When an AI initiative disappoints, the first suspect is usually the model. A new survey suggests the people closest to these projects see it differently.
On 16 September 2026, Collibra released a survey conducted by The Harris Poll from 5 to 11 August 2026. It covered 306 director-level or higher data management, privacy, and AI decision-makers at U.S. companies, with a margin of error of plus or minus 6.4 percentage points.
What the survey found
- 72% agree that the root cause of AI shortfalls almost always traces back to an unaligned or poor data foundation. Among companies with more than $100 million in revenue, the figure is 96%.
- Nine in ten organizations are actively deploying autonomous AI agents, and 76% have hit critical roadblocks moving from pilot to production in the past year.
- 51% spend significant staff hours manually reviewing and correcting agent output before it goes live.
- 87% say their teams regularly re-verify that the context their agents rely on is still accurate and current.
- 53% moved their AI function’s reporting line closer to the data organization in the past year, and 51% are investing in data lineage and documentation.
Collibra’s CEO, Felix Van de Maele, described the cost of skipping this work as a “hallucination tax.” Collibra sells data governance software, so it has an interest in the conclusion. Independent research points the same way: an August 2026 MIT Technology Review Insights report found that AI systems can reach only 45% of company data on average.
Two kinds of AI programs
The survey describes two groups. In the first, data trust is an afterthought, and people pay for it every week by correcting outputs by hand. In the second, teams are reorganizing around data: moving AI closer to the data organization, documenting where data comes from, and checking that the context their agents use is still current.
The second group is not avoiding AI. It is building it on something it can explain.
Lessons learned
- Start with one data domain. Pick the domain behind your highest-value use case and fix ownership, quality, and definitions there first, instead of attempting an enterprise-wide cleanup.
- Write down what the data means. A semantic layer that defines revenue, customer, or active account stops two agents from giving two different answers to the same question.
- Treat agent context as a product. Give every source an owner and a refresh schedule. Stale context produces confident, wrong answers.
- Invest in lineage. If you cannot trace an answer back to its source, you cannot defend it to an auditor, a customer, or a regulator.
- Track rework hours. Time spent correcting AI output is the clearest signal that the foundation needs attention before you scale.
What this means for your AI program
AI built on data you trust is not slower. It is the version that reaches production and stays there. Our AI-Ready Data Foundation package fixes one priority data domain and adds an Ask-Your-Data agent in six weeks, and the AI Strategy & Opportunity Sprint scores your data readiness before you commit to a build.
Sources
- Collibra: New survey from Collibra by The Harris Poll finds 72% of tech decision-makers feel AI initiatives today are falling short (16 September 2026)
- CIO Dive: AI failures often trace back to poor data foundation, survey finds
- MIT Technology Review Insights: Scaling AI agents with trustworthy data (12 August 2026)