Why 72% of AI Projects Miss ROI Targets—and How to Fix It

A recent Gartner survey delivered a sobering reality check for technology leaders: in IT infrastructure and operations (I&O), only 28% of AI projects met their expected ROI, while 20% failed entirely. The remaining half sit in an unresolved gray zone—neither generating real returns nor formally canceled.

The study surveyed 782 I&O leaders, 57% of whom admitted to experiencing at least one outright AI project failure. Meanwhile, global AI spending is projected to surpass $2.5 trillion in 2026—a 44% surge over the previous year. Enterprise capital is deploying faster than ever, but clear returns remain elusive.

The Failure Isn't the Tech—It's Runaway Scope

Melanie Freeze, Research Director at Gartner, summarized the root cause plainly: most failed initiatives suffer from inflated expectations and overly broad scope. Leaders expect AI to instantly automate complex processes, slash overhead, and resolve years of technical debt. When results do not materialize overnight, executive confidence collapses and projects stall.

The highest failure rates concentrate in areas like auto-remediation, self-healing infrastructure, and agent-driven cross-system workflow automation. These happen to be the most heavily marketed AI promises—and the most operationally complex.

By contrast, more mature environments like IT Service Management (ITSM) and cloud operations report AI success rates of 53%. When workflows are standardized, use cases are clearly bounded, and data quality is controlled, AI consistently delivers on its promises.

Across hundreds of enterprise implementations, the pattern holds: AI failure rates scale directly with project ambition and inversely with use-case maturity.

98% of Boards Demand ROI, While 71% of Budgets Face Cuts

Gartner's findings reflect a broader trend across enterprise tech. A February survey by The Harris Poll revealed that 98% of technology executives face intense board pressure regarding AI ROI. Furthermore, 71% warned that their AI budgets will face freezes or cuts if they fail to show measurable returns within the first half of the year.

Another study of nearly 6,000 enterprise executives across the US, UK, Germany, and Australia painted a similar picture: over 80% reported seeing zero tangible impact on employment or productivity, despite 69% of those companies actively using some form of AI.

This gap does not signal an "AI winter." Instead, it marks an overdue calibration phase. Enterprises are shifting from the vague assumption that "AI can do everything" to the disciplined question of "what does AI actually do well today?"

The Two Core Bottlenecks: Skills and Data

Among I&O leaders who experienced project failures, 38% blamed ongoing skill gaps, and another 38% pointed to poor data quality or availability. These two obstacles tied as the leading causes of project failure.

The data challenge is rarely an absence of information; it is fragmented, inconsistent, and unstructured data spread across siloed systems. Advanced foundation models cannot produce reliable outputs when fed low-quality inputs. Similarly, the talent gap is rarely about learning how to prompt a tool—it is about lacking the methodology to map model capabilities to specific, repeatable business workflows.

The Starbox AI View: How the 28% Win

Across our enterprise deployments at Starbox AI, we see a clear pattern: successful AI projects are rarely the largest; they are the most targeted.

Gartner's data validates this principle. ITSM succeeds because its operational boundaries are clean, its workflows are structured, and its results are directly measurable. We see the exact same dynamic in content marketing—one of the fastest paths to an early enterprise AI win. Marketing workflows have clear inputs and outputs, and performance can be tracked across each stage of the customer journey without the systemic chaos of autonomous infrastructure.

To address the twin bottlenecks of data and skills, we rely on a four-dimensional knowledge architecture: 1. Pre-trained knowledge provides broad contextual reasoning. 2. Industry knowledge adds vertical expertise. 3. Enterprise proprietary knowledge grounds the system in internal business reality. 4. Real-time data keeps outputs accurate and current.

Rather than forcing every internal team member to become an AI engineer, skill gaps are solved by embedding expert methodologies directly into specialized agents.

When 72% of AI projects fail to return their investment, the technology is rarely at fault—teams are simply fighting on the wrong battlefield. Securing early, measurable wins in bounded operational areas before expanding into autonomous complexity remains the most reliable strategy for enterprise AI adoption.

As Gartner suggests, manage AI initiatives like discrete products: assign clear ownership, set measurable criteria, and maintain cross-functional accountability. Capital is already moving toward AI; the only real question is where to deploy it, how to measure it, and who owns the outcome.