Beyond the 95% headline

The MIT State of AI in Business 2025 report has been widely quoted for one shocking stat: 95% of generative AI pilots fail to deliver measurable impact on the P&L. On its own, the number suggests AI is all hype. 

But just like the Gartner study we explored in part 1 — where over 40% of agentic AI projects stall — the real lesson isn’t about failure. It’s about survivability. The MIT study shows us exactly where the ROI lives, and the conditions that separate collapse from scale.

Key takeaways from the MIT study include:

  • Most budgets are concentrated in sales and marketing pilots, but ROI is lowest there. The real returns lie elsewhere — in functions often overlooked.
  • Back-office automation produces the highest returns by streamlining processes, reducing outsourcing, and cutting costs.
  • Specialized vendor-led projects succeed ~67% of the time, while internal builds succeed only ~33%.

In other words, these genAI failures are avoidable.

Why most projects stumble

The MIT analysis reinforces what many practitioners have been observing for months:

  • Hype over hard work. Projects often emphasize flashy use cases rather than investing in fundamentals like observability, validation, and integration.
  • Garbage in, garbage out. Weak data quality and rigid processes can derail AI initiatives long before model performance comes into play.
  • Automation over AI theater. In many cases, workflow redesign and automation deliver more value than layering on a chatbot.
  • It’s not the tech, it’s the adoption. Without proper training, governance, and clearly defined ROI goals, even strong tools struggle to gain traction.
  • Build vs. buy realities.  Internal builds often underestimate the cost of integration and stall in pilots (~33% success rate), while specialized vendors succeed ~67% of the time by focusing on workflow fit and adoption.

GenAI doesn’t fail in the lab. It fails in the enterprise — when it collides with vague goals, poor data, and organizational inertia.

From failure modes to survival patterns

Projects that survive don’t just avoid mistakes — they’re built to overcome mistakes through domain focus and deep workflow fit. In Part 1 of this series, we introduced Trullion’s AI Survivability Matrix — a framework for mapping which projects endure and which collapse, based on two levers: domain specificity and workflow integration.

The new MIT report validates this lens with hard data. When we plot their findings onto the matrix, the patterns are clear:

  • Hype experiments (low specificity, low integration): home to the 95% failure rate. Flashy pilots that never escape the lab.
  • Generalist agents (low specificity, high integration): copilots bolted into workflows but without the financial or compliance nuance to deliver ROI.
  • Niche point solutions (high specificity, low integration): narrowly scoped tools that can’t scale beyond a single task.
  • Survivors & scalers (high specificity, high integration): vertical AI for back-office domains — deeply integrated, compliance-aware, and often delivered by specialized vendors.

This last quadrant is where MIT finds success rates nearly 2× higher than other approaches. It’s also where ROI lives — and where the real story behind the “5% that succeed” begins.

Why the 5% matters

It’s easy to fixate on the 95% of failures. But the lesson from MIT’s study is that the 5% that succeed matter far more. These survivors:

  • Cut costs in finance and compliance through automation
  • Replace outsourcing with in-house efficiencies
  • Deliver compliance-ready, audit-proof outputs
  • Scale across workflows, compounding ROI over time

Our point of view

Just as Gartner showed that nearly half of agentic AI projects stall, MIT states that 95% of AI pilots fail. AI doesn’t fail because the technology is flawed. Rather, it fails when it’s misapplied.

The winners aren’t generic copilots or one-off pilots. They’re domain-specific, workflow-integrated solutions in the back office, built and implemented by specialized partners who understand the stakes.

That’s where the practical ROI can be found — and where the few survivors will scale.