AGENTIC RAG
Retrieval-augmented generation (RAG) systems have emerged as the connective tissue between enterprise data and AI. RAG is “a pragmatic stratum for working with internal data,” said Olga Kokhan, CEO at Tinkogroup. “It offers the capacity to relate large language models directly to an organization’s structured and unstructured data, providing grounded, contextual outputs in real time.”
Next up is agentic RAG, or RAG systems capable of reasoning across enterprise data, deciding which sources or tools to use, validating outputs, and presenting answers with supporting evidence, said Philip Miller, AI strategist at Progress Software. “Agentic RAG turns knowledge retrieval into an active workflow. An AI agent can decompose a business question, query multiple structured and unstructured data sources, reconcile conflicting information, call approved tools, and surface an answer with citations, lineage, and context. In other words, it moves AI from answer generation toward governed action over enterprise data.”
Benefits: RAG enables easier access. “Analysts-dominated teams now can query complex datasets in natural language and get usable answers quickly,” said Kokhan. “It also accelerates decision making without a complete redesign of current data infrastructure.” In addition, in organizations drowning in dashboards, documents, tickets, policies, customer records, and operational data, the next generation of agentic RAG “can help business users get from question to insight faster while giving data teams a way to preserve governance, access controls, and auditability,” said Miller. “It can also reduce pressure on scarce data-engineering and analytics teams by making trusted data more self-service. The winners in enterprise AI will not be the companies with the biggest models; they will be the companies that connect agents to the most trusted, governed, and actionable data.”
Adoption issues: Data readiness may slow down RAG performance. “If your underlying data is inconsistent, poorly labeled, or fragmented, RAG will expose those faults immediately,” said Kokhan. “Companies are beginning to understand that better outputs depend less on the model and more on the quality and structure of the data behind it.” Miller pointed to “poor metadata, fragmented systems, unclear permissions, stale documentation, and weak evaluation practices that will lead to inconsistent or risky outputs.”
UNIVERSAL CONTEXT ENGINES
The passivity of data platforms is disappearing, being transformed into systems of action. This means “no longer just generating a dashboard or answering a question,” said Yasmeen Ahmad, product leader for Google Cloud’s Data Cloud. “We are converging real-time operational data with deep historical context so that autonomous agents can perceive, reason, and securely act on the business’s behalf in the same instant.”
Dynamic context engines, built for autonomous AI, will gradually replace passive data catalogs built for human analysts, Ahmad said. “Raw data alone is no longer enough. If an AI agent doesn’t understand the deep business semantics of your organization, for instance, the nuanced difference between your definition of ‘revenue’ versus ‘projected revenue,’ it is forced to guess. AI agents cannot understand the unique logic of an organization and are prone to making flawed recommendations. Context engines map and infer this business meaning across the entire cross-cloud data estate, spanning data stores, data platforms, and applications.”
Benefits: Context engines offer a range of benefits, including eliminating hallucinations, unlocking dark data, and offering automated intelligence at scale, Ahmad said. The context engine, supported by a well-governed semantic layer, allows for “strict semantic guardrails and verified, golden-query patterns. This essentially eliminates the hallucination of data connectivity or the fabrication of financial metrics.” Underlying business logic is locked in and verified, he added. Such engines also extract information from sources such as PDFs, documents, and images.
Adoption issues: “The most common hurdle organizations face right now is hitting the context ceiling,” Ahmad said. You can’t deploy autonomous AI agents on top of aging, fractured data architectures. “Legacy systems only inventory static, technical schemas rather than deep business meaning, so the AI is left without the context it needs to function. The result is unacceptably high latency, stale insights, and, ultimately, flawed AI recommendations that erode executive trust.” Trust overall is an issue with autonomous systems. “If an organization’s context engine cannot provide mathematically verifiable, human-readable lineage for an agent’s reasoning, compliance and risk teams will inevitably lock down the AI’s permissions, throttling its ROI and regressing it back into a glorified chatbot.”
OPEN SOURCE FORMATS
Today’s generation of leading platforms are built on open source formats, such as Delta Lake and Iceberg, which have greater flexibility in managing, organizing, and scaling data estates. “These formats enable both traditional big data workloads as well as agentic use cases by providing a uniform storage layer that simplifies access,” said Wade Tsai, distinguished engineer at EY. “The adoption of open formats and standards will help ensure that technology investments will not become obsolete, much like what HTTP and TLS did for the internet.”
Benefits: Open source formats “simplify interoperability and data sharing across different platforms, enabling unified data strategies with different technology providers,” said Tsai. This enables enhanced abilities “to identify strategic data assets, create data products from those assets, and support a data marketplace.”
Adoption issues: “As open formats evolve, they may define new features that are not easily supported by platforms that leverage them,” Tsai cautioned. “Or worse, introduce breaking changes that require migrations to new formats, which will take time and funding to accomplish. Also, fragmentation of the open formats could create a chaotic landscape that could lead to incompatibility and slow overall adoption.”