A new joint study from Google Cloud and MIT Technology Review reveals that enterprise AI agent programmes are being held back not by the models themselves, but by the data foundations underneath them. Surveying 300 data and technology executives, the report finds that organisations willing to grant agents access to more than 70% of their enterprise data see dramatically more trustworthy outcomes than those that don't — yet most companies are still granting less than half.
The Adoption Curve Is Accelerating Fast
AI agents are already used to some degree by 83% of organisations, though only 10% are deploying them widely across the business, according to the survey — with the bulk of respondents drawn from companies generating over US$500m in annual revenue.
That is set to change quickly. Every organisation surveyed intends to use AI agents within the next two years, and 69% plan to deploy them widely. Andi Gutmans, Vice President and General Manager of Data Cloud at Google Cloud, and Ryan Polivka, Director of Product Marketing for Data Cloud at Google Cloud, describe this as a coming leap from early pilots to enterprise-grade operations.
More than half of respondents (55%) say their legacy data systems are the primary obstacle preventing them from scaling agentic AI across the enterprise.
"The so-called 'modern' data stacks that are glued together with disjointed parts were not built for the agentic era. When we analyse these legacy architectures under agentic load, we see they break."— Andi Gutmans, VP/GM, Agentic Data Cloud, Google Cloud
The Bottlenecks of Legacy Architecture
The report identifies four specific factors holding legacy platforms back: entrenched data silos, difficulty accessing and managing unstructured data, insufficient access to real-time data, and a lack of business context and semantics.
Organisations that grant agents access to more than 70% of enterprise data see the strongest outcomes — yet the average company still grants access to only 45% of its data for AI purposes. Andi Gutmans notes that fragmented architectures "create walled gardens that fracture access controls and dissolve governance perimeters," leaving agents to act on raw data without business understanding, which drives hallucinations.
The trust gap shows up clearly in the numbers: only 51% of executives overall trust their AI. But among those granting agents access to more than 70% of their data, every single respondent reports "consistently" or "mostly" accurate results — compared with three-quarters of restrictive organisations reporting inaccurate outcomes.
Priorities for the Next 12 Months
Surveyed organisations point to three priorities for scaling AI agents over the coming year: improving access to all types of data, strengthening data and AI model governance, and replacing batch processing with streaming and event-driven pipelines.
As Andi and Ryan put it, bolting AI onto an old architecture is "slow, expensive and frustrating for teams" — a patchwork approach that lacks deep integration and can quickly turn into a spiralling financial liability rather than a competitive advantage.