New MIT study says most AI projects are doomed...

A new MIT study suggests the biggest problem in enterprise AI isn’t model quality—it’s execution. After analyzing 300 public AI deployments, interviewing 150 leaders, and surveying 350 employees tied to recent AI integrations, the study found that 95% of AI-driven projects failed to deliver rapid revenue acceleration. The result: most initiatives produced little to no measurable bottom-line impact, fueling fresh skepticism among investors who have been betting on AI to sustain market momentum.

The study also points to a pattern in how companies implement AI. Organizations that tried to build their own AI tooling saw higher failure rates than firms that bought from third parties. The implication is blunt: “build it yourself” often leads to worse outcomes than adopting existing tools—especially when teams lack the operational know-how to integrate AI into real workflows.

That execution gap shows up in the kinds of breakdowns described. Failures weren’t blamed on AI models being “too dumb.” Instead, the integration process ran into brittle workflows, insufficient context, and misalignment with day-to-day operations. In other words, even capable models can underperform when they’re bolted onto processes that don’t provide the right inputs, guardrails, and feedback loops.

The transcript connects these findings to broader market anxiety. It references a wave of Silicon Valley talk about an “AI bubble,” and notes that investor confidence has been shaken by the 95% failure figure. The discussion also ties the hype cycle to hiring and spending behavior—citing Mark Zuckerberg’s reported freeze on AI hiring at Meta shortly after aggressive AI talent moves. The underlying question becomes whether investors are overexcited about AI returns relative to what deployments actually achieve.

Still, the narrative doesn’t end in total pessimism. It highlights a counterexample: Ignite, an enterprise software company whose CEO Eric Vaughn reportedly replaced 80% of developers with AI in 2023 and later claimed the move produced 75% profit margins with no regrets. That success story is used to reinforce the study’s central takeaway: AI can generate value, but the difference between failure and payoff often comes down to how teams deploy it—workflow design, context, and operational fit—rather than whether the underlying model is impressive.

The transcript ultimately frames the moment as a transition period. With many projects failing to translate AI into revenue, it argues that enterprises may still need human programmers for the foreseeable future, while vendors and integrators that can reduce integration risk stand to benefit. The “crack-like” analogy for AI coding—initial productivity spikes followed by escalating errors—serves as a warning about overreliance and the need for disciplined engineering practices even when AI accelerates drafting code.