Summary

• Confluent surveyed 4,625 IT leaders and found 72% say a lack of real time data infrastructure is stalling efforts to scale AI.
• Data quality, lineage, timeliness, fragmented ownership, and infrastructure limits are holding back agentic AI deployment.
• The findings shift attention from model investment to the data foundations needed for production systems and business value.

Confluent says poor data infrastructure is becoming a major constraint on enterprise AI, with nearly three quarters of IT leaders reporting that a lack of real time capability is slowing efforts to scale AI initiatives.

The company’s 2026 Data Streaming Report is based on responses from 4,625 IT leaders across 14 countries. It found that 72% say a lack of real time data infrastructure is stalling AI growth, while many organisations are reaching the point where additional AI spending cannot compensate for weak data foundations.

The barriers are familiar across enterprise technology estates. Confluent found that 72% of IT leaders have encountered at least three challenges when scaling AI initiatives. Common problems include insufficient infrastructure for real time data processing, uncertainty around data lineage, timeliness and quality, and fragmented ownership of data.

Those weaknesses are also slowing agentic AI. Two thirds of IT leaders cited data infrastructure and data quality issues as barriers to agentic adoption, while only 32% said they have agentic AI in production. Most are experiencing delays.

Shaun Clowes, chief product officer at Confluent, said: “Most organisations do not have an AI investment problem, they have a data problem.” His diagnosis is blunt, but it reflects a practical constraint that has followed enterprise AI from experimentation into production: models need timely, governed, and contextual data before they can be trusted with operational work.

Many organisations have spent the past two years buying AI tools, testing copilots, and running pilots. Production deployment is harder because systems need live context, accurate records, governed data flows, security controls, and integration with existing applications. A more capable model cannot rescue an AI workflow if the data feeding it is stale, inconsistent, or poorly owned.

A customer service agent cannot act reliably if it cannot see current account information. A supply chain model cannot make useful recommendations if inventory, logistics, and demand data arrive too late. A risk system cannot be trusted if lineage is unclear. An internal productivity agent can create new problems if it pulls from inconsistent sources and acts without sufficient context.

Clowes said that, as organisations move beyond experimentation and deploy AI across critical business processes, “those gaps become harder to ignore”. That is where the report moves from general AI enthusiasm into the machinery of enterprise implementation. The issue is not whether companies want to use AI, but whether their data estates can support systems that need to understand what is happening across the business as events change.

Confluent’s report points towards data streaming as part of the answer. The company says 88% of IT leaders rank data streaming as a high investment priority, compared with 82% citing AI and machine learning technologies. It also says 94% have seen, or expect to see, data streaming increase the impact of AI investments, while 90% say it helps ease AI adoption.

The report includes a business value claim as well. Half of organisations reported at least 5x return on investment from data streaming investments, while 88% reported 2x ROI or more. Respondents linked data streaming to customer experience, AI innovation, risk management, and faster time to market.

As vendor research, the findings should be read with the necessary caveat that Confluent sells data streaming technology. The underlying diagnosis is still credible. Enterprise AI projects repeatedly run into data access, governance, ownership, latency, and integration problems. Buying more AI capability does not solve those weaknesses if the system lacks reliable business context.

Agentic AI makes the data layer more important because agents do not merely answer questions. They may retrieve information, call tools, trigger workflows, make recommendations, or take action. That makes data quality part of the control environment. If the wrong data reaches the agent, the wrong action may follow.

The management challenge is as significant as the platform challenge. Fragmented data ownership reflects organisational structure as much as technology. Departments own systems, teams define fields differently, and business units protect their own processes. A data streaming platform can improve technical flow, but governance still requires decisions about accountability, quality, access rights, and priorities.

The next phase of enterprise AI will reward operational discipline rather than visible experimentation. Organisations that can connect models to live data safely, govern access, monitor outputs, and integrate AI into workflows are more likely to convert pilots into measurable productivity or service improvements.