According to a recent survey of 101 large organizations, infrastructures designed to provide business context for AI agents are being built faster than they can be trusted. The retrieval-augmented generation (RAG) method has now become the primary source of context, and native provider tools like OpenAI File Search and Vertex AI Search have surpassed specialized vector databases. However, a significant portion of organizations have observed that agents often deliver incorrect yet confident answers due to the lack or inconsistency of contextual data.
To address this issue, companies are developing semantic governance layers and hybrid architectures, but most of these infrastructures are still in the implementation phase. Although native retrieval systems are used more often, many organizations intend to maintain their independence by utilizing specialized or multi-architecture tools. Most organizations select retrieval systems based on ease of data access and operational performance, and after deployment, they monitor answer accuracy and security more closely. Overall, the report indicates that increased retrieval alone is not sufficient to bridge the trust gap, and there is a need for management, shared definitions, and genuine access control.