A Box-commissioned Harris Poll survey of 1,640 IT decision-makers, conducted April 30-May 8, found 83% said their organizations were running AI agents, but only 36% of respondents using or testing agents had connected them to trusted internal content across many use cases. Nearly half reported an AI-related data-exposure incident, pointing to a gap between deployment and governance.
Box's 2026 State of AI in the Enterprise report found a large gap between reported agent adoption and the content controls needed to operate those agents safely. The Harris Poll surveyed 1,640 IT decision-makers in the United States, United Kingdom, France and Japan from April 30 through May 8, 2026.
According to the Box-commissioned survey, 83% of respondents said their organizations were running AI agents, including 19% that described agents as operating autonomously at scale. Yet among organizations using or experimenting with agents, only 36% said they had connected them to trusted internal content across many use cases.
Content access is the bottleneck
The survey found broad agreement that agents need company context: 96% called access to company-specific content and knowledge important or very important. The operational foundations were less common. Respondents cited security and privacy concerns, regulatory requirements, fragmented data, integration difficulty, missing access controls and poorly organized content as barriers to connecting agents with enterprise information.
These are not merely retrieval problems. An agent acting across documents, applications and workflows needs the same identity, permission, provenance and audit controls that govern human access. Box's data suggests many organizations are deploying agents before those controls are consistently available.
Governance trails deployment
Nearly half of respondents, 49%, said an AI tool had surfaced content to someone who should not have been able to access it; 16% characterized an incident as significant. Only 34% reported formal standards for how agents access company data, while 39% said they had comprehensive visibility across sanctioned and unsanctioned AI use.
Those figures are self-reported, not independently observed incident rates. Box also sells enterprise content-management products, so its framing has a commercial interest in content governance. The underlying research was conducted online among full-time IT professionals at companies with at least 100 employees who had managerial responsibility and input into AI decisions. The raw data was not weighted, and respondents selected their own maturity tier, so the results should not be generalized to every enterprise.
What teams can take from the findings
For data and platform teams, the useful signal is the mismatch between agent reach and control coverage. Before broadening an agent's access, teams can inventory trusted content, inherit existing permissions, log actions and sources, and test whether users can retrieve information outside their authorization. The survey does not prove that better content infrastructure alone produces business value. It does show that reported adoption is ahead of the mechanisms needed to make agent decisions traceable and access-controlled. For enterprise deployments, that makes governance part of the runtime architecture rather than a policy added after launch.
Key Points #
- 1Box's survey found 83% of respondents said their organizations were running AI agents, while only 36% of those using or testing agents had connected them to trusted internal content across many use cases.
- 2Forty-nine percent reported an AI-related data-exposure incident, but only 34% said they had formal standards governing agent access to company data.
- 3The Harris Poll survey covered 1,640 IT decision-makers in four countries, used unweighted self-reported data and was commissioned by Box, limiting how broadly the results can be generalized.
Scoring Rationale #
The survey provides a useful cross-country snapshot of enterprise agent deployment, content readiness and access-control gaps. Its practitioner value is meaningful but bounded because the findings are vendor-sponsored, self-reported and based on unweighted respondents rather than observed deployments.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.