AgentsAutonomy & tool use
Neo4j-sponsored survey says 34% of enterprise agent projects reach production and blames missing organizational knowledge
A Neo4j-sponsored MIT Technology Review Insights survey of 300 executives says about a third of agentic AI projects reach production. It blames stalled projects on missing organizational knowledge. Independent benchmarks find that graph retrieval, the sponsor's proposed fix, helps only on certain tasks.
AT Pulse research deskOctober 6, 2026 · 5 min read · AI-assisted analysisThe survey reports that on average 34% of agentic AI projects reach production, compared with 61% at 'production leaders.' It links the gap to stronger knowledge capabilities (semantic, episodic and procedural), especially semantic knowledge. The figures are self-reported, the link is a correlation, and the report is sponsored content from a graph-database vendor that recommends a graph-based 'knowledge layer.' [1] [7]
Independent academic benchmarks give a narrower picture of graph-based retrieval. GraphRAG does better on multi-hop reasoning, but plain RAG does better on single-hop, detail-oriented questions, and one benchmark finds GraphRAG often underperforms vanilla RAG on real-world tasks. Graphs cost more to build, results depend on graph quality, and routing queries between the two methods or combining them beats either one alone. [2] [3] [4]
Other sources also point to high attrition, but give different reasons. Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing cost, unclear business value and weak risk controls. A separate survey sponsored by Google Cloud found that AI can reach about 45% of company data on average and that only half of organizations trust their agents' decisions. [5] [6]
two vendor-sponsored surveys and a Gartner forecast all point to high attrition among agent projects.
Read the full assessment
Implication: data fragmentation and governance may matter more than model choice, but graph investments should be tested against plain RAG on real workloads.
Executive brief
According to a survey of 300 data, AI and technology executives, only about 34% of enterprise agentic AI projects reach production on average. The report says the main thing holding the rest back is a lack of organizational knowledge and context, not model quality (MIT Technology Review). Neo4j, a graph-database vendor, sponsored the report, and it recommends a "knowledge layer" that includes knowledge graphs. Independent academic benchmarks give a narrower picture: graph-based retrieval helps with some tasks and often does worse than plain RAG on others (Xiang et al.). Treat the survey as directional evidence from a sponsor with a commercial interest in the answer.
What changed and event timeline
RAG vs. GraphRAG systematic evaluation
Han et al. found the two methods have complementary strengths. GraphRAG wins on multi-hop reasoning. The latest revision is dated March 2026 ().
GraphRAG-Bench asks "when to use graphs."
The paper reports that GraphRAG frequently underperforms vanilla RAG on real-world tasks and sets out to identify when graphs actually help. It was later accepted at ICLR 2026 (;).
Gartner forecasts agent project cancellations
Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027. It cites rising costs, unclear business value and weak risk controls ().
Google Cloud-sponsored survey on agent data
An earlier MIT Technology Review Insights survey of 300 executives found that AI can reach on average 45% of company data, and only half of organizations trust their agents' decisions ().
Neo4j-sponsored "knowledge layer" report
This report finds 34% of agentic projects reach production, compared with 61% at "production leaders." It ties the gap to stronger knowledge capabilities, especially semantic knowledge (;).
Capabilities and access
- This is a survey report, not a product or model release. No model or version is named.
- The full report is available as a download through MIT Technology Review and Neo4j's whitepaper page.
- It scores organizations on three kinds of "agentic knowledge capability":
Read the full section
- This is a survey report, not a product or model release. No model or version is named.
- The full report is available as a download through MIT Technology Review and Neo4j's whitepaper page.
- It scores organizations on three kinds of "agentic knowledge capability":
- Semantic knowledge: what the data means in the organization's context.
- Episodic memory: a record of past events and interactions.
- Procedural knowledge: how tasks are done.
- Neo4j presents its graph technology as the way to build the recommended knowledge layer (Neo4j).
Technical analysis for researchers and developers
- Survey method: The published pages give the sample size (300) but no fieldwork dates, industry mix or definition of "production" (Neo4j).
- GraphRAG trade-offs: Han et al. report that GraphRAG costs much more to build. The quality of the graph strongly shapes results.
- Evaluation caveat: When LLMs are used as judges, they show strong position effects (the order in which answers are shown sways their preference).
Read the full section
- Survey method: The published pages give the sample size (300) but no fieldwork dates, industry mix or definition of "production" (Neo4j). The link between knowledge capability and production rate is a correlation, not a cause.
