Big Tech has poured an estimated $400 billion into artificial intelligence. AI agents are the “it” factor of the year. Executives are heralding the recuperative (and productivity) powers of the technology.
And yet businesses that have adopted AI so far have little to show for it.
The starkest assessment comes from an MIT report this week: It found 95% of generative AI pilot programs at companies have little or no impact on the bottom line, based on an analysis of 300 public AI deployments, a survey of 350 employees, and 150 interviews with leaders.
At issue isn’t the quality of the AI models, but a “learning gap” for both tools and organizations – in particular, flawed enterprise integration. Widespread use of shadow AI, in the form of unsanctioned tools like ChatGPT, and the ongoing challenge of measuring AI’s impact on productivity and profit, also contributed to organizational obstacles, research discovered.
While ChatGPT excels for individuals because of their flexibility, it often stalls in enterprise use since they don’t learn from or adapt to workflows, Aditya Challapally, lead author of the report and research contributor to project NANDA at MIT, told Fortune.
Further, the report exposes a major disconnect in the allocation of resources. More than half of GenAI budgets go to sales and marketing tools when the greatest ROI comes from back-office automation, MIT found.
The high failure rate isn’t “a sign that GenAI doesn’t work, it’s a sign that most organizations are still learning how to make it work,” Mo Cherif, senior vice president of AI and innovation at Sitecore, said in an email.
The small sliver of companies that have successfully implemented GenAI tend to be run by under-20 founders who “pick one pain point, execute well, and partner smartly with companies who use their tools,” Challapally said. AI tools from specialized vendors and building partnerships succeed about two-thirds of the time, he said, as does empowering line managers.
The most advanced organizations, meanwhile, are already experimenting with agentic AI systems, he said. (
Nearly eight in 10 U.S. companies are adopting AI agents, though most solutions aren’t built for high-risk or regulated spaces like legal, privacy, and forensics, according to a PwC survey in May.)
Another report, from market researcher Gartner, predicts 40% of agentic AI projects will fail by 2027, potentially wasting millions in corporate investments. “Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied,” said Anushree Verman, senior director analyst at Gartner.
A January 2025 Gartner poll of 3,412 webinar attendees revealed 19% said their organization had made significant investments in agentic AI, 42% had made conservative investments, 8% made no investments, and the remaining 31% are taking a wait-and-see approach or are unsure.
Many vendors are engaging in so-called “agent washing” — the rebranding of current products such as AI assistants, robotic process automation (RPA), and chatbots without substantial agentic capabilities. Gartner estimates only about 130 of the thousands of agentic AI vendors are real.
“Companies have about 18 months to get their data house in order before the AI agent wave makes or breaks them,” Teradata Chief Technology Officer Louis Landry said in an email. “The truth is, almost anyone can build an AI agent, but when it comes to enterprise ROI, most whistle past the hard part: data integration. That’s where the real labor is, and without it, projects are bound to fail. The winners and losers will be defined by how well the data is organized and integrated.”
Empromptu.ai CEO Shanea Leven puts it succinctly. “Here’s the quiet part no one says out loud: AI fails in enterprise because most exec ideas aren’t solving real problems and the systems they’re plugging into are messy,” she said in an email. “Like mobile, only AI-native ideas will survive. And they’ll need accurate output teams can trust. Until we treat AI accuracy like uptime, most GenAI apps will stay stuck in the lab. The real unlock isn’t a bigger model, it’s better thinking and infrastructure that adapts to the real world.”
