A stunning 95% of enterprise architects and cloud infrastructure leaders have delayed or cancelled AI projects in the past year due to problems with data governance, compliance, or regulatory issues, according to a report published by Cloudera in partnership with Wakefield Research. Separate research from Google and MIT reached similar conclusions, finding that legacy data systems are blocking enterprises from scaling AI agents and damaging return on investment. Both studies point to the same underlying crisis: companies built their data infrastructure for human users, but AI agents require an entirely different foundation.

The Cloudera report surveyed 1,500 enterprise architects and cloud infrastructure leads, while the Google/MIT study polled 300 IT executives and heads of product, IT, data, and AI. More than half of respondents in the Google/MIT research said they've paused or delayed AI agent deployments to fix foundational data problems like silos or lack of context. Over half also reported that outdated data systems are preventing them from scaling agentic AI and are having a "significant negative impact" on their AI ROI. A wide majority of Cloudera's respondents said AI integrations have changed their data storage and architecture practices, AI workloads have increased infrastructure costs, and current data architecture requires a "significant overhaul" to meet AI goals. Nearly all respondents—98%—to the Google/MIT survey are already using agentic AI or plan to soon, with one in 10 using it widely and nearly three-quarters deploying it in limited fashion. The most common applications right now are customer service, IT systems management, and IT security. Looking ahead, enterprises plan to use AI agents in HR, finance, and supply chains.

The reports identify several specific barriers to AI agent deployment. Google and MIT point to entrenched silos where information sits in disconnected systems or is "pocketed away" in different departments with no integration layer, as well as difficult-to-access "dark" or unstructured data contained in formats like images, video, or PDFs. Insufficient access to real-time data is another obstacle, as legacy batch processing architectures make in-time action challenging. Perhaps most critically, agents often receive basic metadata rather than enterprise-specific semantics, so they struggle to make relevant connections or suggestions. "Without this deep understanding, data cannot be highly relevant to specific use cases," the report noted. Cloudera's research describes the need for a "great AI re-architecture," warning that "even if enterprises are ready to use AI, many are coming to the realization that the foundational infrastructure it relies on is not."

The performance gap between leaders and laggards is striking. Google and MIT found that enterprises giving AI systems access to more than 70% of their data—"data leaders"—see dramatically better results than "data laggards" who share 30% or less. Every single data leader says their agents make "mostly" or "consistently" accurate and relevant decisions, while just 22% of data laggards report that level of trust. To successfully scale agents, the Google/MIT report recommends that enterprise leaders first improve access to data by discovering, inventorying, and classifying both structured and unstructured information. After governance principles are applied, information can then be extracted and connected to AI agents. The report also urges companies to "put a premium on context" by giving agents enterprise-specific data, replace batch processing with streaming and event-driven pipelines, and think AI-native. Cloudera found a notable trend toward hybrid environments, with 66% of respondents moving AI workloads from public cloud back to on-premises or private cloud environments over the last 12 months. Another 25% plan to place more emphasis on a hybrid-first approach, 24% plan to increase on-premises spend, and 22% plan to increase edge spend. The bottom line: enterprises that don't rebuild their data foundations for AI will continue to abandon projects they can't make work. The divide between organizations that grant agents frictionless data access and those that don't may soon separate market leaders from those left behind, as governance designed for human users proves inadequate when thousands of agents interact with enterprise systems simultaneously.