The biggest obstacle to enterprise AI agents may not be the AI models themselves, but the data infrastructure underneath them. Companies have spent years building databases and systems designed primarily for human employees, but autonomous agents need information that is accessible, contextualized and available in real time. Research from Google/MIT and Cloudera points to the same problem: organizations are enthusiastic about deploying agents, yet data silos, governance requirements and legacy architectures are delaying or even cancelling AI projects.
The scale of the problem is significant. A Cloudera survey of 1,500 enterprise architects and cloud-infrastructure leaders found that 95% had delayed or cancelled AI projects because of data governance, compliance or regulatory issues. Separately, more than half of the IT executives surveyed by Google/MIT said they had paused or delayed agent deployments because of foundational data problems such as silos and insufficient context. Legacy systems can also introduce latency, making it difficult for agents to make decisions quickly enough for real-time workflows.
The article identifies several specific weaknesses: data trapped in disconnected departmental systems, inaccessible unstructured information such as PDFs and images, outdated batch-processing architectures, and insufficient business context. An agent might technically be able to access a database but still lack the semantic information needed to understand what the data means or how it relates to a particular business decision. This is why simply connecting an AI agent to more systems does not automatically make it more capable or trustworthy.
Interestingly, the organizations getting the best results are those with what Google and MIT call “data leader” architectures. These companies give AI systems access to more than 70% of their enterprise data, compared with 30% or less among “data laggards.” The difference in trust is striking: all of the data leaders surveyed said their agents make mostly or consistently accurate and relevant decisions, compared with only 22% of data laggards. The broader lesson is that agentic AI is becoming an infrastructure problem before it becomes an AI-model problem—companies that want autonomous systems to act reliably will first need to modernize how their data is stored, connected, governed and delivered.
