Connecting AI agents to enterprise knowledge
MIT Technology Review MIT Technology Review Insights ● Covered by 15 sources
Enterprise AI agents keep stalling because they don’t really know the company’s data. A survey of 300 execs says only 34% of projects reach production.
Based on reporting by MIT Technology Review, MIT Technology Review Insights — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AI agents can crunch data all day and still miss the point. MIT Technology Review’s latest report says that’s the core problem inside enterprises: these systems often lack the knowledge needed to understand what the data actually means in a specific business context. Without that layer of context, agents are shaky at reasoning, decision-making, and action-taking. And shaky agents do not get shipped.
The report is based on a survey of 300 data, AI, and technology executives. It breaks knowledge into three parts: semantic knowledge, episodic memory, and procedural knowledge. However you slice it, the finding is blunt. On average, only 34% of agentic AI projects make it into production. Even firms that are supposed to be good at this are struggling.
There is a split between the average company and a small set of production leaders. Those leaders get 61% of their agentic projects beyond pilot on average, and they also show stronger knowledge capabilities, especially around semantics. The report ties that advantage to better production outcomes. In other words: the companies that know more about their own data are the ones getting agents to work.
The biggest obstacle is not a lack of ambition. It is fragmentation. Data sharing across systems is inadequate, and 55% of respondents cite that as a top barrier to expanding agents’ access to knowledge. Security and privacy concerns also loom large, especially for the production leaders, where 72% call them a major concern. Legacy data systems show up too, along with the more embarrassing issue that the agent simply lacks the context it needs.
Most organizations are trying to build a stronger link between their data and their agents. The report says executives think the biggest payoff will come from reinforcing the structural foundation between the two, with experts pointing to a knowledge layer as the most promising approach. The spending plans sound practical rather than flashy: retrieval tools such as ingestion pipelines, AI-ready APIs, retrieval-augmented generation, AI evaluation agents, and knowledge graphs. That is a lot less glamorous than the agent hype cycle, but probably a lot closer to what gets systems into production.
My take — AI-written commentary, not fact-checked reporting
This is the part of enterprise AI that the hype crowd hates: the model is not the hero, the wiring is. Open or closed, it does not matter much if the company cannot connect its own data without turning security teams into full-time firefighters. Europe’s endless paperwork at least has one useful side effect: it keeps people from pretending a half-connected agent is ready for production.
Read more about this at: MIT Technology Review