70% of Enterprises Run Three or More AI Models. Most Can’t Fully See What They’re Doing

Datadog’s Yadi Narayana says organizations across Asia-Pacific are scaling AI systems rapidly, but many still lack the visibility, governance, and operational maturity needed to manage increasingly complex AI environments.

Enterprise AI adoption across Asia-Pacific is entering a new phase. Over the past two years, organizations across the region have moved aggressively from experimentation to deployment, integrating large language models, AI copilots, automation systems, and agent-based workflows into customer service, software development, fraud detection, internal operations, and knowledge management.

However, as enterprises move AI systems into production, a new challenge is emerging: operational complexity. According to Datadog‘s State of AI Engineering 2026 report — based on real-world data from thousands of organizations running AI in production — nearly seven in ten companies (69%) now use three or more AI models in production environments, reflecting a growing shift toward multi-model strategies rather than reliance on a single provider. The report also found that approximately 5% of AI model requests fail in production, with nearly 60% of those failures linked to exceeded rate limits, creating performance bottlenecks and reliability concerns as AI deployments scale.

The trend reflects a broader transformation in enterprise AI. Organizations are no longer simply testing models. They are building increasingly complex AI ecosystems involving multiple vendors, agent-based architectures, cloud environments, and regional compliance requirements. As a result, observability is becoming a growing priority for enterprises seeking visibility into how AI systems perform, where failures occur, how costs are accumulating, and whether deployments remain reliable at scale.

Speaking exclusively with AsiaTechDaily, Yadi Narayana, Field CTO for Asia Pacific and Japan at Datadog, said many organizations across the region are encountering operational challenges as AI adoption accelerates.

AI Adoption Is Scaling Faster Than Operational Readiness

While enterprise AI adoption is increasing globally, Narayana noted that maturity levels across Asia-Pacific remain uneven.

“Across Asia-Pacific we’re seeing the same structural pressures that show up in our global data, but the maturity profiles across markets differ,” he told AsiaTechDaily.

According to Narayana, ASEAN markets provide a particularly clear example of this divergence.

“Singapore is further along on governance and observability, while Malaysia, Indonesia and Thailand are pushing deployments into production faster than their operational readiness.”

The observation highlights a growing distinction between AI adoption and AI operational maturity. Across Southeast Asia, enterprises are increasingly deploying AI into real-world business environments. However, many organizations are still developing the governance frameworks, monitoring capabilities, and operational processes required to manage these systems effectively once they move beyond pilot programs.

This challenge becomes more significant as AI deployments expand across multiple business functions and jurisdictions. Unlike earlier software deployments, AI systems often operate dynamically, interact with external models, consume large volumes of data, and generate outcomes that can be difficult to trace or explain without adequate monitoring infrastructure.

The Rise of Multi-Model and Agent-Based Architectures

One of the most significant shifts occurring within enterprise AI is the move toward multi-model environments. Rather than relying on a single AI provider, organizations increasingly use different models for different workloads, balancing performance, latency, reasoning capability, and cost considerations. Datadog’s report found that more than 70% of organizations now use three or more models, and the share of companies operating six or more models has nearly doubled. Agent framework adoption — tools such as LangChain, LangGraph, and Pydantic AI — also doubled year-over-year. At the same time, enterprises are experimenting with agent-based architectures capable of carrying out multi-step tasks with limited human intervention. These developments are creating new operational challenges.

“As multi-model and agent-based architectures emerge, teams across the region are running into reliability issues, limited visibility and inconsistent performance, as well as rising token costs without robust optimisation disciplines,” Narayana said.

The complexity stems from the fact that organizations are no longer monitoring a single application stack. Instead, they must manage interconnected AI systems that may involve multiple models, APIs, orchestration layers, data pipelines, and autonomous agents operating simultaneously. As AI systems become more distributed, maintaining visibility into how these environments behave becomes increasingly difficult.

Why Visibility Is Becoming a Business Requirement

One of the most pressing challenges emerging from enterprise AI deployments is the lack of operational visibility. Traditional monitoring tools were largely designed for infrastructure, applications, and cloud environments. AI introduces a different set of questions. Organizations increasingly need to understand which model handled a request, why latency increased during a workflow, where token consumption is rising, how agent decisions are affecting outcomes, and whether AI systems are meeting governance requirements. According to Narayana, many enterprises still lack these capabilities.

“Many teams are moving from proofs of concept into customer service, software development, fraud detection, knowledge management and internal productivity use cases, but may not yet have a mature way to see which model was called, why latency spiked, where tokens are being wasted or how much each workflow costs.”

