Across Southeast Asia, generative AI pilots are stalling—not from a lack of model power, but from broken retrieval. Agentic RAG bridges this gap: autonomous agents that verify facts, enforce governance, and execute end-to-end workflows. For CIOs in 2026, this turns fragile experiments into auditable, scalable profit centres.
With Gartner warning that 60% of AI projects will be abandoned due to poor data and weak controls, agentic RAG is no longer optional—it is the only practical path from pilot to production. In markets like Singapore, where data residency and compliance are non-negotiable, retrieval intelligence is now the bedrock of ROI.
The pilot trap: Why AI experiments stall
The pattern is painfully consistent across the region. Budgets get approved, pilots get built, demos get applauded—and then nothing ships. According to Gartner, by the end of 2025, at least 50% of generative AI projects were abandoned after proof of concept due to poor data quality, inadequate risk controls, escalating costs, or unclear business value. S&P Global’s 2025 Voice of the Enterprise survey found that 42% of companies abandoned most AI initiatives last year, more than double the 17% recorded the year before.
Ed Keisling,chief ai officer at Progress Software, puts it bluntly: “The devil is really in the details when it comes to scale.” He explains that while RAG is easy to understand in concept, it is “very hard to do at scale“—and this is where most implementations fail.
The problem, industry analysts argue, is rarely the model itself. Rather, it is the complexity of enterprise data landscapes: legacy systems, multilingual content, and disparate cloud storage. For organisations in Southeast Asia, this challenge is magnified by multiple languages, different embedding strategies, and tightening data residency laws.
From broken retrieval to agentic RAG
Traditional RAG systems are starting to show their limits, particularly when enterprises try to combine structured and unstructured data while ensuring governance and output quality. To address these limitations, agentic RAG is emerging as the critical evolution.
Agentic RAG builds on traditional pipelines by adding reasoning capabilities and introducing greater flexibility in how information is retrieved and reasoned over. The goal is to improve relevance, accuracy, and alignment with user intent. Critically, this approach standardises the pipeline and raises confidence at each step.
As Keisling observes: “One of the core challenges with any agentic system is the potential amplification of wrong answers. If you do a retrieval and get 80% accuracy, over a multi-turn agentic workflow, you may end up with 80% of 80% of 80%.” Without the right checks and balances, the result becomes a coin flip in terms of correctness.
This is where the “agentic” layer proves indispensable. By embedding verification, governance, and reasoning at each stage, agentic RAG turns fragile experiments into systems that can be trusted to operate at scale.
Governance and data sovereignty: The APAC imperative
For enterprises across Southeast Asia, data residency and compliance are non-negotiable. Governments in the region have introduced measures that keep personal and strategic information within national borders, pushing many organisations towards sovereign AI infrastructure. Vietnam’s Decree 53, for example, requires data relating to Vietnamese citizens to remain in-country, making sovereign AI infrastructure essential.
Agentic RAG solutions address these concerns by giving organisations control over data residency and data access. Keisling explains that such solutions provide “choice over data residency, whether you want to run in your own environment or in a regional data centre”—as well as flexibility in which models are used for ingestion or extraction based on privacy requirements.
This control extends to auditability. Jason Mantell, APAC director of solutions architects at Cloudian, notes that RAG will be “especially critical for government and financial services, where data volumes are immense but tuned to be shielded from the open internet.” It effectively operates as “a mature search engine for enterprise data, moving beyond simple keyword matching to true semantic understanding”.
Measuring ROI: Groundedness, relevancy, and trust
When measuring the return on retrieval intelligence, Keisling advises CIOs to focus on metrics such as groundedness—whether answers are based on the retrieved data—and relevancy—whether answers match the questions being asked. Monitoring these metrics over time is essential.
“You want to be able to verify the data being returned, particularly as you incorporate it into day-to-day workflows,” he says. The key with any RAG system is being able to ground answers in verifiable facts and trace which answers map to which data in the repository.
Progress Software’s approach includes quality metrics for every generated answer. Capabilities like REMi (RAG Evaluation Metrics), which use an LLM to evaluate other LLMs’ outputs, are critical to building trust and ensuring consistent, high-quality results. Without such evaluability, organisations cannot confidently deploy AI in production environments.
