YC Partners Debunk the '95% of AI Projects Fail' Study

Why YC Partners Think the MIT Study Favors Startups

This is the Light Cone (YC partners Jared, Gary, Diana) dismantling the viral "95% of AI projects fail" narrative. The MIT study got weaponized by AI doomers, but actually confirms everything YC has been saying about startup opportunities.

"Apple can't make a good calendar app." Infinite resources, infinite access to smart people, and the calendar app is still garbage. If that's true for Apple, how could internal IT, Deloitte, or Ernst & Young build sophisticated AI systems? Two-thirds of the projects in the study were internal builds or consulting agency work - and that's where the failures concentrated.

The real finding: outside vendors succeed more often. When enterprises bought products from startups (like Greenlight, Tactile, Castle AI), success rates were dramatically higher than internal projects. The study proves startups have a structural advantage, not that AI doesn't work.

"Engineering teams at these orgs don't believe in AI." This is Jared's provocative observation that triggered angry emails. Large companies would prefer to buy from established vendors, but those vendors can't build the products because their engineers think AI is overhyped, don't use codegen tools, and retweet MIT studies as confirmation. If your engineers don't believe, you can't build products that work.

The knock-on effect: startups get shots they never had. Enterprises can't build internally, can't go to established companies. The only option is startups. YC has examples: Castle AI closed a major bank by winning a bakeoff against an incumbent whose "AI" was just slapped on top. Reduct closed a FAANG company 154 days after Demo Day.

"We've been working on this internally for years." That's what enterprises say before the startup wins. The internal IT system took years and millions; the startup did it in months with a REST API.

10 Insights From YC on Enterprise AI Adoption

• Two-thirds were internal/consulting builds - That's where failures concentrated
• Outside vendors succeed more - Startups structurally advantaged
• "Grumpy engineers" - Incumbents can't build AI products because engineers don't believe
• Apple can't make calendar - Internal IT has always been bad; not AI-specific
• Castle AI won bakeoff - Incumbent's AI was "slapped on top"; Castle was AI-native
• Reduct: FAANG in 154 days - Found through YC launch post
• Champions live vicariously - Find employees who dreamed of startups but didn't do it
• Acquired founder champions - YC founders acquired by big cos help startups get in
• "Switching costs prohibitive" - $5B fintech CIO quote; that's the moat
• If you're really good, you can be the 5% - Not random; smart founders win

What This Means for AI Startups Selling to Enterprises

The "95% of AI projects fail" narrative masks a massive opportunity: most failures are internal IT and consulting projects, not startup products. Enterprises can't build AI internally, incumbents' engineers don't believe - leaving a vacuum that only startups can fill.