The issue is becoming increasingly important as AI spending grows. Unlike conventional software systems, AI workloads often generate variable costs depending on model usage, token consumption, inference demands, and orchestration complexity. Without visibility into those variables, enterprises can struggle to optimize performance or control spending. Industry analysts increasingly view AI observability as a critical layer for enterprise AI governance — organizations cannot effectively manage systems they cannot fully observe.

Observability Is Expanding Beyond Infrastructure Monitoring

Historically, observability focused on collecting and correlating logs, metrics, traces, and infrastructure telemetry. Today, the concept is expanding. Modern AI observability increasingly includes monitoring model behavior and output quality, hallucinations and prompt injection attempts, agent actions and decision paths, workflow reliability, cost and token utilization, and governance and compliance controls.

Datadog has significantly expanded its AI observability capabilities over the past year, including new tools focused on LLM monitoring, AI agent observability, workflow tracing, and AI system evaluations. The company’s broader strategy reflects growing enterprise demand for visibility into AI behavior rather than simply infrastructure performance.

The shift is also being reflected across the wider technology industry. Recent research and enterprise deployments suggest that monitoring AI systems increasingly requires visibility into reasoning chains, model interactions, agent actions, and workflow execution — rather than traditional application telemetry alone. As agentic AI systems become more common, observability is evolving into a foundational layer for governance, accountability, and operational trust.

Stabilization Is Becoming the Next Enterprise AI Challenge

For many organizations across Asia-Pacific, the immediate goal is no longer simply deploying AI. The next challenge is ensuring those deployments remain reliable, governable, and economically sustainable as they scale. This is particularly relevant in Asia-Pacific, where enterprises often operate across multiple jurisdictions with varying regulatory requirements, infrastructure environments, and data governance frameworks.

“That creates a practical scaling challenge because the region’s speed of adoption can expose operational gaps faster,” Narayana said.

He noted that enterprises increasingly face challenges around data governance, compliance, infrastructure management, and cost control as AI systems expand across multiple markets.

“This becomes more important in multi-market environments, where data, compliance and infrastructure decisions often vary across jurisdictions.”

According to Narayana, the priority for many organizations is now stabilization.

“As such, the immediate challenge for many enterprises in this region is stabilisation: shoring up observability, governance and cost control so AI can scale without undermining reliability or margins.”

The Next Phase of Enterprise AI in Asia-Pacific

The first wave of enterprise AI adoption focused largely on experimentation and deployment. The next phase is likely to focus on operational maturity. As organizations expand AI usage across business functions, the complexity of managing multi-model systems, autonomous agents, governance requirements, and infrastructure costs is increasing rapidly. Visibility into these environments is becoming less of a technical preference and more of an operational necessity.

For enterprises across Asia-Pacific, success may increasingly depend not on how quickly AI systems are deployed, but on how effectively those systems can be monitored, governed, optimized, and trusted at scale.

The region’s AI adoption momentum remains strong. However, as deployments move deeper into production environments, observability is emerging as one of the key foundations required to support the next stage of enterprise AI growth. Datadog’s findings suggest that the challenge facing many organizations is no longer whether they can implement AI, but whether they can operate increasingly complex AI systems with the visibility and control needed to sustain long-term performance, reliability, and business value.

Quick Takeaways

• Enterprise AI adoption across Asia-Pacific is moving rapidly from experimentation to production, creating new operational challenges around reliability, governance, and cost management.
• According to Datadog, organizations are increasingly deploying multiple AI models simultaneously, making AI environments more complex to monitor and manage.
• Datadog’s Yadi Narayana says AI adoption is scaling unevenly across ASEAN, with markets such as Malaysia, Indonesia, and Thailand pushing deployments into production faster than their operational readiness.
• Multi-model and agent-based AI architectures are introducing new challenges, including limited visibility, inconsistent performance, rising token costs, and workflow reliability issues.
• Many enterprises still lack mature mechanisms to understand which models are being used, why latency spikes occur, where tokens are being wasted, and how much individual AI workflows cost.
• AI observability is evolving beyond traditional infrastructure monitoring to include model performance, agent behavior, workflow tracing, governance, and cost optimization.
• As AI deployments expand across multiple markets and jurisdictions, observability is becoming increasingly important for managing compliance, infrastructure complexity, and operational risk.
• For many Asia-Pacific enterprises, the next phase of AI adoption is focused on stabilization, ensuring AI systems can scale reliably without compromising performance, governance, or margins.