprise can trust that agent to perform it repeatedly, under real operating conditions, without compromising the business it was designed to improve.