95% of AI Projects Fail — Stay on Track and Measure Value Added
I came across this Fortune article published in August 2025, which reports the results of an MIT study titled GenAI Divide: State of AI in Business 2025. The statistic struck me.
95% of enterprise generative AI pilot projects deliver no measurable impact on organizational benefit.
I’m shocked. Not because the number teaches me something fundamentally new about AI itself, but because it confirms, with solid data, what many practitioners have felt for 18 months without daring to say it out loud. AI doesn’t fix everything.
What the Study Says
The MIT NANDA report draws on 150 executive interviews, a survey of 350 employees, and analysis of 300 public generative AI deployments. The methodology is solid.
The main finding is brutal:
- 5% of projects actually accelerate revenue.
- 95% stall, with no measurable effect.
And the most important point, worth emphasizing: the cause of failure is NOT the quality of AI models. They’re the same GPT, Claude, and Gemini everyone has. The problem is elsewhere.
The Real Cause According to MIT
The report speaks of a learning gap — a gap that affects both the tool AND the organization. In plain terms:
- The tool doesn’t learn from your specific context. It doesn’t know what YOU do, what YOUR client expects, or how YOUR processes work.
- The organization doesn’t learn from the tool. It buys it, turns it on, and moves on.
This is exactly the same problem we see with certifications, security policies, or business continuity programs. Install the solution. Check the box. Nobody comes back to verify whether it actually works.
Buying a tool has never replaced the discipline of using it.
Three Numbers That Should Disturb You
The report contains figures worth reading slowly.
1. Buying an AI solution from a specialized vendor and building a real partnership succeeds about 67% of the time. Developing in-house without dedicated expertise succeeds only 33% of the time. Half as often.
2. More than half of enterprise AI budgets go to sales and marketing. Yet measurable ROI is found in less glamorous areas: administrative automation, eliminating outsourcing, operations optimization.
3. Success is not a question of tool. It’s a question of feedback loop.
Staying on Track
Here’s where I want to go. The lesson of this study is that an AI project is not an event. It’s a process you must FOLLOW from start to finish.
Concretely, that means asking three governance questions from the outset.
- What value added are we targeting? Not “we want to do AI.” Rather: “we want to reduce customer service request processing time by 30%.” A measurable, dated, assigned target.
- How will we measure that value? An indicator defined BEFORE deployment, not after. With a baseline for comparison.
- Who is responsible for follow-up and at what frequency? A named executive, quarterly reviews at minimum, a clear decision to continue or stop.
If you launch an AI project without a clear answer to these three questions, you’re in the 95% pool. Not the 5%.
The Parallel With Information Security
I’ll put on my usual hat here. This study faithfully reproduces what I’ve observed for years in security programs.
Organizations buy firewalls, antivirus, SIEM tools, ISO certifications. They never verify whether these investments produce the expected benefits. Policies are written, filed in SharePoint, then forgotten. Indicators aren’t kept up to date. Management reviews are done by intuition.
AI falls into the same trap. Faster, because media pressure is stronger. But it’s the same trap.
A program that isn’t measured is a program that doesn’t improve.
What to Remember
Before launching the next AI project in your organization, ask yourself one question.
Am I ready to measure whether it works, and to stop the project if it doesn’t?
If the answer is no, you just funded a PowerPoint.
AI is a powerful tool. Like any tool, its value depends on what you do with it, the discipline you apply, and the rigour with which you validate results.
Staying on track is exactly that.
Sources