Model capabilities are advancing rapidly and token economics are improving, yet a fundamental challenge persists in deploying AI agents across enterprises. The context gap—the difficulty of supplying agents with relevant, timely information—remains a significant barrier preventing organizations from achieving meaningful returns on their AI spending.

To understand how the industry is addressing this problem, The New Stack spoke with Sri Desikan, an Elastic executive working on search capabilities. His perspective sheds light on why context delivery has become the critical path to moving beyond proof-of-concept demonstrations toward reliable, production-grade automation.

Understanding the Context Problem

The core issue is straightforward but consequential. According to Desikan, "frontier models are getting better and better and better," yet without proper context, "all that money you're spending on tokens is going to waste." The ability to furnish agents with needed information precisely when required represents the essential difference between impressive prototypes and systems that deliver sustained business value.

Enterprises recognize the limitation acutely. While AI models can query MCP servers to retrieve information, simply connecting a company's data repository to a bridge falls short of solving the problem. Desikan emphasizes that you cannot simply "shove all this data into [a model's] context window." This approach carries dual penalties: it consumes substantial computational resources through excessive token usage and introduces unacceptable latency.

The Pre-Indexing Solution

The path forward involves restructuring enterprise data into an AI-optimized index that maintains awareness of data locations and interdependencies. Desikan describes this as constructing a "container which stores metadata about everything in your enterprise." Through this approach, agents can locate pertinent information more rapidly while substantially reducing token consumption—in some cases, dramatically so.

An additional advantage emerges from pre-computing context through indexing. Organizations can "make sure that stale information" doesn't contaminate the agent's knowledge base, enabling companies to "know how the LLM is going to think." This predictability becomes increasingly valuable as agents assume greater operational responsibility.

Broader Considerations

Employing AI systems to prepare data for consumption by other AI agents introduces novel questions around data privacy and agent autonomy. These concerns warrant careful consideration as organizations implement such strategies. Nevertheless, the thoughtful approaches emerging across the industry suggest that challenges confronting enterprises today need not persist indefinitely—solutions are actively being developed and refined.