Most enterprises are investing heavily in agentic AI, but many still struggle to deploy it at scale or to create sustained business value.
This refrain in the AI adoption narratives unfortunately seems to hold true over the last 12 months. Why? Because the true cost of deploying AI without the right architecture can be prohibitive.
Retrieval-augmented generation (RAG) improves on standalone LLMs by generating responses using sources outside its training data. This approach can dramatically reduce the occurrence of AI hallucinations.
But is agentic RAG the missing bridge between APAC’s AI experimentation and real ROI?
In this interview with Eudald Camprubí, Software Fellow, Progress Software, we discover that what separates experimental AI from production-ready is not ‘model intelligence’, but ‘retrieval intelligence’.
Gartner predicts that, in 2026, organizations will abandon 60% of AI projects that aren’t supported by high-quality, AI-ready data. How does an effective retrieval strategy help resolve this major pain point?
Eudald Camprubí (EC): Gartner highlights a simple but critical idea: AI is only as good as the data it uses. However, organizations often underestimate how important it is to prepare their data properly.
Making data AI-ready means more than just storing it. It involves organizing and enriching it so AI can understand it. This includes tasks like indexing different file types, extracting key information, identifying entities, generating summaries and even describing images or tables automatically.
This preparation is essential for effective retrieval, which is how AI finds the right information to answer a question. Different teams, such as marketing or legal, need different types of information, so retrieval must adapt to each use case.
By ensuring data is well-prepared and applying the right retrieval strategies, companies can make their AI projects more reliable and valuable in real-world use.
What is the role of agentic RAG in closing the loop between AI pilots and real workflows to make AI strategies more time- and cost-efficient?
EC: Agentic RAG provides essential capabilities to help enterprises confidently deploy AI solutions in production in a cost-efficient way and within weeks rather than months.
First, it offers quality metrics for every generated answer. In practice, no organization will deploy an AI solution without ensuring that outputs are accurate, grounded in their own data and free from hallucinations. Capabilities like REMi (RAG Evaluation Metrics), using an LLM evaluating other LLMs outputs, are critical to building trust in the system and ensuring consistent, high-quality results.
In addition, experimentation is key. Organizations need the ability to compare different models and retrieval strategies in controlled environments, understanding both output quality and cost (such as token consumption) without impacting production.
Equally important is providing intuitive tools that can be used not only by technical teams but also by business users. This allows all stakeholders to understand, fine-tune and control the AI pipeline, helping reduce costs, save time and ensure that deployments deliver real business value.
Does a retrieval strategy matter more than an LLM? Why?
EC: I like to say that retrieval is the king of RAG. It is an essential component, and without it, it would not be possible to generate high-quality outputs from internal data using an LLM.
However, as mentioned earlier, the most important aspect of AI is still the data itself. Ensuring that data is properly prepared and AI-ready is critical when starting any AI project, especially those that rely on querying internal organizational data.
Once the data is AI-ready and properly stored, retrieval becomes the key factor. It involves understanding user intent and selecting only the most relevant information needed to answer a query.
Even with the best LLM, without the right context gathered during the retrieval phase, the output will not be reliable.
In this sense, while LLMs are important, retrieval is the foundation of any RAG system and essential for effectively using both structured and unstructured data.
