A Third of Deployed AI Agents Hampered by Lack of Access to Internal Data

AI agent projects struggle with access to enterprise knowledge: on average, only 34% make it to production. Organizations that are better equipped with knowledge report 61% of projects moving beyond the pilot phase and target investments in data foundations and graphs.

Production hindered by fragmentation and legacy systems

In enterprises, the production of AI agents remains limited to an average of 34% of projects. The main points of failure cited include legacy data systems, security and privacy concerns, as well as a lack of knowledge and context. Data fragmentation is identified as the number one challenge for expanding agents' access to knowledge by 55% of respondents. Among the most advanced organizations, 72% consider security and privacy to be a major issue. Even high-tech companies face these difficulties.

High-performing leaders through knowledge capabilities

A group of production leaders manages to advance an average of 61% of their agent projects beyond the pilot phase. These organizations possess stronger knowledge capabilities than others, particularly in the semantic dimension. Enhanced knowledge capabilities are associated with a higher production rate.

Investment priorities: foundations, RAG, and graphs

Most companies are looking to strengthen the link between their data and their AI agents. The executives surveyed expect that strengthening the structural foundations connecting data and agents will have the most significant impact on the quality of decisions made by the agents. Experts believe that a knowledge layer is a preferred means to achieve this goal. Investment priorities include retrieval technologies such as ingestion pipelines, AI-ready APIs, retrieval-augmented generation, as well as AI evaluation agents and knowledge graphs.

Why knowledge matters for the agent

Enterprise AI agents often suffer from a lack of knowledge, understood as the ability to interpret the meaning of data in the context of the organization. This understanding is necessary for reasoning, decision-making, and action; its absence increases the risk of erroneous and unreliable decisions. This lack is among the major reasons why use cases for agentic AI do not reach production. Competitive pressure makes it urgent to address this deficit, lest investments already made be wasted and more efficient competitors gain an advantage with their agents. These findings are based on a survey conducted with 300 executives and a report aimed at assessing agentic knowledge capabilities—semantic, episodic memory, procedural—exploring access to knowledge challenges, and examining measures taken to address them.