- GraphRAG trade-offs: Han et al. report that GraphRAG costs much more to build. The quality of the graph strongly shapes results. Routing queries to the right method, or running both in parallel, consistently beats either one alone (arXiv).
- Evaluation caveat: When LLMs are used as judges, they show strong position effects (the order in which answers are shown sways their preference). This can distort GraphRAG comparisons (arXiv).
- Reproducibility: GraphRAG-Bench publishes its code, datasets and leaderboards (GitHub).
Claims and evidence
- 34% of projects reach production; 61% for leaders. Sponsor-backed survey, self-reported (MIT Technology Review). No independent confirmation was found.
- Data fragmentation is the top barrier (55%); 72% of leaders cite security and privacy.
- A knowledge layer or graph is the fix. This is a recommendation from the sponsor and the experts it interviewed.
Read the full section
- 34% of projects reach production; 61% for leaders. Sponsor-backed survey, self-reported (MIT Technology Review). No independent confirmation was found.
- Data fragmentation is the top barrier (55%); 72% of leaders cite security and privacy. Same survey, same caveats.
- A knowledge layer or graph is the fix. This is a recommendation from the sponsor and the experts it interviewed. Independent benchmarks support graphs only for multi-hop and reasoning-heavy tasks (arXiv; arXiv).
- Data gaps limit agents. A separate MIT Technology Review Insights survey, sponsored by Google Cloud, points the same way. It is also vendor-sponsored, so it is not independent confirmation (MIT Technology Review).
Context and prior work
- Gartner's cancellation forecast puts more weight on cost, unclear value and risk controls than on missing knowledge (Gartner).
- Two MIT Technology Review Insights reports have now been sponsored by vendors whose products match their conclusions. Neo4j's stresses knowledge graphs (August report; October report).
- The academic GraphRAG literature tests retrieval quality on benchmarks. It does not measure whether enterprise deployments succeed.
Read the full section
- Gartner's cancellation forecast puts more weight on cost, unclear value and risk controls than on missing knowledge (Gartner).
- Two MIT Technology Review Insights reports have now been sponsored by vendors whose products match their conclusions. Google Cloud's report stresses data access and governance. Neo4j's stresses knowledge graphs (August report; October report).
- The academic GraphRAG literature tests retrieval quality on benchmarks. It does not measure whether enterprise deployments succeed.
Limitations, safety and contested findings
- MIT Technology Review labels the piece sponsored content produced by Insights, its custom content arm, not by its editorial staff (MIT Technology Review).
- The production rates are self-reported averages. No sampling details are public.
- One benchmark reports that GraphRAG often underperforms vanilla RAG (arXiv). That undercuts any general claim that graphs are the main fix.
Read the full section
- MIT Technology Review labels the piece sponsored content produced by Insights, its custom content arm, not by its editorial staff (MIT Technology Review).
- The production rates are self-reported averages. No sampling details are public.
- One benchmark reports that GraphRAG often underperforms vanilla RAG (arXiv). That undercuts any general claim that graphs are the main fix.
- Production leaders worry more about security and privacy (72%). Giving agents wider access to knowledge also widens their exposure to data.
Business and practitioner implications
- Two separate sources point to high attrition: 34% reaching production in the survey, and more than 40% forecast to be cancelled by Gartner.
- These are cited as barriers in both sponsored surveys.
- Use graphs selectively. Budget for the cost of building them, and consider routing queries between plain RAG and GraphRAG.
Read the full section
- Expect most pilots to stall. Two separate sources point to high attrition: 34% reaching production in the survey, and more than 40% forecast to be cancelled by Gartner.
- Fix data fragmentation and governance first. These are cited as barriers in both sponsored surveys.
- Use graphs selectively. Fit them to multi-hop, relationship-heavy questions. Budget for the cost of building them, and consider routing queries between plain RAG and GraphRAG.
- Test on your own data. Run head-to-head evaluations on your own workloads, and control for LLM-judge bias, before buying "knowledge layer" products.
Sources
Read the full section
- Connecting AI agents to enterprise knowledge — MIT Technology Review Insights
- MIT report: Why AI agent projects stall before production — Neo4j
- Scaling AI agents with trustworthy data — MIT Technology Review Insights
- Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
- When to use Graphs in RAG (arXiv 2506.05690)
- GraphRAG-Benchmark repository
- RAG vs. GraphRAG: A Systematic Evaluation and Key Insights (arXiv 2502.11371)