Most enterprise agentic-AI projects are stalling in production, and the problem is data, not the models

The gap between agentic AI in a demo and agentic AI in production is starting to show up in the numbers. In Confluent's annual Data Streaming Report, a survey of 4,625 IT leaders across 14 countries, 77 percent of those running agentic AI in production say projects have stalled over data-related problems, and 61 percent report outright abandonment. Adoption is real and still growing, but so is the stall rate.

The blockers IT leaders named are mostly not about the models. A skills gap and limited organizational readiness topped the list at 69 percent, followed closely by worries about LLM reliability and non-determinism at 68 percent, data infrastructure and quality at 66 percent, and governance, risk and compliance at 65 percent. Delays of one to five months are common, and some efforts stop indefinitely. About 32 percent of enterprises now run agentic AI in production, up from 29 percent a year earlier, so this reads as the painful leading edge of adoption rather than its collapse.

The recurring theme underneath is data, and it is more than a quality complaint. An autonomous agent acts on whatever data it is given, which turns shaky provenance and stale information into a safety problem rather than a reporting inconvenience: the agent will take real actions on data that nobody has verified. That is the same trust gap surfacing elsewhere this week, from a new Google standard for agents to verify the tools they connect to, to a benchmark on which the best model still clears only about a third of expert science tasks.

The finding lines up with the broader research mood. Gartner has predicted that more than 40 percent of agentic AI projects will be canceled by 2027, and a separate survey from Transcend put the share of enterprises delaying, scaling back, or abandoning AI over data-permission and governance gaps at 81 percent. None of this says agents do not work. It says the hard part is not the demo, it is wiring an autonomous system into messy real data and real accountability, and right now the capability keeps arriving faster than the plumbing and the trust needed to run it safely.