A new study from MIT has sent tech executives reeling after finding that despite tens of billions invested, the majority of AI pilots are failing to deliver any meaningful gains.

Coming out of the prestigious tech school’s Project Nanda initiative, which claims to be ‘building the foundational infrastructure for an Internet of AI Agents’, the State of AI in Business 2025 study found that 95% of organisations are getting zero return from their AI strategies.

After an investigation comprising more than 300 publicly disclosed AI programmes, interviews with 350 employees, and a survey of over 150 senior leaders across major industry, MIT’s team discovered that just a slim minority (5%) of AI pilots are extracting millions in value.

The vast majority of businesses, including enterprise firms which have invested heavily in the tech, are seeing no measurable impact on their bottom line, according to the study.

MIT found that one big problem in realising ROI on these AI initiatives is that most companies are unable to scale up their pilots. For example, although 80% of firms have explored the use of tools like ChatGPT and Copilot, only 40% report deployment.

The issue is even more evident for enterprise-grade systems, with just 5% of firms reaching the production stage despite 60% exploring the use of these bespoke tools.

Why are AI pilots stalling?

According to MIT’s study, the biggest problem lies in the functionality of the tools themselves. 

The overwhelming failure rate of enterprise AI pilots are put down to firms pursuing made-to-measure tools that can’t adapt to employee workflows, and compare poorly to generic LLMs and chatbots that workers generally consider more flexible, familiar and user-friendly.

Employees said that they trusted generic AI more (65%) and found that they received better answers (85%) than when using tools prescribed by their company, despite limitations like a lack of adaptability or the ability to remember context that systems like ChatGPT have suffered from until now. 

Recommended reading

• Report: Job Fears Driving Secret AI Adoption at Work
• AI Is Driving Data Breaches, IBM Finds
• Organisations Struggle to Cope with Shadow AI

Enterprise users in particular reported that their firm’s custom AI tools were brittle and overengineered, with one CIO telling researchers, “Maybe one or two are genuinely useful. The rest are wrappers or science projects.”

For complex, high-stakes or mission-critical projects, the vast majority of users (90%) said that they would still rely on a human, relegating AI to quick tasks like emails, summaries or basic analysis.

Along with comments from OpenAI CEO Sam Altman comparing the AI boom to the ‘90s dot-com bubble, MIT’s widely cited report has been linked by numerous outlets to a slump in tech stocks over the last week, with major firms like Nvidia and Palantir suffering drops.       

Likewise, Torsten Sløk, chief economist for investment firm Apollo Global Management, warned last month that the AI bubble is already fit to burst, with the top ten firms making up the S&P 500 ‘more overvalued than they were in the 1990s’ as a result of promises in an AI-driven future.