Plenty of Data, Not Enough Understanding
There is a persistent and largely underreported problem eating into enterprise AI investments: AI agents that can process enormous volumes of data but cannot actually understand what that data means inside a specific organization. Data and knowledge are not the same thing. Data is the raw material; knowledge is the contextual interpretation of that material – what a number means in this company, what a process step implies in this workflow, what an exception signals in this particular business environment. Without that layer of interpretation, AI agents make decisions that are flawed, inconsistent, and ultimately too unreliable to deploy at scale.
A new report from MIT Technology Review Insights, based on a survey of 300 data, AI, and technology executives, quantifies how badly this gap is hurting organizations. On average, only 34% of agentic AI projects make it into production. That means roughly two-thirds of enterprise agent initiatives stall before they ever deliver value – a figure that should alarm any organization counting on AI efficiency gains to stay competitive.
What “Knowledge” Actually Means for AI Agents
The report measures organizational knowledge capabilities across three dimensions: semantic knowledge, which concerns how well agents understand meaning and relationships in data; episodic memory, which covers an agent’s ability to recall context from prior interactions; and procedural knowledge, which describes whether agents understand how tasks and processes are supposed to unfold. These are not abstract technical categories – they map directly to whether an agent can reason about a situation, not just pattern-match against a database.
The organizations that consistently get agents into production score higher across all three, particularly in semantics. The report identifies a segment it calls “production leaders” – organizations where an average of 61% of agentic projects advance beyond the pilot stage, compared to the overall average of 34%. That gap is not explained by budget or sector alone. It tracks closely with these organizations’ stronger knowledge infrastructure, especially their ability to give agents meaningful context around the data they ingest.
Where the Majority of Organizations Are Getting Stuck
Data fragmentation is the most commonly cited barrier to expanding agent knowledge access, flagged by 55% of survey respondents. When data does not flow adequately across systems – when it sits in silos separated by incompatible formats, ownership structures, or access controls – agents cannot build a coherent picture of what they are working with. They end up operating with partial information, which produces partial and often wrong conclusions.
Legacy data systems compound the problem. Organizations built their data infrastructure long before agentic AI was a real concern, and much of that infrastructure was never designed to serve the kind of continuous, contextual querying that AI agents require. Retrofitting those systems for agent readiness is expensive, slow, and organizationally complicated – it requires cooperation across IT, security, and business units that don’t always operate with aligned incentives.
Security and privacy concerns also rank as a major constraint, though the report reveals an interesting split. Among production leaders – the higher-performing organizations – security and privacy concerns are cited as a top challenge by 72%. That number is actually higher than among the broader population, which suggests that organizations sophisticated enough to push agents into production are also sophisticated enough to recognize how much sensitive data those agents are now touching. They aren’t less worried; they’re more aware of what’s at stake.
Even well-resourced, high-tech firms struggle with these failures. The report makes clear this is not a resource problem that more spending alone can fix. The gap between data availability and contextual knowledge access exists even where data infrastructure is relatively mature. The shortcoming is structural, not budgetary.
What Separates the Organizations Getting It Right
Production leaders share a common orientation: they invest in the connective tissue between data and agents rather than treating those as separate concerns. The executives interviewed for the report describe a “knowledge layer” – an architectural concept that sits between raw data stores and the agents that consume them – as the most direct path to improving agent decision quality. Rather than expecting agents to derive context from raw data on their own, the knowledge layer encodes organizational meaning, relationships, and rules in a form agents can actually use.
Investment priorities among organizations working to close the knowledge gap concentrate in a few areas: ingestion pipelines that make data consistently available to agents; AI-ready APIs that expose organizational knowledge in structured formats; retrieval-augmented generation (RAG), which lets agents pull relevant context at inference time rather than depending on what’s baked into their training; AI evaluation agents that assess output quality; and knowledge graphs that map relationships between data entities in ways that support richer reasoning.
The Cost of Staying Stuck at 34%
The competitive dimension is hard to ignore. Organizations that cannot advance more than a third of their agent projects to production are watching competitors who can do better pull ahead on operational efficiency, automation capacity, and AI-derived decision quality. The investment already sunk into stalled projects doesn’t disappear – it sits as a liability on the books while the capability gap widens.
The report does not offer a simple fix, because there isn’t one. Improving agentic knowledge access requires coordinating data strategy, security policy, architectural investment, and organizational change simultaneously. What the data does confirm is that the organizations treating knowledge infrastructure as a first-class engineering priority – not an afterthought to model selection or agent design – are the ones that reach 61% production rates while the rest hover around 34%.
The real question may not be whether an organization can afford to invest in a knowledge layer, but whether it can afford to keep deploying agents without one – particularly as competitors already past the pilot stage begin compounding their advantages quarter by quarter.
