Agentic AI is forcing enterprises to confront a problem they have been working around for years: their data infrastructure was not built for systems that can act on their own.
For much of the past 18 months, the enterprise technology industry has talked about agentic AI as if production deployment were simply a matter of time. But moving agents into production is exposing a harder reality. If the data behind them is fragmented, stale or poorly governed, the consequences are no longer confined to a bad dashboard or a slow report. They can show up in live decisions, automated workflows and customer-facing systems.
That is the context for Confluent’s 2026 Data Streaming Report. Based on responses from 4,625 IT leaders across 14 countries, the report finds that just around a third (32%) of organizations are currently using agentic AI in production, up from 29% in 2025. That suggests progress, but not a breakthrough.
The more revealing finding is what happens once organizations do reach production. Among those already using agentic AI, more than three-quarters (77%) report stalled projects in response to challenges. A further 61% report project abandonment as a problem.
Reaching production, then, is not the end of the challenge. In many cases, it is where the harder problems begin.
Why agentic AI is getting stuck
Ask most executives why agentic AI projects stall, and you’ll likely hear about model quality, cost, or change management. The survey data tells a broader story.
The biggest existing barrier to agentic AI success, cited by 69% of respondents, is a skills gap and lack of organizational readiness. LLM reliability and non-determinism follows closely at 68%. But the data issue is almost as prominent: 66% cite data infrastructure and quality as a challenge.
As I’ve written before, AI agents are fundamentally dependent on the data they’re given. An agent that draws on fragmented, stale, or poorly governed data will produce unreliable outputs regardless of the quality of the underlying model. The survey captures this clearly: 74% of respondents identify data silos as a major or frequent challenge, 72% flag inconsistency of data sources, and 71% cite uncertain data lineage, timeliness, or quality.
These are persistent infrastructure issues that organizations have been struggling to manage for some time. What’s changed is that agentic AI has made the consequences of getting this wrong much more visible.
The infrastructure gap is widening
One finding from the report deserves particular attention from IT and data leaders. The proportion of respondents identifying insufficient infrastructure for real-time data processing as a major or frequent challenge has risen from 61% in 2025 to 72% in 2026. That’s a 15-point year-on-year increase in those calling it a major issue specifically.
The direction of travel here is concerning. As organizations push more AI workloads into production, the demands on data infrastructure increase. Batch processing architectures and fragmented pipelines may have been adequate for some analytics use cases, but they are poorly suited to the continuous, low-latency data access that agentic systems require.
What practical progress looks like
For organizations looking to move agentic AI from pilot to durable production deployment, the data suggests a few priorities worth acting on.
Fix the data foundation before scaling the AI
The temptation when pilots stall is to iterate on the model or the agent architecture. But if the underlying data is siloed, inconsistent, or delayed, model refinement alone will not solve the problem. An honest audit of data quality, lineage, and governance is a more productive starting point than most teams expect.
Treat real-time data access as a first-order requirement
Agentic AI systems make decisions based on the current state. An agent operating on data that is hours or days old is working from a fundamentally incomplete picture. Organizations that are still relying on batch pipelines to feed AI systems will find this becomes an increasingly visible limitation as use cases mature.
Address the skills gap with infrastructure choices, not just hiring
The 69% of respondents citing skills gaps as a barrier reflects a real constraint, but to my mind, it’s partly an infrastructure problem in disguise. Managed platforms that reduce operational complexity can lower the specialist knowledge required to build and maintain data pipelines, making organizational readiness less dependent on recruitment alone.
Don’t mistake a working pilot for a solved problem
The finding that 77% of organizations running agentic AI in production still report stalled projects is a useful corrective to the assumption that reaching production is the finish line. Sustained performance requires continuous data quality, governance, and monitoring. And this requires investment well beyond the initial deployment.
The good news
Agentic AI is not creating these data infrastructure problems, but it is making them harder to ignore. Yet, despite the challenges, there are signs that priorities are starting to shift. For the first time, IT leaders in the survey rank investment in data streaming platforms ahead of AI and ML solutions as a top strategic priority. Organizations at last appear to be looking beyond the agent itself, and towards the data foundations that determine whether it can operate reliably in production.
Explore the full findings at: https://www.confluent.io/resources/report/2026-data-streaming-report/