Nearly half of AI projects are stalling due to data problems, new study finds

For the first time, investment in data streaming platforms has overtaken AI and machine learning solutions as a top strategic priority among IT leaders—a shift that reflects a growing recognition that better models aren’t the answer if the data feeding them is broken. 

That’s one of the central findings of Confluent’s new 2026 Data Streaming Report, released this week and based on a survey of 4,625 IT leaders across 14 countries. While 88% of respondents ranked data streaming as a high investment priority, only 32% of organizations said they have agentic AI running in production. 

Andrew Sellers, VP of Technology Strategy and Enablement at Confluent, said the gap comes down to what the models are missing. “For most applications, the latest frontier models are generalizable enough to enable sophisticated applications if they have relevant context,” Sellers told IBM Think. “Most customers I speak with have data that is siloed, poorly governed and not readily discoverable without significant curation. Customers can create impressive demos with AI, but they struggle to deliver relevant context in real time.” 

The report found that 72% of IT leaders cited insufficient infrastructure for real-time data processing as a barrier to AI adoption—up from 61% in 2025—with data silos and governance gaps among the biggest drivers. 

Nearly half of respondents reported indefinite delays or outright abandonment of agentic AI projects. For those organizations, Sellers’ advice is simple: stop optimizing the model and fix the data. “The best investments they can make target improving enterprise data so that it is trustworthy, discoverable, contextualized and secure, rather than fine-tune a particular model,” he said. 

Then there’s the skills problem, which is getting worse. The report found that 71% of IT leaders cited a skills and expertise gap as a barrier to AI adoption, up from 66% last year. Sellers said the nature of AI development itself is part of what’s driving it: traditional QA tools assume software behaves deterministically—that the same input always produces the same output. AI doesn’t. “To build effective AI applications, developers must be experts at data engineering and distributed systems more than ever before.” Sellers said. 

The report points to shift-left data integration—moving governance and processing closer to the data source—as the practical path forward. Sellers’ first concrete step: annotate critical data with schema and business metadata, starting with basics like who owns it and what it’s for. “Data governance simply isn’t realistic without adding structure and meaning,” he said.