The path to profit: Intentionality in ingestion
Keisling emphasises that the secret to successful AI deployment lies in intentionality—particularly in how data is ingested and prepared. “You need to, as a human, go through and identify the most important artefacts from a knowledge perspective, then be very deliberate about labelling those documents, with dates, with the specific fields you care about, so that retrieval is much more likely to return the right answer.”
This means addressing core data challenges: how documents are chunked, indexed, and embedded. As Keisling notes, language models are probability machines—if they cannot find an answer, they will generate one, often with high confidence. Agentic RAG mitigates this by enabling the construction of context from specific information sources before passing it to the LLM.
For organisations wrestling with legacy systems and fragmented data, this represents a fundamental shift. IDC predicts that by 2026, two-thirds of the top 1000 businesses in Asia will leverage GenAI and RAG to improve decision-making by 40% through domain-specific knowledge discovery. This is not merely a technology upgrade; it is a business transformation.
Build or buy: A regional reality check
For regional enterprises without custom AI stacks, the build-vs-buy decision looms large. Keisling advises that all major vendors—AWS, Microsoft, and Google—now provide Agentic RAG capabilities with data residency controls. However, the complexity of different document types, languages, and embedding strategies can quickly become overwhelming.
“If building the solution will differentiate your business, direct your engineers towards it. If it’s secondary to your core business and you want to leverage it to go and build something else, I’d suggest buying it from a vendor.”
This pragmatic approach reflects the reality that most enterprises are not AI infrastructure specialists—and should not have to become them. The average enterprise RAG implementation takes 8–12 months and costs between $500,000 and $2 million before delivering any business value, with many projects abandoned before reaching production.
The leadership imperative
Looking ahead over the next 12 to 18 months, Keisling identifies a critical leadership challenge: “Large language models raise the floor for everyone. We can all see what’s possible. But if you raise the floor for everyone, you haven’t really differentiated yourself from competitors.”
The way to truly differentiate, he argues, is to use an organisation’s distinctive competency—its “secret sauce”—to drive the top line better with AI. This means ingesting all available knowledge and leveraging it safely and securely.
“There’s a lot of AI hype out there, and I think a lot of the failed projects come from rushing in to solve problems with AI, not necessarily picking the right problems, or trying to apply AI to problems that don’t need it.” Ed Keisling
For CIOs across Asia-Pacific, the message is clear. The window for experimentation is closing. As organisations in the region pivot from pilot projects to large-scale, sovereign AI infrastructure, those investing in modern data platforms capable of handling frontier-model demands will be best positioned to turn AI from aspiration into operational value and lasting competitive advantage.
Click on the PodChats player to learn more about turning APAC’s AI pilots into profits in 2026
- What is RAG?
- Given that most regional AI pilots never scale, what specific architectural weaknesses does agentic RAG fix that traditional RAG or fine-tuning cannot?
- In markets with fragmented data landscapes—legacy systems, multilingual content, and disparate cloud storage—how does agentic RAG ensure consistent, high-quality retrieval at enterprise scale?
- What out-of-the-box governance and audit trails does agentic RAG provide to satisfy both local data residency laws (e.g., Singapore’s PDPA) and board-level risk controls?
- For CIOs managing lean teams, how does agentic RAG reduce the operational burden of maintaining retrieval pipelines, monitoring hallucinations, and orchestrating multi-step agent workflows?
- How can agentic RAG help move beyond isolated use cases (e.g., customer support) toward fully autonomous, end-to-end processes spanning finance, supply chain, and compliance?
- How should CIOs in Singapore and across Southeast Asia measure the ROI of retrieval intelligence compared to simply upgrading large language models?
- For regional enterprises without custom AI stacks, what vendor or open-source scaffolding for agentic RAG offers the fastest path from pilot to profit while preserving data sovereignty?
- What organisational, data, and leadership shifts must CIOs prioritise over the next 12–18 months to ensure agentic RAG transitions from a technical capability into a sustained source of competitive advantage?
