Data governance failures undermine enterprise AI, survey finds

Poor data infrastructure is the culprit behind most struggling enterprise AI projects, according to research commissioned by Collibra. In a Harris Poll survey of 300 professionals working in data management, privacy and AI decision-making roles, 72% identified a weak data foundation as the primary reason their AI initiatives underperformed.

The findings are prompting organizations to restructure how they operate. Companies are realigning their organizational charts to bring AI functions under closer supervision by data teams. Among surveyed decision-makers, 53% reported moving AI reporting lines nearer to their data organizations. Within larger enterprises specifically, 58% are working to establish explicit accountability mechanisms for AI system outputs.

Felix Van de Maele, co-founder and CEO of Collibra, emphasized the stakes of poor governance: "Enterprises need to know which agents are operating, who owns them, what data and systems they can access, and what they are authorized to do. Without that visibility, governance gaps aren't discovered until something goes wrong."

The supervision burden

Data forms the essential foundation for AI agents to function autonomously. When that foundation crumbles, human workers must step in to oversee tasks the agents were supposed to handle independently.

The survey revealed that 87% of respondents have their teams regularly validating whether the information available to AI agents remains accurate and up-to-date. Beyond validation, more than half of organizations report that staff members dedicate substantial time—often hours—to manually reviewing and fixing the outputs generated by AI agents.

Constructing a data infrastructure suited for AI agents requires deliberate choices: standardizing definitions across the organization and embedding policies that agents can evaluate before taking action, according to Van de Maele.

Historically, enterprise data systems were built with human users in mind. Analysts would review dashboards and run reports, and when something seemed off, they would question the data and mentally fill in missing context. Van de Maele noted that AI agents operate differently: "When a definition is ambiguous, or context is missing, an agent can still produce an answer and act on it with confidence. So the work isn't simply cleaning up more data. Enterprises need to make context available in a form machines can understand and use: What does this data mean, can I trust it right now, and what am I allowed to do with it?"

Data access and regulatory readiness

An August report from Google Cloud and MIT Technology Review Insights underscores the data access challenge. The research found that AI systems currently reach only 45% of the data available within a typical enterprise. Additionally, just half of organizations express confidence that their AI agents produce outputs that are both relevant and accurate.

Beyond operational performance, data quality and accessibility matter for regulatory compliance. Nine in 10 business leaders are preparing for incoming AI regulations in the United States and internationally. Among enterprise leaders, 51% are investing specifically in data lineage and documentation capabilities to meet anticipated regulatory demands.

The push for governance comes as prominent figures in AI development have called for caution. Leaders from OpenAI, Anthropic, Microsoft and SpaceX's AI division have advocated for slowing advanced AI development, citing inadequate safety frameworks relative to the capabilities of frontier models.

Policy action is accelerating on multiple fronts. Senator Bernie Sanders plans to introduce legislation that would impose a temporary halt on advanced AI development. In California, Governor Gavin Newsom recently signed two bills creating requirements for independent audits of large language models. Across the Atlantic, the EU AI Act's provisions on AI literacy and restrictions on high-risk use cases became effective in August.