Future Alpha 2026 – 85% of AI Projects Fail, And It Has Nothing To Do With The Technology

John Thomas Foxworthy opened his presentation at Future Alpha 2026 with a history lesson that doubled as a warning. The term “artificial intelligence” was coined in 1956 by John McCarthy at Stanford. McCarthy needed funding. The federal money for his research had dried up, so he invented a term that sounded ambitious enough to attract new dollars. Seventy years later, Foxworthy argued, that pattern has repeated with every generation of AI branding.

Foxworthy, who is founder of the Global Institute of Data Science (sponsored by Carnegie Mellon University) and teaches at Caltech and UC San Diego, walked through the history of AI. The perceptron (arguably the first artificial intelligence) arrived in 1957. Machine learning was coined in 1959 for a checkers program. Expert systems came in the 1980s. Statistical learning became machine learning in the 1990s when someone dropped “statistical” and replaced it with “machine.” The word “machine” means software program. The word “learning” means estimation. “So it’s a software program that estimates,” Foxworthy said. “That’s it.”

Deep learning, which won Geoffrey Hinton a Nobel Prize in Physics in 2024, is based on work from 1981: linear models stacked on top of each other. Convolutional neural networks date to 1989. LSTMs to 1995. Data science was officially coined in 2001 by William S. Cleveland at Princeton, in a paper that was, once again, asking for funding.

“If I look at the variation of the math we’ve been using for 70 years, it’s about this much,” Foxworthy said, holding his fingers close together. “If I look at the variation of the labels we call the same thing in a different way, it’s about this much,” spreading his arms wide.

The latest label is “agentic.” Foxworthy noted that agentic used to refer specifically to reinforcement learning. Now it refers to everything. “Agentic is kind of like the word artistic,” he said. “It’s referring to the quality of the agent. Just like artistic refers to the quality of an artist.”

The 85% Failure Rate

Foxworthy returned to one stat repeatedly: 85% of AI projects fail. This is not one study or one consultancy’s opinion. Four institutions with nothing in common, Rand Corporation, Gartner, McKinsey, and MIT Sloan School of Management, arrived at roughly the same number over nine years of data collection. McKinsey’s 2024 report alone covered 1,353 companies. The failure rate for AI across all studies ranges from 70 to 95%.

An audience member asked the obvious question: is it a technology problem?

“It is not a tech issue,” Foxworthy said. “All these organizations asked. Year after year after year. And the answer is no, no, and no.”