Across nearly every major industry, enterprise AI spending continues to climb, yet a persistent paradox has emerged: demos work, models are more capable than ever, and deployments still stall. Boardrooms that once debated whether to invest in AI are now grappling with a more uncomfortable question: why, after years of effort and significant budget, so few implementations actually scale. The gap between what AI can do in a controlled setting and what it reliably delivers inside a real organization has become one of the defining challenges of the current technology cycle.
At the center of that conversation is Gilad Salinger, CEO and Co-Founder of Naboo, a company that has built its thesis around a provocative diagnosis of why enterprise AI keeps failing. In Salinger’s view, the industry has been solving the wrong problem. They are focused on model benchmarks and retrieval improvements while ignoring the foundational layer that those systems actually run on. The result, he argues, is that organizations have been feeding world-class reasoning engines a fundamentally incomplete picture of themselves.
The Knowledge That Never Got Written Down
Salinger points to a specific pattern he observed that crystallized the issue for Naboo. When engineers encountered questions that AI systems consistently fumbled, their senior colleagues could answer them fluently, not because they had access to better documents, but because they carried context that had never been captured anywhere. The tradeoffs, the historical incidents, the constraints that shaped a particular decision, all of it lived in memory and in the scattered exhaust of daily work.
“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.” That realization led Naboo away from trying to improve how organizations retrieve what they’ve documented, and toward reconstructing what was never documented in the first place: signals that organizations already produce but never compile into anything a machine can reason against.
The distinction between retrieval and reconstruction sits at the heart of Salinger’s critique of where the enterprise AI market has been heading. Retrieval, he argues, is fundamentally a lookup operation. You surface the artifact, from the design document and the config to the write-up, but you don’t recover the logic that produced it. “Retrieval is a lookup. Reconstruction is comprehension,” he says. Agents that act on lookups, in his framing, end up making decisions that are locally correct but globally wrong, because they lack visibility into constraints that shaped the current state of a system years before they were pointed at it.
Why Agents Changed Everything
The urgency around this problem has sharpened significantly as AI systems move from conversational assistants to autonomous agents capable of taking consequential actions. A chatbot that returns an imperfect answer can be corrected by a human in the loop. An agent that ships a code change without knowing that a downstream billing job depends on a behavior nobody ever encoded in a document creates a different class of problem entirely.
“The industry is converging because retrieval was built for humans who read, and agents do not read; they act,” Salinger says. This shift is why he believes the broader market is beginning to move toward what he calls compiled context, knowledge that has been structured and resolved before an agent ever runs a task, rather than assembled at query time from loosely related material.
Salinger sees this convergence reflected in moves by major players. Microsoft’s GraphRAG was built because flat vector search couldn’t handle questions that span multiple documents. Pinecone, whose core product is a vector database, has begun compiling structured artifacts in advance so agents query a resolved layer rather than raw chunks. These are not the moves of companies doubling down on retrieval. They are acknowledgments that something underneath retrieval needed to change.
The Governance Stakes
The implications extend beyond engineering workflows into compliance and accountability. As AI agents take on more autonomous roles in regulated industries, the question of what the system knew when it made a decision, and whether that reasoning is auditable, is becoming a procurement requirement, not a feature request. “Explainability without traceability is theater,” Salinger says, “and traceability requires capturing the reasoning in the first place, which most architectures simply do not.”
Output logs, which most vendors currently provide as their answer to observability, tell you what the system said. They do not tell you what the system knew, what it failed to see, or whether a human reviewer could have identified the error from the trail. For security and compliance teams now involved in AI procurement decisions, those are the questions that matter, and the ones that current logging practices leave unanswered.
The Long Game
Salinger believes the organizations best positioned for the next phase of enterprise AI are those that have stopped treating context as something the model brings to the table and started building it as infrastructure they own. Much like cybersecurity evolved from a per-application feature into a foundational layer with its own budget and governance, he expects semantic infrastructure to follow the same arc, driven by the same forcing function: the cost of getting it wrong keeps rising.
“In two years, the models will be even more interchangeable than they are today,” he says, “and the durable advantage will lie entirely in the quality of the context that an organization can provide its agents.” The compiled context layer, in his view, is as foundational as the database. The companies that win the agent era, he argues, won’t be distinguished by which model they access. They’ll be distinguished by how well their agents actually understand the organization those agents are operating inside.
