Data Federation and Context for Enterprise AI
Starburst SVP Jitendra Aswani argues that data federation, not centralization, is the critical infrastructure for scaling enterprise AI. The discussion highlights how fragmented data and missing context cause agent failures, proposing a new architecture for agentic workflows.
The Shift from Centralization to Federation
Enterprise data architecture is undergoing a fundamental paradigm shift. For decades, the dominant strategy was centralization: moving data into lakes or warehouses to simplify access. However, as Jitendra Aswani, SVP of Engineering at Starburst, argues, this model fails to scale with the increasing entropy of modern data landscapes. Traditional enterprises now manage between 52 and 200 distinct data sources, a fragmentation that centralization exacerbates rather than solves. The result is a costly, brittle ETL ecosystem where pipelines break, data shapes change, and insights are delayed by days. Federation, by contrast, places a virtual compute layer over existing data sources, allowing queries to run where the data resides. This approach preserves data locality, reduces latency, and eliminates the operational debt of constant data movement.
The Agentic Data Challenge
The rise of AI agents has intensified the urgency of this architectural shift. Agents do not behave like human analysts; they generate queries at machine speed, often firing unbounded requests when they lack access to the right data or context. Aswani notes that agents operating with limited visibility become "overzealous," continuing to query indefinitely rather than acknowledging data gaps. This behavior places immense strain on compute infrastructure, leading to performance degradation and increased costs. More critically, it leads to incorrect business decisions. When agents cannot access the full data landscape, they provide incomplete or contradictory answers, undermining trust in AI-driven insights.
Context as the New Competitive Moat
Beyond raw data access, the key differentiator for enterprise AI is context. Aswani distinguishes between a data catalog, which tracks where data lives, and a context layer, which encodes business rules, ontologies, and semantic definitions. For example, the definition of "churn" or "fiscal year" varies by industry and company. Without this nuanced context, agents cannot make accurate decisions. Starburst is investing in a "context federation" strategy, unifying scattered business knowledge into a coherent graph that agents can query. This approach transforms AI from a simple query tool into a strategic partner capable of executing complex workflows. The implication for businesses is clear: those who invest in robust data federation and context engineering will unlock the true potential of agentic AI, while those who rely on legacy centralized architectures will face escalating costs and diminishing returns.
Key insights
-
Data centralization is no longer viable for scaling enterprise AI due to increasing data entropy and fragmentation. Federation allows for querying data in place, reducing ETL overhead and latency.
Impact: Enterprises can reduce infrastructure costs and improve data freshness by adopting federation, enabling faster decision-making and more agile AI deployments.
-
AI agents generate significantly higher query volumes than humans, often firing unbounded queries when data access is restricted. This behavior strains compute resources and leads to inefficient workflows.
Impact: Organizations must implement auto-scaling compute and robust guardrails to handle agentic traffic, preventing performance degradation and cost overruns.
-
Context is distinct from metadata; it includes business rules, ontologies, and semantic definitions that are critical for accurate AI decision-making. Without a unified context layer, agents produce inconsistent results.
Impact: Investing in context federation creates a competitive moat by enabling AI agents to understand business nuances, leading to more reliable and actionable insights.
-
Traditional BI dashboards are insufficient for dynamic business questions. Conversational analytics interfaces allow users to ask natural language questions and trigger workflows directly from insights.
Impact: Conversational analytics democratizes data access, enabling non-technical users to derive value from data and accelerate business processes through direct action triggers.
-
Data fragmentation is a persistent challenge in traditional enterprises, with 52 to 200 data sources per organization. This fragmentation hinders AI adoption by limiting agent access to comprehensive data views.
Impact: Solving data fragmentation through federation is essential for enterprises to fully leverage AI, as isolated data sources prevent agents from making holistic business decisions.
Action items
-
Audit current data architecture to identify fragmentation points and evaluate the feasibility of implementing a federation layer over existing data sources.
Impact: Identifying fragmentation opportunities allows enterprises to reduce ETL costs and improve data accessibility for AI agents, enhancing overall system efficiency.
-
Implement auto-scaling compute infrastructure to handle the increased query volumes generated by AI agents, ensuring performance stability during peak agentic activity.
Impact: Auto-scaling prevents performance degradation and cost overruns caused by machine-speed query generation, ensuring reliable AI operations.
-
Develop a context layer that integrates business rules, ontologies, and semantic definitions from various sources, creating a unified knowledge graph for AI agents.
Impact: A robust context layer enables AI agents to make accurate, context-aware decisions, reducing errors and improving the reliability of AI-driven insights.
-
Deploy conversational analytics interfaces that allow users to ask natural language questions and trigger workflows, replacing static dashboards with dynamic, interactive tools.
Impact: Conversational analytics increases user engagement and accelerates decision-making by providing immediate, actionable insights without requiring technical expertise.
-
Establish guardrails and monitoring for AI agents to prevent unbounded query generation and ensure they operate within defined data access boundaries.
Impact: Guardrails improve agent efficiency and reduce resource waste, while monitoring provides visibility into agent behavior, enabling continuous optimization of AI workflows.
Quotes
“The federation is the only model that scales with entropy.”
“Every enterprise AI project that's failing is fighting the same hidden battle.”
“Context is more than metadata.”