The gap between data and knowledge is the real bottleneck for agentic projects. MIT Technology Review Insights surveyed 300 data, AI, and technology executives: on average, only 34% of agentic AI projects reach production. The reported failure points are legacy systems, security and privacy, and lack of knowledge and context. Even technology companies do not distance themselves from the average: selling innovation does not make it easier to put it into practice on internal processes.

The report identifies three kinds of knowledge that agents must acquire: semantic knowledge, episodic memory, and procedural knowledge. The first concerns what data means within a company context; the second, accumulated cases; the third, the steps to carry out tasks. Without these three dimensions, an AI agent does not decide: it completes. And systems that complete without understanding produce errors that are difficult to trace.

The most interesting figure is not 34% itself but the correlation with production leaders. A small group of organizations takes an average of 61% of projects beyond the pilot phase. These companies do not necessarily have more raw compute; they show stronger knowledge capabilities, especially semantic knowledge. They first solved the problem of making agents understand what data means inside the organization, rather than increasing model parameters. This is a shift in the bottleneck: from compute capacity to integration quality.

Data fragmentation is cited by 55% of respondents as the top challenge. Corporate systems do not share information sufficiently; departments create silos that agents cannot query coherently. Production leaders, by contrast, cite security and privacy as a major concern in 72% of cases. This is not a contradiction: those who have already consolidated data access begin to treat compliance constraints as the next layer of complexity. For organizations evaluating self-hosted deployment, this reverses the hierarchy of priorities: the starting point is not the GPU but the architecture that lets agents read data where it resides, with access controls and audit trails.

The planned investment priorities confirm the shift in gravity. Executives point to retrieval technologies: ingestion pipelines, AI-ready APIs, and retrieval-augmented generation. They combine these with AI evaluation agents and knowledge graphs. This is not a model-centric feature set; it is an integration stack. Knowledge graphs, in particular, give structure to relationships among entities, processes, and business rules, reducing the chance that an agent produces a formally plausible but contextually wrong answer.

The experts interviewed in the report identify a knowledge layer as the main way to strengthen the link between data and agents. This intermediate architecture has deployment implications: the closer the layer is to the data, the smaller the surface exposed to external providers. In regulated environments, keeping retrieval and knowledge graphs on infrastructure under the organization's control is not an operational detail; it is a condition for closing the gap between pilot and production.

The report suggests that the next competitive advantage will not be measured in teraFLOPs but in the ability to turn fragmented corporate archives into queryable knowledge. Organizations that treat agents as an API problem risk remaining stuck at 34%.