MIT Media Lab, Project NANDA · Report

The GenAI Divide: State of AI in Business 2025

The source of the famous “95% of GenAI pilots fail” headline, and worth reading for what it actually says.

Published
Reading time
About 30 min
Authors
Aditya Challapally, Chris Pease, Ramesh Raskar, Pradyumna Chari
Read the original

The original MIT link now redirects elsewhere; this is a widely used mirror of the PDF.

What is it about?

This preliminary report draws on 52 organizational interviews, 153 survey responses and a review of more than 300 public AI initiatives. It claims that only about 5% of organizations get measurable P&L value from task-specific generative AI tools despite $30–40 billion in enterprise spending, and blames a “learning gap”: most tools don't retain feedback, adapt to context or fit real workflows.

Its more useful findings are less quoted. Workers at more than 90% of surveyed companies use personal AI tools, while only 40% of companies bought an official subscription. Solutions from external partners reached deployment about twice as often as internal builds, and the clearest returns came from back-office work such as cutting outsourcing and agency spend.

Key numbers

  • 60% of organizations evaluated task-specific tools, 20% piloted them, 5% reached production with measurable impact.
  • Among organizations that actually piloted, roughly one in four succeeded.
  • External partnerships reached deployment about 67% of the time versus about 33% for internal builds (self-reported).
  • A lawyer at a firm with a $50,000 contract-analysis tool kept using ChatGPT because it “consistently produces better outputs”.

Why read it

  • You will be asked about “the MIT 95%”. This is how to answer precisely.
  • The shadow-AI finding is a ready argument for auditing what employees already use before buying more tools.
  • The buy-versus-build and back-office ROI sections are directly useful for portfolio decisions.

Who should read it

  • CIOs fielding board questions about AI failure rates
  • Procurement leaders deciding between buying and building
  • Anyone who has quoted the 95% figure

About the author: Aditya Challapally

Lead author; works on Microsoft Copilot; Stanford Online instructor

Challapally describes his path as a data-science internship at Uber, AI consulting at McKinsey, then Microsoft, where he works on Copilot agent capabilities. He co-teaches Stanford Online's course “Mastering Generative AI for Product Innovation”.

His advice to companies is refreshingly unglamorous: most organizations don't need anything beyond the basics. One irony worth noticing is that his own report places Copilot-style tools in the low-customization, low-learning corner.

Read it carefully: common misquotes

  • The 95% refers to all organizations in the sample, most of which never ran a custom pilot. It is not “95% of pilots”.
  • It is a self-published, non-peer-reviewed preliminary report, and some figures contradict each other inside the PDF.
  • Critics note that the authors work on agent frameworks of the kind the report recommends.

How it shapes the CIO's AI Playbook

Notes by Pavel Veselov, last reviewed . Summaries are my own; quotes and figures belong to the original authors.

All 11 sources in the CIO's AI Library