Agentic AI adoption is real, but production deployment is not keeping pace with boardroom ambition. Most organisations can point to an agent pilot; few can point to one operating with meaningful autonomy inside a core business process.
The gap is not primarily a model capability problem: The reasoning, planning, and tool-use abilities behind today’s agents are already sufficient for a wide range of enterprise work. The real challenge is that enterprises are moving from deploying individual agents to operating systems of agents, and the architectural foundations for that transition are still immature.
The pattern is familiar: A team builds a pilot, it delivers promising results, stakeholders identify adjacent opportunities, then progress slows. Organisations that deploy a handful of agents successfully struggle to expand across business functions because production introduces coordination requirements, governance obligations, and architectural dependencies that simply do not surface at pilot scale.
Three constraints are emerging repeatedly: coordination, control, and enterprise design.
The first scaling bottleneck - Coordination
Most successful pilots involve a small number of agents inside a narrowly defined task. Production is different: Agents must interact with multiple systems, run longer processes, and increasingly collaborate with other agents and humans. The question is no longer whether one agent can complete a task but whether a system of agents can operate reliably over time and across shifting context.
Traditional orchestration assumed predictable execution: A workflow engine invoked a service and triggered the next step in a sequence defined in advance. Agentic systems break that assumption, since an agent decides at runtime which tool to use or whether to hand off to another agent.
This is why orchestration is not a single problem but three: Execution orchestration keeps long-running work operating through infrastructure failures; cognitive orchestration coordinates nondeterministic activities such as planning and tool selection; and process orchestration coordinates agents within workflows that contain human approvals, compliance requirements, and audit obligations.
Most pilots succeed because they only require a light form of cognitive orchestration, usually contained within a single agent framework. Production introduces the other two layers: Aloan application or insurance claim may involve multiple agents, multiple systems, and regulatory controls over extended periods, and reasoning capability alone is not enough. The issue is rarely that agents cannot perform the work. It is that the coordination mechanisms needed to integrate them into real business operations remain immature.
The second scaling bottleneck - Control
As coordination complexity increases, control complexity follows close behind. An enterprise running a handful of agents can often govern them through manual oversight and platform-native controls. One running dozens or hundreds across business units faces a harder challenge: understanding what those agents are doing, what they can access, and whether their actions stay aligned with organisational intent.
The central risk is that agents are becoming more connected, and as this happens, governance shifts from a model oversight problem to a systems oversight problem. The platforms used to build and orchestrate agents enforce their own policies and telemetry but only within their own boundary, and deployments rarely stay confined to one platform.
The result is agent sprawl: Different teams deploy agents on different tools, each governed by whatever that platform offers, leaving visibility fragmented, policy enforcement inconsistent, and audit trails incomplete.
This is not unique to agentic AI. Similar patterns emerged as cloud computing expanded across providers, with organisations eventually recognising that governance needed to exist independently of the platforms being governed. The challenge is sharper in Asia Pacific specifically, where the regulatory backdrop is fragmenting rather than converging.
Singapore’s Model AI Governance Framework for Agentic AI, the first of its kind globally, sits alongside Indonesia’s data localisation requirements and Vietnam’s AI law (which took effect in March 2026), three materially different regimes inside one region.
A control architecture that cannot adapt to local policy without being rebuilt for each market will not hold up against that backdrop. The challenge is no longer governing an individual agent. It is governing a distributed system of autonomous actors across an equally distributed, and unevenly regulated, enterprise.
The third scaling bottleneck - Enterprise design
The coordination and governance challenges expose a deeper issue: Most enterprises are still organising their agent initiatives in ways that do not scale. The organisational model often remains rooted in traditional project thinking, where each new use case becomes a separate implementation effort. That works during experimentation, but over time, organisations accumulate isolated agents, duplicated capabilities, and inconsistent governance, an estate of point solutions that becomes harder to coordinate.
Use cases are useful for securing funding, but they are a poor foundation for an enterprise-scale agent ecosystem, since every new one encourages teams to rebuild capabilities that already exist elsewhere.
A more durable approach is to organise around reusable cognitive skills: capabilities such as document extraction, classification, or risk assessment, defined independently of the specific agent or workflow that consumes them.
Agents become compositions of skills, while business processes become orchestrations of agents and humans working together. This creates a more stable unit for governance, since the same skill can be reused across workflows under consistent controls.
Given the jurisdictional fragmentation described above, this matters in practice, not just in principle: A centrally governed skill can be adapted to local regulatory policy, whether Singapore’s, Indonesia’s, or Vietnam’s, at deployment, without separate implementations per jurisdiction.
What is actually working now, and what does that tell us?
The organisations generating measurable value today are not necessarily those deploying the most agents. They are the ones deploying agents where the surrounding architecture is mature enough to support them: customer service triage, internal IT operations, and structured back-office processes, where scope is well defined and governance is easier to enforce.
This is not the ceiling of what agentic AI can do. It reflects current enterprise readiness: Models are increasingly capable of more complex work, but the orchestration and governance needed to deploy those capabilities safely have not matured at the same pace. Whether agentic AI works is increasingly settled. What remains unsettled is whether the systems around it have matured enough to support it everywhere else.
The shift leaders need to make
The first phase of the agentic AI market asked whether agents could perform meaningful work. Increasingly, yes. The next phase asks a different question: Can enterprises operate systems of agents reliably at scale?
- Leslie Joseph, principal analyst, Forrester.
That has less to do with model capability than with architecture — whether organisations can coordinate autonomous actors across complex processes, govern them consistently across platforms, and organise their capabilities to support reuse rather than fragmentation.
This is why measuring progress by the number of deployed agents is increasingly misleading. An enterprise running 100 isolated agents may be less mature than one running 10 inside a well-governed, orchestrated system. The relevant measure is whether agents operate as part of a coordinated, controllable architecture, not how many exist.
The opportunity remains significant, particularly for financial services, telecommunications,and government, where automation can create substantial value and many are already experimenting aggressively. The organisations that pull ahead are unlikely to be those that simply deploy the most agents.
They will be those that build the orchestration and governance foundations required to scale them — and in markets where the regulatory map keeps shifting, those foundations need to be portable across jurisdictions from the outset.
That gap will not close because the next generation of models becomes marginally more capable. It will close when enterprises become capable of operating agentic systems with the same discipline, visibility, and control they expect from every other critical technology platform.
Leslie Joseph is the principal analyst at Forrester. This byline has been written exclusively for iTnews Asia.
