Connecting AI agents to enterprise knowledge

A report from MIT Technology Review Insights, the publication's custom content arm rather than its newsroom, surveyed 300 data, AI and technology executives. On average, only about a third of organizations' agentic AI projects, 34%, make it into production. Legacy data systems, security and privacy concerns, and a lack of knowledge and context are named as the main points of failure.

The report draws a distinction we find useful: data is not the same as knowledge. Knowledge is the understanding of what data means inside a specific organization. Without it, agents make flawed and unreliable decisions, and companies end up with expensive demos that never ship.

We think this framing is right, even with the caveat that it comes from a sponsored research product and not independent reporting. Anyone who has watched an agent pilot stall knows the pattern. The model can write, summarize and call tools just fine. What it can't do is know that two systems use different names for the same customer, or that a field everyone ignores is actually the one finance relies on. That is institutional knowledge, and it was never written down.

The most interesting finding is the gap between leaders and everyone else. A small group of organizations gets an average of 61% of agentic projects past the pilot, and they rate stronger on knowledge capabilities, especially semantics, meaning a shared understanding of what terms and data actually refer to. Correlation isn't causation, and we'd want to know more about how those leaders were defined. But the direction makes sense: agents work better when the company has done the boring work of describing itself.

There's a downside worth naming. The fix most executives point to is a knowledge layer, built from data ingestion and retrieval tools, AI-ready APIs, retrieval-augmented generation, evaluation agents and knowledge graphs. That means piping more of the business into systems agents can reach. Leaders in the survey were notably more likely to flag security and privacy as a major concern, cited by 72% of that group. We read that as a sign of maturity, not timidity. The teams furthest along are the ones who have seen what an agent with broad access can touch.

Where this leads if it scales: the winners in enterprise AI won't be whoever buys the smartest model. They'll be whoever has the cleanest, best-governed map of their own operations. That favors organized companies, and it quietly raises the cost of staying messy.

The survey measured agentic knowledge capabilities across three areas: semantic knowledge, episodic memory and procedural knowledge. Data fragmentation, meaning inadequate sharing of data across systems, was the most commonly cited top challenge to expanding agents' access to knowledge, at 55% of respondents.