As companies move to deploy agent AI in real business operations, the central question is beginning to change. The issue is no longer simply how well an AI model can generate answers, but whether those answers can be traced to verifiable data and defended in audit, legal and compliance settings.
That was the argument put forward by Datastreams at the “Data Strategy Seminar in the Age of Agent AI,” held by the National Information Society Agency (NIA) in Seoul on March 26. Lee Young-sang, chief executive of Datastreams, delivered a presentation titled “Understanding AI Governance and Strategies to Strengthen AI Agents Through Data Fabric,” arguing that data governance and data architecture are becoming the decisive variables in successful AI adoption.
The seminar was organised around a practical concern now facing both public institutions and private companies: how to secure quality, reliability and accountability as AI systems move beyond generative assistance and into autonomous decision-making and task execution. Datastreams’ position was that the limitations of large language models cannot be resolved simply by upgrading the model itself. In real operating environments, the more consequential task is to design the data system and control structure within which the AI operates.
From model intelligence to data accountability
Lee said the more important question is not how intelligent an AI appears to be, but what data it has been designed to rely on when making decisions. In areas with very low tolerance for error, such as financial calculations, legal notices and civil complaint responses, a probabilistic LLM on its own remains structurally insufficient, he argued.
That point carries particular weight for organisations seeking to use AI not merely as a support tool but as part of an actual work system. In those settings, fluency is less valuable than consistency, traceability and institutional trust. The standard for adoption, in other words, is shifting from answer quality alone to the reliability of the data basis behind the answer.
Datastreams presented data fabric as the architectural response to that problem. Rather than allowing an LLM to produce answers freely, the company proposed a structure in which the model is directed to search only data sources that contain verified answers. According to the company, that structure combines metadata-based data virtualisation, data quality management, data lineage and business metadata so that AI systems operate only on verifiable data.
The company said that even when a user submits a natural-language query, the actual answer can be generated on the basis of database search results and the underlying SQL. That, it argued, makes it possible to build systems capable of handling audits, complaint responses and even legal review. In effect, explainability is embedded in the architecture itself rather than added after the fact.
AI governance as a pre-deployment requirement
Lee also framed AI governance not as a control mechanism applied after deployment, but as a policy and process framework that must be designed in advance. He argued that data quality management standards must also evolve beyond simple error removal and toward ensuring the consistency required for AI learning and inference.
That view reflects a growing tension in the market. Many organisations are accelerating AI adoption, but practical deployment is often slowed by unresolved questions around data provenance, accountability and operational trust. The more AI is expected to perform real tasks, the harder it becomes to rely on model capability alone.
Datastreams also cited a case involving Korea Expressway Corporation. According to the company, an earlier LLM-based question-and-answer system showed limits in actual business use, but the adoption of data governance and a data fabric structure helped move the system beyond one-off responses toward a tool that could preserve work context and support repeated use. The case suggests that evaluation standards for AI in the public sector are beginning to shift from simple answer generation to the ability to function as a repeatable and accountable work instrument.
The presentation also included a policy proposal aimed at smaller businesses. Lee argued that unlike large corporations, many small and medium-sized enterprises do not have the resources to build their own legal and governance systems, and therefore require a government-led “national AI compliance service system.” Such a framework, he said, would help companies identify and use the laws, regulations and ethical standards relevant to AI adoption through LLM and retrieval-augmented generation systems.
As agent AI spreads, the importance of public infrastructure that supports not only technical adoption but also regulatory compliance and accountability is likely to grow. Datastreams’ broader message was that the axis of AI competition is moving away from model performance by itself and toward trust structures that can withstand real operational demands.
That shift may prove especially important in sectors such as public administration, finance, legal services and civil affairs, where explainability and audit readiness are not optional. In those fields, the decisive factor may not be what answer AI gives, but what data that answer is based on.
