Agentic AI is ready—but enterprise data is not
Enterprise AI agents are autonomous software systems that can interpret context, make decisions, and take actions across business applications and data platforms without constant human prompts, but their reliability depends on clean, well-governed, and well-orchestrated data that most organizations still lack. In keynotes at recent industry summits, leaders made it clear that the agentic era has moved from theory to practice: “The agentic era is not a roadmap item,” Andi Gutmans told attendees, arguing that autonomous agents are already driving live operations for telecom networks. Yet when Sridhar Ramaswamy described today’s data landscape, his verdict was blunt: “Our data is a mess”. That tension—agents ready for prime time, data still stuck in silos—explains why many enterprise AI ambitions stall before they reach meaningful deployment.
From pilots to production: AI infrastructure requirements get real
The uncomfortable truth is that most AI initiatives fail not because the models are weak, but because the operational groundwork is missing. Gutmans says the most common reason projects never reach production is scope: leaders try to “boil the ocean” instead of starting with sharply defined use cases and clear success metrics. Ramaswamy adds that data must be secured, governed, compliant, and discoverable before any agent can be trusted to operate across finance, logistics, or other core functions. That is an infrastructure checklist, not a brainstorming exercise. Building digital twins of telecom networks and using agents to drive autonomous operations only works when data platform capabilities, graph mapping, full-text search, geospatial data, and agentic orchestration are integrated into a coherent operating environment. Enterprises that skip this step stay stuck in perpetual pilot mode.
Messy data estates: the hidden barrier to autonomous agent deployment
Every executive now admits the same problem: their data estate is a mess. Ramaswamy describes exploding volumes of structured, unstructured, and partially structured data trapped in silos across the enterprise. Gutmans echoes the point, noting that 90% of enterprise data is unstructured and not catalogued at all. One customer arrived with 20,000 database tables they wanted to make agent-ready—a scope far beyond what any team of human data stewards could curate. Without a knowledge layer that enriches and contextualises this chaos, enterprises cannot safely allow agents to act on their behalf. Google’s response is a Knowledge Catalog that uses agents to infer relationships from query logs and build structured knowledge fast enough to support other agents. Snowflake aligns with this philosophy by deploying AI close to governed data and breaking down silos across multiple clouds. The message is stark: if you do not fix data, you do not get autonomy.
Orchestration, control planes, and agents policing agents
Raw model power no longer differentiates serious enterprise AI strategies; orchestration does. Ramaswamy points to the Model Context Protocol, an open standard for connecting AI models and agents to external tools, data sources, and business applications. On top of that, he argues for an agentic control plane that coordinates actions across software and departments, governing how agents work together in finance, logistics, and beyond. Gutmans describes a complementary evolution: agent‑verification architectures where agents quality‑assure other agents, even voting on whether an answer meets a high bar before it reaches the user. A year ago, he says, foundation models could not support this level of reasoning; now Google has rewritten its first‑party agents to take advantage. This is operating system‑level design for AI, with policy, routing, and verification baked in. Enterprises that fixate solely on model choice miss the real leverage point: who orchestrates which agent does what, when, and under whose rules.
What this means for everyday employees—and what comes next
For workers, the implications are both exciting and demanding. “With Snowflake Intelligence, every user in your enterprise now has an AI agent,” Ramaswamy said, arguing that tasks once reserved for specialised teams can now be handled by agents embedded in daily workflows. Sanofi’s Emmanuel Frenehard reports agents already supporting procurement, HR, IT support, and sales. Gutmans goes further: every individual contributor can have a team of agents working in parallel on their behalf. Operators are already comfortable letting agents autonomously escalate customer support issues or place orders at the £50,000 level, while still insisting on human oversight for commitments closer to £20 million. Trust, he notes, will grow over time, pushing the pendulum toward more autonomy. Industry leaders expect compound gains to continue, urging enterprises not to think only about today’s pilots but “about what’s going to be” in the agentic era. The conclusion is clear: the future belongs to organizations that treat data quality and AI infrastructure as strategic assets, not afterthoughts.
