For Southeast Asian CIOs, the window to experiment with AI is closing fast. The imperative now is to industrialise intelligence, moving from proof of concept to production-scale agentic systems that deliver tangible business value.
Yet, as the “token economy” dictates, success hinges on a foundational element: your data platform. The key to controlling costs and ensuring trusted, real-time outcomes lies not in the fragmented data stack many organisations use today, but in a unified architecture that makes retrieval accuracy a competitive advantage. This demands a new strategic focus.
The AI readiness divide is a warning for ASEAN CIOs
The stakes have never been higher. IDC research commissioned by MongoDB reveals a stark “AI readiness divide” across Asia/Pacific, separating a “Leaders Cohort” from the “Mainstream Cohort.” The leaders, who have embedded modernisation into their strategy, generate nearly three times more digital revenue than their peers.
This is not a theoretical gap; it is measurable in business performance.
For CIOs in the region, the urgency is clear. IDC research shows that 95% of organisations have reported project delays, and 90% have experienced failed modernisation initiatives, with poor data quality as a consistent culprit.
Sources: IDC Asia/Pacific Big Data and Analytics Software Forecast, 2025–2028; IDC FutureScape: Worldwide CIO Agenda 2026 Predictions Doc # US53865325; IDC’s Asia/Pacific Modernization Survey, 2025, sponsored by MongoDB, n = 1,400Dr William Lee
As IDC’s Dr William Lee states, “Research shows that many organisations are being held back by their existing rigid legacy architectures that do not have the flexibility and scalability to handle the high volume of unstructured data required for AI”. As a result, 43% of organisations say they cannot build new applications without major modernisation first.
The fragmented stack is an architecture designed for failure
Thorsten Walther, managing director of CXO Advisory Asia at MongoDB, paints a vivid picture of the problem: “The biggest difference between a pilot and a production is that production has to work with live business data, not curated data sets.”
He observes a recurring pattern: “The pilots don’t fail because the model isn’t smart enough. They fail because the right facts aren’t in front of the model at the right moment.”
Walther calls the root cause “exactly the wrong setup.” Many enterprises have built a fragmented data stack, stitching together separate databases, vector stores, search engines, and governance tools.
“Every extra piece adds latency, adds synchronisation problems, risk and complexity,” he warns. This fragmentation is the primary obstacle to achieving the retrieval accuracy agentic AI needs to succeed.
The token economy: Retrieval accuracy as cost control
AI’s economic model reinforces this imperative. Walther explains that the “token economy” makes retrieval accuracy a direct cost-control lever. He highlights the cost of failure:
Thorsten Walther
“The model gets incomplete or irrelevant information, so it gives a weak answer. Someone retries, adds another prompt, or a human steps in. Each loop costs money, and at enterprise scale, it adds up fast.” Thorsten Walther
With agentic AI, the problem escalates. “An agent doesn’t just answer; it acts on what it retrieves. If the context is wrong, the mistakes go straight into your business process,” Walther states. The most effective cost lever, he argues, “isn’t a cheaper model; it’s better retrieval filtering.”
The rise of agentic systems and “agent memory”
The transition from chatbots to agentic AI systems represents a fundamental shift. Walther notes: “A chatbot answers a question and nothing changes. An agent executes a transaction, and the state of your business has changed. That’s a completely different bar or architecture and a completely different complexity.”
This new paradigm demands what Walther calls “agent memory,” blending immediate context, operational knowledge, and insights learned across the agent fleet. He warns against architectures where agents act on “last night’s data” because they rely on synchronised copies.
“If the architecture grants agents direct access to the operational data, they go into the operational data with transactional integrity when they write and the full audit trail of what they did,” he advises.
This is critical, as auditability is non-negotiable in regulated environments.
The unified platform: A strategic solution
To overcome these challenges, Walther advocates for a strategic shift: “Bring operational data, search, and retrieval together on one platform. Fewer copies of your data, fewer moving parts, and you get lower latency and better governance and, most importantly, reduced complexity.”
This is not merely a technical preference but a strategic necessity. The Leaders Cohort identified in the IDC study treats AI readiness as an enterprise capability, not a standalone initiative. They invest in modern data platforms that support hybrid architectures and AI workloads without adding complexity.
As Walther points out, “AI enablement is now the number one reason organisations choose the database.”
Security, compliance, and data sovereignty
For regulated enterprises in ASEAN, security, compliance, and data sovereignty concerns further complicate the path to AI adoption. Walther addresses this head-on: “Regulated organisations shouldn’t have to choose between responsible AI innovation.”
He emphasises that “autonomy is earned, not assumed.” Before an agent goes into production, it needs memory, context, proper retrieval, transactional rights, and a full audit trail. “Every read, every write, every decision must be explainable.”
This becomes exceptionally difficult when data is copied across five different environments. As Singapore’s cybersecurity agency recently published additional guidance on securing agentic AI systems, the need for a governable architecture is front of mind.
A call to strategic action
For CIOs in Southeast Asia, the message is clear: the era of AI experimentation is over. The move to production-grade, agentic systems requires a fundamental rethink of the data platform. The fragmented data stack of the past blocks innovation and creates risk.
The strategic imperative is to unify operational data, search, and retrieval on a single, flexible platform. This is the foundation for accurate retrieval, cost control, and the secure, auditable deployment of agentic AI.
As Walther concludes, “every company now is an AI company, whether they plan for it or not.” CIOs must choose whether to lead this transformation or be left behind.
Click on the PodChats player for more on why retrieval accuracy is king and use cases to make that happen.
As we move from AI experimentation to production, how can we rationalise a fragmented data and retrieval stack to reduce latency and governance risk, particularly given the need to manage highly dynamic, unstructured data?
Given that the “token economy” makes retrieval accuracy a direct cost-control lever, how can we implement a unified data platform to improve retrieval quality and reduce expensive LLM retry loops?
How do we ensure our data architecture provides the schema flexibility needed for rapid AI iteration, avoiding the rigidity of relational models that slows development and creates technical debt?
With the rise of agentic AI, how can we architect a system for “agent memory”—blending short- and long-term context—that allows agents to act on the current state of data, not stale copies?
How can we improve retrieval accuracy by natively combining semantic understanding with precise keyword search, while also using reranking to refine results, without adding external systems that create sync delays?
For regulated enterprises, how can we bring these production-grade AI retrieval capabilities inside our compliance framework, without forcing a choice between innovation and data sovereignty?
What is our strategy to move beyond the complexity of managing separate databases, search engines, and vector stores to a single, unified platform that reduces operational overhead?
How do we build a flexible, “production-ready” data foundation that lets us pivot quickly as model providers and agent frameworks evolve, without locking us into a rigid stack?
How can we best equip our developers and agentic workflows with the skills and best practices needed to avoid common pitfalls like over-normalisation, ensuring agents build on a robust data model?
With a significant focus on the ASEAN market, how can we leverage local partnerships and expertise to accelerate our AI modernisation journey and address specific regional regulatory and data challenges?
