AI systems in the average large organisation can reach just 45 per cent of company data. In financial services that figure falls to 33 per cent, and in healthcare to 40 per cent. The numbers come from Scaling AI Agents with Trustworthy Data, an MIT Technology Review Insights report sponsored by Google Cloud. It draws on a global survey of 300 senior executives.
Respondents were chief data and analytics officers, CIOs, CTOs, chief AI officers, and heads of product, IT, data and AI. Most of their organisations earn $US500 million or more in annual revenue.
The report's central finding links data reach to confidence in output. Only 51 per cent of surveyed executives trust the accuracy and relevance of their AI agents' decisions.
The survey identifies 25 organisations, described as data leaders, whose AI systems can access more than 70 per cent of enterprise data. All of them report that their agents make mostly or consistently accurate and relevant decisions. Among data laggards, where AI reaches 30 per cent or less, just 22 per cent trust agent decisions.
The gap widens on delivery. Sixty-six per cent of laggards say legacy data systems prevent them scaling agentic AI. Sixty-eight per cent say legacy latency stops agents deciding at speed. Among leaders, 8 per cent report either constraint.
Across all respondents, 55 per cent say their current data platform is the primary bottleneck preventing enterprise-wide agent scaling. Fifty-four per cent have paused or delayed agent deployment specifically to fix foundational data issues such as silos and missing governance or business context. Half say legacy systems are significantly hurting return on investment from agent projects.
Dark data is the constraint
The report defines the missing material as dark data. That is largely unstructured content in images, video, PDFs and social media posts. It also covers structured data locked in operational systems, including customer support logs, server logs, IoT feeds and old HR records.
Entrenched data silos are the biggest platform challenge, cited by 50 per cent. Difficulty accessing and managing unstructured data follows at 40 per cent. Lack of access to realtime data sits at 34 per cent, and lack of business context and semantics at 32 per cent. Disjointed data and AI governance is cited by 28 per cent.
The report argues that metadata alone is insufficient. "A technically clean dataset is not enough," said Thomas Hoy, director of clinical data management at HCA Healthcare. "It also needs business definitions, ownership, lineage, quality signals, evidence, security, and usage constraints. When that context travels with a governed data product, users and AI-enabled applications can interpret the data more reliably."
Adoption is not slowing
Eighty-three per cent of organisations use AI agents to some degree today. Ten per cent are in widespread use and 73 per cent in limited use. Within two years every surveyed organisation expects to be using agents, and 69 per cent expect widespread use.
Current deployment concentrates in customer service at 62 per cent, IT systems management at 52 per cent and IT security at 51 per cent. Human resources is the fastest-growing planned use case, with 53 per cent intending to deploy agents in onboarding and benefits management within 12 months.
Success rates lag deployment. Forty-six per cent report success with agents in customer service, and 42 per cent in IT systems management and IT security.
The report closes with three recommendations. Activate all data including dark and unstructured assets, put a premium on business context, and prioritise AI-native systems when replacing platforms.