Most AI projects quietly go nowhere. Here's what I keep seeing.
Nina Brenes··6 min read
- MIT's Project NANDA found 95% of enterprise generative-AI pilots deliver no measurable return (2025).
- The cause is organizational, not technical. It is how AI gets implemented, not which model is used.
- An agent is a demo. An operation is a workflow that actually runs, with your standard inside it and someone who owns whether it works.
- Over half of AI budgets go to sales and marketing, while the real return hides in the back office nobody automates.
You saw the demo. The agent answered on its own, sounded brilliant, promised to save you hours. You bought it. Three months later nobody uses it and you are not quite sure why. You are not alone, and it is not your fault, or the model's.
The uncomfortable number: 95%
MIT's Project NANDA study (2025) analyzed 300 enterprise AI deployments, interviewed 150 leaders, and surveyed 350 employees. The finding: 95% of generative-AI pilots produce no measurable return. And the reason is not technical. It is organizational, a problem of how AI gets implemented, not which model gets used.
An agent is not an operation
An agent is a demo: a single piece that does something impressive in isolation. An operation is a workflow that actually runs, with your business's standard built into it, with review, with someone who owns whether it works. The demo sells you the moment. The operation gives you the result, every Tuesday, whoever is in the room.
Where the return actually hides
The same MIT study found something revealing: more than half of AI budgets go to sales and marketing, but the real return sits in the back office, in operations and finance, where almost no one automates. Companies buy AI for what is visible, not for what pays.
The way out is not a better agent
It is to stop buying agents and start building an operation: one workflow, scoped honestly, with governance from day one and a human where it matters. That is the whole point of AI operations, and it is what separates the 5% from everyone else.
FAQ
Why do AI pilot projects fail?
Mostly for organizational reasons, not technical ones. MIT's 2025 study found 95% deliver no measurable return because companies deploy isolated agents instead of building governed workflows that people actually adopt.
What is the difference between an AI agent and an AI system?
An agent performs a task, often impressively, in isolation. A system, or operation, is the workflow around it: the standard, the review, the ownership, the governance that make the output reliable enough to run the business on.
Why does my company use AI but see no results?
Usually because AI is being used as scattered tools, not as an operation. Tools get abandoned; operations compound. The fix is to turn one recurring, high-friction process into a governed workflow with a clear owner.
What does it take to move an AI pilot into production?
A written playbook of what good looks like, review loops, governance, adoption by the people who do the work, and one accountable owner. Without those, a pilot stays a demo no matter how good the model is.
- MIT report: 95% of generative AI pilots at companies are failing — MIT Project NANDA, via Fortune
- From Potential to Productivity: Latin America in the Intelligent Age — McKinsey & World Economic Forum
Nina Brenes
AI partner for purpose-led, human-first founders
Certified AI Operator (The AI Exchange) and Anthropic Claude Partner. Nina spends her own time in AI conferences and daily practice so the founders she works with don't have to.
Certified AI Operator · Anthropic Claude Partner
Connect on LinkedIn