As enterprises across Asia Pacific push beyond traditional Retrieval-Augmented Generation, Siddon Tang, GM of Asia Pacific, SVP of Engineering and Product, TiDB, says the success of Agentic RAG may depend less on AI models and more on whether organisations modernise the data foundations that support autonomous intelligence.
Retrieval-Augmented Generation has quickly become the foundation of many enterprise AI projects across APAC. It helps large language models speak the language of the business by grounding responses in internal knowledge. For uses like chatbots, knowledge lookup and customer support, it has been good enough.
But the region is now pushing into something more ambitious. Instead of simply retrieving answers, organisations want AI that can work across systems, understand context, make decisions and take the next step without waiting for a human. This is where Agentic RAG enters, enabling AI agents that can plan, reason and coordinate tasks on their own.
The promise is real. However, the reality is that many enterprises discover the same barrier. Agentic systems struggle not because the AI is immature, but because the underlying data layer was never designed for this kind of intelligence.
Why APAC is ready for Agentic RAG
APAC has the perfect conditions to embrace autonomous AI. It is a region marked by rapid digital growth and massive data generation across banking, retail, logistics and manufacturing. Businesses operate across diverse languages, regulatory environments and markets, making automation even more valuable.
Many organisations also have the advantage of leapfrogging. Instead of decades of technical debt, there is an opportunity to build more modern architectures from the start. This gives APAC a chance to become a global reference for practical, scalable AI.
Yet all of this potential runs into the same hurdle: legacy data systems that cannot support the speed, scale and intelligence Agentic RAG depends on.
How Agentic RAG stretches legacy architecture
Agentic AI behaves very differently from traditional software. When enterprises plug early prototypes into existing systems, several limitations surface almost immediately.
First, unpredictable and high-volume workloads become a challenge. AI agents run many tasks simultaneously. They fetch context, generate temporary data, coordinate with other agents and reason in loops. This creates sudden workload spikes. Legacy environments, which typically scale vertically or depend on manual tuning, struggle to keep up.
Second, multiple workload types happen at once. Agentic RAG does not separate workloads neatly. An agent may read operational data, analyse patterns, fetch documents, process embeddings and write results back into a system all within seconds. Traditional divisions between transactional systems, data warehouses and search engines create delays that undermine the quality of reasoning.
Third, context becomes fragmented and stale. Autonomous agents need a unified picture of what is happening right now. When operational databases, analytics platforms and document stores are disconnected, context becomes inconsistent. Even small lags or mismatches can lead to incorrect decisions.
Finally, there is limited room for experimentation. Agents evolve quickly. Teams need to try new reasoning methods, adjust knowledge sources and test different workflows safely. Many legacy databases are rigid, making it difficult to create isolated environments or roll back changes. This slows down iteration and increases risk.
These issues are not theoretical. They appear the moment organisations attempt to scale beyond small RAG pilots into real-world agentic automation.
What a modern data architecture should look like
Supporting Agentic RAG requires a shift in how enterprises think about data infrastructure. It is no longer enough to bolt on additional search tools or optimise a single database. The foundation itself needs to reflect how AI systems operate.
Hybrid, multi-modal retrieval is essential. Enterprises work with tables, PDFs, logs, emails, embeddings and more. An agent needs to search semantically, match keywords and filter by business rules in one place. Hybrid retrieval becomes critical for grounding decisions accurately.
Real-time understanding is equally important. Agents must act on the present, not yesterday’s data. Architectures that unify operational and analytical processing are becoming necessary, simply because autonomy depends on fresh information.
Elasticity also matters. AI workloads spike, flatten and spike again. Scaling needs to feel automatic. Systems should grow to support reasoning bursts and shrink during idle periods to keep costs under control.
Finally, open and composable design is key. Agentic AI sits at the centre of a larger ecosystem. It interacts with orchestration frameworks, event streams, vector indexes and monitoring tools. An open architecture allows enterprises to adopt new capabilities quickly and avoid being boxed into rigid stacks.
A more intelligent future depends on the foundations we build today
The move from RAG to Agentic RAG signals an important shift. AI is no longer just retrieving information. It is beginning to participate in the decision-making fabric of the enterprise. Whether this evolution reaches its full potential will depend less on new models and more on whether organisations prepare the data layer to support intelligence that is always observing, always reasoning and always learning.
Enterprises that modernise now will not just adopt Agentic RAG. They will shape how it transforms the way work gets done.
