Enterprise AI has entered a strange phase. The demos are polished, the models are smarter than ever, and yet many deployments still stall before they ever reach production. Across industries, organizations are discovering that the challenge is no longer whether AI models can reason, summarize, or generate. The challenge is whether they actually understand the business context they are operating within.
That realization is beginning to reshape how companies think about AI infrastructure. Instead of focusing solely on larger models or bigger context windows, some organizations are turning their attention toward the hidden layer underneath AI systems: the organizational knowledge substrate. For Naboo CEO and Co-Founder Gilad Salinger, this shift represents one of the most important changes happening in enterprise AI today.
The Limits Of Retrieval
According to Salinger, the industry’s core issue is no longer model intelligence.
“The models stopped being the bottleneck,” he explains. “We hit a point where frontier models were good enough that the demo always worked and the deployment always stalled.”
What changed, he argues, was not the quality of the reasoning engine but the quality of the information those systems could access. Enterprises were feeding advanced AI systems documents, dashboards, and repositories filled with outputs, but not the reasoning behind them.
“A model can read every document in the company and still not know why the system was built the way it was,” Salinger says. “That is a substrate problem, not an intelligence problem.”
This insight became especially clear while observing how experienced engineers answered questions AI systems consistently missed. Senior engineers could explain not only what decisions were made, but why they were made, including tradeoffs, abandoned alternatives, historical incidents, and organizational constraints. Much of that reasoning never existed in formal documentation.
“The reasoning, the rejected alternatives, the constraints, all of it lived in people’s heads and in the exhaust of the work,” Salinger explains. “The review comment, the ticket, the argument in chat.”
Why Bigger Context Windows Are Not Enough
Many organizations still believe enterprise AI performance can improve simply by expanding context windows or refining retrieval systems. Salinger believes that assumption is beginning to collapse.
“Both assume the answer exists as text somewhere,” he says. “When the reasoning was never written down, more windows and better search just deliver the same incomplete picture with higher confidence.”
Instead of retrieval alone, Naboo focuses on reconstructing organizational reasoning from fragmented signals spread across tools like Git repositories, issue trackers, incident reports, design comments, and internal chats.
The distinction matters because retrieval only surfaces artifacts, while reconstruction attempts to rebuild the logic behind them.
“Retrieval is a lookup,” Salinger says. “Reconstruction is comprehension.”
That difference becomes critical when AI systems move beyond chatbots into autonomous agents capable of taking action. A chatbot generating an imperfect answer may create frustration. An autonomous agent making a technically correct but contextually incomplete decision can create operational risk.
The Rise Of “Compiled Context”
Salinger believes the broader industry is already shifting toward this new architecture.
He points to companies like Microsoft, Google, and Pinecone increasingly emphasizing compiled knowledge systems rather than relying entirely on retrieval at query time.
“The industry is converging because retrieval was built for humans who read, and agents do not read; they act,” he says.
This approach centers around what Naboo calls “decision trails,” structured representations of organizational reasoning that connect decisions, tradeoffs, contributors, dependencies, and downstream impacts into a continuously reconstructed graph.
“It is a graph, not a log,” Salinger explains. “A decision trail tells you why, and what it touches.”
The goal is not simply documentation. It is creating auditable semantic infrastructure that AI systems can reason against before acting.
Why Enterprise AI Still Struggles To Scale
Salinger argues that many organizations fail to operationalize AI because they continue treating context as something provided by the model itself rather than infrastructure the enterprise must own.
“Pilots succeed because someone hand-curates the context for one narrow use case,” he says. “That does not generalize.”
Organizations that scale successfully, by contrast, invest in shared context layers that multiple agents and applications can access consistently. Without that layer, every new AI initiative essentially starts from scratch.
“The tell is simple,” Salinger says. “If your second AI project was as hard as your first, you did not build infrastructure; you built a demo twice.”
From Feature To Foundation
As AI agents become more autonomous, Salinger believes semantic infrastructure will evolve much like cybersecurity did over the last two decades, from an optional feature into foundational enterprise architecture.
“Right now, it is treated as something each AI application figures out for itself,” he says. “That is exactly how security was treated before everyone learned that bolting it on per app does not work.”
In his view, the long-term competitive advantage in enterprise AI may not come from model access at all.
“Context was the actual product, and the model was the commodity,” Salinger says. “The companies that win the agent era will not be the ones with the best model access. They will be the ones whose agents actually understand the organization.”
