An August 2025 MIT report made waves when it revealed that 95% of generative AI pilots at companies are failing. The statistic sparked debate across the tech industry, but the report left one critical question unanswered: why?
New research from CData Software provides the answer. According to their State of AI Data Connectivity: 2026 Outlook report, which surveyed over 200 enterprise data and AI leaders, only 6% believe their current data infrastructure is ready to support AI initiatives.
“The era of AI being constrained by models is over. Today, AI is constrained by data,” said Amit Sharma, CEO and co-founder of CData. “The organizations winning with AI aren’t the ones with the best algorithms; they’re the ones with connected, contextual, and semantically consistent data infrastructure.”
The problem isn’t the sophistication of AI models. It’s the pipes that feed them data.
Time Lost on Data Plumbing
The numbers tell a stark story. A full 71% of AI teams spend over a quarter of their time on data plumbing—modeling data, implementing ETL pipelines, and configuring connectors—instead of building innovative features.
This isn’t just an annoyance. Integration challenges have forced 55% of organizations to delay AI launches, and 33% have had to completely rethink their data architecture. Another 51% report increased engineering costs because of these delays.
The survey identified data quality and integration as the top blockers to AI success, cited by 73% of respondents. Security and compliance concerns came in second at 66%. Meanwhile, 50% of organizations report a lack of pre-built connectors for enterprise applications.
The Real-Time Problem
Here’s where it gets more complicated. AI agents and customer service automation require real-time access to data. All of them. 100% of organizations deploying these use cases say real-time data is essential for AI agents.
Yet 20% still lack real-time integration capabilities, and most of those who have begun implementing it remain in the early stages. This gap creates a significant bottleneck for autonomous AI adoption at scale.
The correlation between real-time capabilities and AI maturity is clear. All organizations at the highest level of AI maturity support real-time integration. Among those at the lowest maturity level, only 40% have real-time capabilities in place.
Integration Complexity is Exploding
The average AI use case requires connecting to multiple systems. 46% of organizations need real-time access to 6 or more data sources for a typical AI application. Each connection adds architectural complexity and increases the burden on data teams.
The problem is more acute for AI-native software companies. These providers require three times as many external data integrations as traditional software vendors—46% need more than 26 integrations, compared to just 15% of legacy providers.
This suggests that AI features are inherently more integration-intensive. Companies building AI from the ground up recognize this and architect for scale from day one. Traditional vendors are still catching up.
The Infrastructure Divide
The research reveals a stark divide in AI performance. A full 60% of companies at the highest level of AI maturity have invested in advanced data infrastructure. Meanwhile, 53% of organizations struggling with AI implementations are hampered by immature data systems.
Organizations with mature AI deployments share a common trait: they’ve built centralized, semantically consistent data access layers. All high-maturity enterprises have implemented this capability. Among low-maturity organizations, 80% haven’t even started.
What does semantic intelligence mean in practice? It’s the ability to understand business context—how data relates across systems, what entities mean, and how metrics are defined. Without this layer, AI agents can’t reliably interpret or act on data.
Mitch Ashley, vice president and practice lead, software lifecycle engineering, The Futurum Group, sums up the issue, “The organizations succeeding with AI solved the data foundation first. Everyone else is learning that better models cannot compensate for insufficient infrastructure, broken data plumbing, and treating AI like just another tool.”
Investment Priorities Are Shifting
A full 83% of organizations are now implementing or planning centralized, semantically consistent data access. Only 9% rank AI model development as their top investment priority. The market has fundamentally shifted its focus from models to infrastructure.
“Organizations are realizing that AI success isn’t determined by the sophistication of their models. It’s determined by the maturity of their data infrastructure,” said Sharma. “The companies gaining real value from AI are the ones that invested early in connected, real-time data access. Those that haven’t will find themselves at a significant competitive disadvantage.”
What This Means for Tech Teams
The implications are clear. If you’re building or deploying AI systems, your success depends on three things:
Data connectivity: Can you access the data your models need when they need it? Custom-built APIs and manual data exports won’t scale.
Real-time integration: Batch processing is insufficient for agentic AI. You need live data access with low latency.
Semantic consistency: Your data needs shared definitions and relationships that span multiple sources. Fragmented context produces fragmented results.
The research also reveals a troubling confidence gap. While 73% of software providers say implementing AI features is mission-critical to their roadmap, only 9% are “very confident” their integration strategy can support AI development.
The Path Forward
The enterprises and software providers succeeding with AI aren’t the ones with the most sophisticated models. They were the first to solve the data infrastructure problem.
This means investing in governance, quality, and lineage. It means building or buying robust integration capabilities that support both batch and real-time patterns. It means creating semantic layers that give AI systems the context they need to act reliably.
The MIT report was right about one thing: Most AI pilots are failing. But the solution isn’t better models. It’s a better data infrastructure. Organizations that recognize this reality will pull ahead. Those who won’t keep spending time and money on pilots that go nowhere.
