Just 34% of agentic AI projects make it into production, on average. That number comes from a new MIT Technology Review Insights survey of 300 data, AI, and technology executives. The report says the main problem isn’t the models. It’s the knowledge companies give their agents, or fail to give them.
One caveat up front: MIT Tech Review’s custom content arm produced this report, not its editorial staff. The numbers are self-reported by executives, so read them as a direction, not a precise measure. Still, the pattern lines up with what a lot of enterprise teams have seen over the past year.
📊 The numbers that matter
- 34%: the average share of agentic projects that get past the pilot stage
- 61%: the production rate among a small group of “production leaders”
- 55%: executives who name data fragmentation as a top barrier to giving agents more knowledge
- 72%: production leaders who call security and privacy a major concern
According to the report, even high-tech firms struggle with this. The three biggest points of failure are legacy data systems, security and privacy concerns, and a lack of knowledge and context.
🧠 The bottleneck is knowledge, not intelligence
The survey scores “agentic knowledge capabilities” in three areas:
- Semantic knowledge: what things mean inside your business
- Episodic memory: what happened before, like past interactions and decisions
- Procedural knowledge: how work actually gets done
Production leaders score higher across the board, and especially on semantics. MIT Tech Review reports that their advantage “tracks closely” with their higher production rate.
This matters because a capable model can still be useless if it doesn’t know that “active customer” means one thing in your CRM and something else in billing. Frontier models keep getting better at reasoning. That doesn’t help much when the agent is reasoning over scattered, contradictory data.
🔐 The leaders worry about different things
This is the most interesting finding in the report. Most companies are still fighting fragmentation, meaning their data doesn’t move cleanly between systems. Production leaders put security and privacy at the top instead.
My read is that this signals maturity. Once your data is connected, the hard question changes from “can the agent reach it?” to “should the agent see it, and on whose behalf?” An agent with wide access is a permissions problem waiting to happen. Teams that ship more agents run into that wall sooner.
🛠️ Where the money is going
Executives expect the biggest improvement in agent decisions to come from strengthening the connection between their data and their AI agents. The experts MIT Tech Review interviewed point to a “knowledge layer” as the best way to build it. Investment priorities include:
- Retrieval tech: ingestion pipelines, AI-ready APIs, and retrieval-augmented generation (RAG), which pulls relevant documents into a model’s context before it answers
- AI evaluation agents: automated checks on whether agents’ outputs are any good
- Knowledge graphs: structured maps of how business entities relate to each other
That fits where the wider market is heading. Data platforms, enterprise search vendors, and model labs are all pushing their own version of a context layer. Whoever owns that layer gets a lot of leverage over which agents a company can run.
✅ What practitioners should do now
- Pick one use case and map its data. List every system the agent has to touch before you write a single prompt.
- Fix definitions early. A shared business glossary or a basic ontology (an agreed map of your key terms and how they relate) will do more for you than swapping models.
- Design permissions from day one. The leaders’ focus on security tells you where you’ll end up anyway.
- Build evaluation into the pipeline. You can’t get out of a pilot if you can’t show the agent is reliable.
- Don’t expect RAG alone to fix it. Retrieval only helps if what’s being retrieved is clean and well defined.
The 34% figure should give pause to anyone promising fully autonomous enterprise agents in 2027. Over the next year, the winners will probably be the companies that do the unglamorous data work, not the ones with the flashiest demos. The full survey breakdown is available from MIT Technology Review.