Two out of three agentic AI projects never leave the pilot stage, a new survey finds. Across 300 large companies, the average agentic AI pilot to production rate was 34 percent, according to a report MIT Technology Review Insights published October 5.
New here? I publish one agentic AI governance post every weekday. Subscribe to the blog and it lands in your inbox the moment it goes live.
The survey behind the number
The survey ran in June and July. It reached top data, tech and AI chiefs at firms that earn at least $500 million a year. Even tech companies got only 42 percent of their agent projects live. Healthcare did worse at 27 percent, and consumer goods came last at 26 percent.
Still, some experts think the real number is lower. One analyst told the researchers it is “closer to the single digits.” Instead, a factory tech chief put it near 20 percent for firms that need exact answers.
Then there is a small group that does far better. The report calls them production leaders. They make up 13 percent of the sample. Their agentic AI pilot to production rate averages 61 percent. Everyone else averages 30 percent.
Why pilots stall
So what breaks the agentic AI pilot to production path? The execs named three main blockers. Old data systems topped the list at 55 percent. Then came security and privacy at 53 percent. Thin agent knowledge followed at 46 percent, and weak audit trails at 44 percent.
Here is where the leaders split off. For example, only 28 percent of them blamed thin knowledge, against 49 percent of the laggards. But 72 percent of leaders said security and privacy limit what their agents can see.
My read is simple. The laggards are still fighting their data. The leaders already won that fight, so now they face the real risk. That is a governance question, not a data question.
Two failures, told plainly
The report lists real failures. In April, a coding agent wiped a software firm’s production database in about nine seconds, Euronews reported. Car rental firms that ran on that software could not manage bookings for more than a day.
In another case, a health insurer’s agent denied post-acute care to elderly patients. In short, it overrode their doctors. A later review, summarized by Monte Carlo, found the model missed health problems the doctors saw right away.
Neither story is about a bad model. In both, the agent could take an action that nobody had put behind a gate.
Two teams that made it
A global drug maker took the slow road first. Its fraud team cleaned its data and pulled it into one place before the agent touched it, the team’s director told the researchers. Then it fed outside data into a knowledge graph. As a result, its agents can now trace whole rings of shell pharmacies, not just odd invoices.
An engine maker runs a different play. Its digital chief said the team runs each agent next to a human for four to six cycles. Only when the two outputs nearly match does the agent run alone. Even then, a human signs off on final approvals.
I would copy the second story first. It is a promotion gate with proof attached, and it costs almost nothing to set up.
More on Agentic AI Production
- Cost Per Task: Why I price agents by the cost of each completed task, not by the monthly bill.
- ROI Evidence: The deployments that paid back, in my roundup of agentic AI ROI case studies.
- The Eval Suite: How a nightly eval suite turns a demo into proof you can trust.
- Autonomy Tiers: Why I let agents earn autonomy in bands, one step at a time.
What Anthropic’s data adds
Anthropic’s study of how people really use agents points the same way. It found a human in the loop on 73 percent of tool calls. Only 0.8 percent of actions could not be undone. New users let the agent run alone about 20 percent of the time. After 750 sessions, that rose past 40 percent.
So autonomy grows with use, while trust builds. Nobody grants it on day one, and the firms that ship seem to know that.
What I would do before the next pilot
Every pilot should get a written exit gate before it starts. The gate names the business number, the eval suite, the audit trail and the shadow cycles. If the pilot cannot pass the gate, it ends early and cheap. That alone changes the agentic AI pilot to production math, because dead pilots stop eating budget.
A knowledge layer may help you pass the gate. It is not the gate, and I think the report blurs that line. In my work in regulated loan origination, the gate was the thing auditors asked to see first.
The 34 percent is not a model problem. It is a promotion problem. Instructions in, results out was IT; intent in, outcomes out is agentic AI. Outcomes need a gate, and the gate is what moves an agentic AI pilot to production.
How many of your agent pilots have a written exit gate today? Tell me in the comments.
Was this post useful? One click tells me what to write next.
Subscribe to my LinkedIn Newsletter
- Agentic AI Architect: control design for enterprise agents.
- Agentic AI Case Studies: deployment evidence, one teardown at a time.
- Agentic AI P&L: cost, payback, and risk in a CFO’s voice.
- Agentic AI SDLC: the lifecycle that builds the agents.
- Subscribe to the Blog: I publish a new post every weekday on one agentic AI governance question. Enter your email and each post arrives in your inbox the moment it goes live.
© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026
Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.
Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.
Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.