According to separate research reports from Cloudera and a joint Google/MIT study, data silos, legacy batch processing architectures, and fragmented governance frameworks are causing up to 95% of businesses to delay or cancel their AI deployments.
The Great AI Re-Architecture and Project Delays
The enterprise war on silos during the 2010s successfully stitched together disparate databases into centralized warehouses. Now, autonomous software is rapidly undoing that progress. According to Cloudera’s Wakefield Research survey of 1,500 enterprise architects and cloud infrastructure leads, a staggering 95% of respondents reported delaying or cancelling AI projects—sometimes six or more—over the past year due to data governance, compliance, and regulatory hurdles.
AI workloads are fundamentally altering storage practices. They are driving up infrastructure costs and forcing organizations to admit that legacy systems cannot meet modern demands. As the Cloudera report highlights, foundational architectures require a complete overhaul.
The Google/MIT survey of 350 IT executives and product leads mirrors these findings. More than half of respondents have paused agent deployments to fix foundational issues like data siloes and missing contextual layers. High latency is actively preventing AI agents from executing decisions at high velocity, degrading overall system trustworthiness.
Why Legacy Data Systems Starve Autonomous Agents
Traditional automation follows rigid, hard-coded rules. AI agents, conversely, perceive, reason, and act dynamically based on real-time inputs. They demand multimodal, context-aware data that is instantly available.
Enterprise legacy systems consistently fall short across three major vectors, according to the Google/MIT research:
- Entrenched Silos: Data remains trapped in disconnected departmental systems, obsolete logs, or legacy Internet of Things (IoT) hardware without an integration layer.
- Dark and Unstructured Data: Critical business insights hide inside unstructured formats like images, videos, and PDF documents.
- Batch Processing Latency: Legacy batch architectures prevent in-time action, starving agents of the fresh context required to make accurate decisions.
Without enterprise-specific semantics, agents receive only basic metadata. They fail to build relevant connections across business domains.
Data Leaders Versus Data Laggards
Despite these friction points, adoption is widespread. Nearly 98% of respondents in the Google/MIT survey are either using agentic AI or planning to do so immediately. Current deployments cluster around customer service routing, IT systems management, and security anomaly detection.
A stark divide separates organizations based on data access. Google and MIT classify companies sharing more than 70% of their data with AI systems as “data leaders,” while those sharing 30% or less are “data laggards.”
Among data leaders, 100% report that their agents make mostly or consistently accurate decisions. Among laggards, only 22% share that level of trust. Frictionless access to operating systems and context remains the primary differentiator for scaling AI successfully.
The Resurgence of On-Premises and Private Cloud Environments
Enterprise data is rarely housed in a single location. Cloudera’s research shows that 97% of organizations move data between on-premises, SaaS, cloud, and edge environments monthly, with nearly one-third executing these transfers daily. Concurrently, 73% of respondents state that AI integration significantly increases data governance complexity.
This reality has triggered an unexpected shift back to private infrastructure. Over the preceding 12 months, 66% of Cloudera survey respondents migrated AI workloads out of public clouds and back to on-premises or private cloud environments. Organizations are adopting hybrid-first strategies to balance performance, strict data sovereignty, and cost controls.
As enterprise architectures evolve past legacy bottlenecks, building AI-native data systems designed for streaming pipelines and event-driven architectures is no longer optional.
