A recent survey of 300 tech executives highlights a pretty jarring divide. On average, AI agents only have access to about 45% of a company's data. For the "data laggards," that number drops to 30% or less. Meanwhile, the companies actually seeing ROI—the "data leaders"—have cleared the path for their agents to access over 70% of their data.
The correlation between data readiness and trust is the most telling part. Only about half of all organizations actually trust the decisions their agents make. Yet, 100% of those "data leaders" trust their agents. This proves that the "hallucination" problem in agentic workflows is often just a data access problem in disguise. If the agent can't see the full context or the latest operational state, it's just guessing based on incomplete information.
The bottlenecks killing agentic ROI #
If you're trying to build an AI workflow or a full-scale LLM agent, you'll likely hit these three walls: Legacy System Friction: Most enterprise data is locked in silos that weren't built for API-first agentic access. This prevents agents from making decisions at the speed the business requires.Context Gap: There is a massive difference between having "access" to data and having the "business context" to understand it. Without governance and metadata, an agent might find the right table but misinterpret the column headers.The Scaling Paradox: 68% of struggling companies admit that legacy systems stop their agents from acting quickly. You can't "prompt engineer" your way out of a slow database or a locked API.
To actually get these things to work, the focus has to shift toward a deep dive into data management automation. Improving access to both structured and unstructured data is the only way to move from a "cool tool" to a system that actually automates 50% of business decisions. If you're just layering a fancy agent framework on top of a 20-year-old SQL server with no documentation, you're just automating the process of making mistakes faster.
[Next Is the AI bubble actually bursting or just correcting? →](/en/news/6120/)
[an AI side-hustle playbook](https://tanyan888.com/), with plenty of directly applicable cases.