Introduction: The mandate has arrived,
the readiness has not

Enterprise AI was supposed to get cheaper and more reliable as it scaled. But for most organizations, it has done the opposite. Agents make errors, humans step in to catch them and the cost of every correct answer goes up instead of down. That gap between what AI was expected to cost and what it actually costs is the hallucination tax, and it’s now the most-significant gap between AI experimentation and AI ROI.

In August 2026, Collibra commissioned The Harris Poll to survey 306 U.S. data management, privacy and AI decision-makers, all at the director level or higher, to find out where that tax is being paid. So when we asked them where their AI initiatives actually break, we weren’t asking for a hunch. We were asking the people who own the outcome.

Their answer was consistent. More than seven in ten (72%) agree that when an AI initiative falls short at their organization, the root cause almost always traces back to an unaligned or poor data foundation.

That finding lands at an interesting moment. Nine in 10 decision-makers (90%) say their organization uses autonomous AI agents. But adoption is not the same as production. Nearly 8 in 10 decision-makers (76%) report that their organization has encountered roadblocks moving agents out of pilot in the last twelve months. Among organizations using autonomous AI agents, legacy systems and fragmented data architecture are the leading roadblock to agent integration (41%).

The pattern will be familiar to anyone who has spent time governing enterprise data: the technology has moved faster than the foundation beneath it. What has changed is the cost of that gap. A copilot that produces a bad suggestion wastes an analyst's morning, while an agent that acts on ungoverned context can misroute a customer, approve the wrong transaction or file an incorrect regulatory disclosure.

This is not simply a context problem or a control problem, but both at once. Context is what an agent needs to know before it acts, its sense of meaning, trust and policy; control is what the enterprise needs to be able to prove after the fact, who deployed the agent, what it did and whether it stayed within its boundaries. In this survey, the same organizations leading on context measures are also leading on control measures, which is itself the point: treating them as one problem, not two, is what correlates with relative success.

This report reflects what those 306 decision-makers told us, organized around the three questions every autonomous agent must answer before it acts: what this data means, whether it can be trusted right now and what policy governs how it can be used.

Ownership is consolidating, but unevenly. Ownership sits with the CIO or CTO at 32% of organizations, a CDO at 20% and a combined CDAIO at 19%, 9% have a Chief AI Officer (CAIO) and 6% say ownership lies with someone; while 13% have no single executive owner at all. The takeaway: more than one in ten organizations lack a single executive owner for AI data readiness and governance
Agents are widely used, but many organizations struggle to move them beyond pilot. Nine in 10 decision-makers (90%) report that their organization uses autonomous AI agents; 76% of all decision-makers report roadblocks to moving agents into production. Among organizations using autonomous AI agents, the most common roadblocks are legacy systems and fragmented data architecture making integration difficult (41%) and inability to verify or govern the unstructured data and context agents retrieve (34%)
Confidence in context does not equate to confidence in verification. Among decision-makers, 87% say they regularly re-verify that the context their agents rely on is still accurate and current, yet 39% say they have trouble verifying the quality or security of that context
Manual review tends to break down earlier than many organizations expect. Among decision-makers at organizations using autonomous AI agents, 42% say human review becomes operationally or financially impossible at 16 or more deployed agents. For 19%, it becomes impossible at one to five agents
Regulation readiness is real, but uneven. Nine in 10 decision-makers (90%) say their organization is preparing for AI regulation. decision-makers at organizations with $100 million or more in revenue are roughly twice as likely as decision-makers at smaller organizations to be investing in documentation and lineage to demonstrate AI performance and compliance (72% vs. 35%)

Key highlights

Context is what an agent needs to know before it acts, its sense of meaning, trust and policy; control is what the enterprise needs to be able to prove after the fact, who deployed the agent, what it did and whether it stayed within
its boundaries.