Roughly two out of three agentic AI projects never make it out of the pilot phase. That's the headline finding of a new survey of 300 data, AI, and technology executives, which puts the average share of agentic projects reaching production at just 34%.
The report, produced by Insights, MIT Technology Review's custom content arm — not its editorial staff — points to legacy data systems, security and privacy concerns, and a lack of knowledge and context as the main reasons projects stall.
What separates the winners from the graveyard
The survey splits respondents into two camps. At organizations it labels production leaders, an average of 61% of agentic projects make it past pilot. Those companies lean harder on knowledge capabilities, particularly semantics — the layer that gives an AI agent enough context to act without constant human hand-holding.
📊 Market Data Snapshot
Everyone else is stuck. Data fragmentation was the most commonly cited obstacle to expanding agents' access to knowledge, named by 55% of respondents. It's a familiar problem dressed up in new packaging: the models got good faster than the plumbing underneath them.
Security and privacy show up differently depending on where you sit. Among production leaders, 72% flagged those concerns as major — a higher share than the broader pool. That's a counterintuitive result worth sitting with. The companies furthest along aren't less worried about giving autonomous agents access to internal systems; they're more worried, and apparently better at managing it.
Where the money is going instead
Executives expect the biggest payoff to come from strengthening the structural link between organizational data and the AI agents sitting on top of it. The investment list that follows from that includes retrieval technologies — ingestion pipelines, AI-ready APIs, retrieval-augmented generation — plus AI evaluation agents and knowledge graphs.
None of that is glamorous. It's the unsexy middle layer that decides whether an agent can answer a question about last quarter's contracts without hallucinating, or route a support ticket using context a legacy system never captured in the first place.
The part most coverage will skip
It's worth flagging the provenance here. Insights is MIT Technology Review's custom content operation. That doesn't make the numbers wrong, but it does mean the findings arrived through a channel funded by sponsors with commercial interests in the AI tooling market. Read it as a vendor-adjacent signal, not a peer-reviewed study.
The 34% figure will get quoted widely anyway. It fits a narrative that's been building for months: the gap between agentic AI demos and agentic AI that actually runs a business process at scale.
Why crypto desks should care, eventually
There's no direct crypto angle in the survey — no tokens, no exchanges, no blockchain infrastructure named. The near-term read-through is close to zero. Bitcoin is trading around $85,417 with a slightly bullish tone and a Fear & Greed reading of 73, and this report isn't going to move that.
The second-order story is slower. If enterprises keep tripping over data fragmentation, provenance, and semantic context, demand grows for infrastructure that solves exactly those problems — and a chunk of that infrastructure is being built by crypto projects working on decentralized data marketplaces, verifiable data pipelines, and knowledge graphs. The same pain points cited by 55% of respondents are the ones those teams pitch against.
Whether enterprises actually reach for decentralized tools rather than another centralized vendor is unresolved. The survey's investment priorities lean toward retrieval pipelines and evaluation agents, not blockchain rails. For now, the more honest takeaway is that agentic AI's production gap is real, expensive, and not closing fast — and anyone betting on a near-term AI-and-crypto convergence should watch how the 61% clubs fix their data layer first.
