Enterprise AI agents are often presented as a model-selection problem: choose a stronger foundation model, add tools, and automate work. A new survey-based report argues that the more persistent constraint is organizational knowledge—the context that tells an agent what enterprise data means, which rules apply, and how work is actually done.
That distinction matters because an agent can retrieve a policy document, customer record or inventory figure without understanding its relationship to other systems, prior decisions, approval rules and user permissions. In that setting, an apparently capable agent can still make unreliable recommendations or take the wrong action.
The report, published by MIT Technology Review’s custom-content arm and linked to Neo4j research, surveyed 300 data, AI and technology executives. Its central finding: on average, 34% of organizations’ agentic AI projects reach production.
The production bottleneck is context
Respondents identified legacy data systems, security and privacy concerns, and inadequate knowledge and context as major reasons projects stall. Data fragmentation was the most commonly cited challenge to expanding agent access to knowledge, named by 55% of respondents.
The report separates the issue into three capabilities:
- **Semantic knowledge:** what data, entities and relationships mean in the organization.
- **Episodic memory:** relevant history from prior interactions and decisions.
- **Procedural knowledge:** the steps, rules and workflows required to complete a task.
This is a more demanding problem than placing documents into a retrieval-augmented generation (RAG) system. RAG can surface relevant text, but production agents also need authoritative sources, access controls, up-to-date definitions, traceable decision paths and clear boundaries on when they may act.
For example, an agent handling a customer exception may need to connect account terms, order status, prior service interactions, fulfillment constraints and escalation policy. A correct answer depends on both retrieval and the governed relationships among those sources.
A correlation worth operational attention
The survey identifies a group of “production leaders” whose agentic projects advance beyond pilot at an average rate of 61%. These organizations report stronger knowledge capabilities than other respondents, particularly in semantics.
The finding is correlational, not proof that a knowledge layer alone causes deployment success. The report is also sponsored content rather than independent editorial research. Still, its operational implication is sound: teams should treat knowledge infrastructure as a core dependency for agent deployment, alongside model quality, tool integration and evaluation.
Notably, production leaders were more likely to cite security and privacy as a major concern, at 72%. That may indicate that mature programs encounter the harder governance issues that appear once agents move closer to sensitive systems and real workflows.
What teams should build first
The most useful near-term move is not a broad attempt to model every piece of enterprise information. Start with a narrow, high-value workflow and identify the specific knowledge an agent needs to perform it safely.
That typically means:
1. Map authoritative data and ownership. Identify the system of record for each fact, who maintains it and how freshness will be checked. 2. Define business semantics. Standardize key entities, terms and relationships so the agent does not infer meaning from inconsistent labels. 3. Encode procedural constraints. Make approvals, eligibility rules, exception handling and escalation paths explicit. 4. Apply permissions at retrieval and action time. An agent should not gain broader access merely because it can query a connected system. 5. Evaluate with real cases. Test groundedness, policy compliance, correct tool use and harmful-action prevention—not just response fluency.
The report expects companies to invest in ingestion pipelines, AI-ready APIs, RAG, AI evaluation agents and knowledge graphs. These are complementary components, not interchangeable products. Pipelines and APIs improve access; retrieval finds relevant material; graphs can represent relationships; evaluation determines whether the combined system performs safely enough for production.
What to watch next
The critical measure for enterprise agent programs will be less the number of pilots announced than the share that can operate reliably in governed workflows. As agents gain access to systems of record and the ability to take actions, the quality of an organization’s data definitions, workflow logic and permission model will increasingly determine their usefulness.
For leaders, the question is therefore not simply whether the company has enough data for AI. It is whether that data has been made intelligible, connected and governable enough for an agent to use without creating new operational risk.
