Survey: AI Agents Access Less Than Half of Enterprise Data on Average

A survey of 300 data and technology executives conducted for MIT Technology Review Insights suggests that most organizations are not yet giving AI agents sufficient access to enterprise data. The custom-content report, which was not produced by MIT Technology Review's editorial staff, found that AI agents have access to an average of 45% of company data across all respondents. Organizations labeled 'data laggards' provide 30% or less access, while a smaller group of 'data leaders' provide over 70% access.

The gap matters for trust and performance. Only around half of all surveyed organizations say they trust their agents' decisions to be accurate and relevant, but all data leaders report such trust. Two-thirds of laggards say legacy data systems limit scaling (66%) and prevent agents from making decisions at speed (68%), compared with only 8% of leaders reporting either constraint.

Respondents plan rapid adoption: all expect to use agentic AI within two years, and 69% expect widespread use. Top priorities for scaling include improving access to structured and unstructured data, adding business context to data and AI governance, and—among leaders—automating data management.

The findings should be interpreted with caution. The survey was produced by the custom content arm, the sponsor and methodology are not disclosed in this excerpt, and the self-reported measures of trust and success may be influenced by selection bias.

The reported figures come from a survey presented in sponsored material, so they should be read as an indicator of industry sentiment rather than an independently audited measure of every enterprise deployment. Even so, the underlying problem is familiar: agents cannot reliably act on information they cannot locate, interpret, or access with appropriate permissions. Organizations evaluating agent projects should therefore measure data coverage, freshness, provenance, and authorization alongside model accuracy. A capable model connected to fragmented or poorly governed records can still produce incomplete answers or take the wrong action. Improving the data layer may deliver more practical value than simply switching to a larger model.