Why Most AI Projects Fail – And How Companies Can Avoid the Trap
Recent headlines have painted a sobering picture of artificial intelligence (AI) adoption. A widely circulated MIT Media Lab/Project NANDA report claims that ‘95%’ of investments in generative AI (GenAI) have produced zero returns. For executives and business leaders, the statistic raises critical questions: Is AI just hype? Or are most companies simply using it the wrong way?
While I am skeptical of the methodolgy and the definition of success of the MIT NANDA’s study, it corrobrate directionally other industy research and my own interviews with senior finance colleauges in Pharma and Medical Device.
AI itself is not flawed—the issues arise from how businesses approach it. There are repeated patterns of failures that prevent companies from turning AI pilots into scalable, profitable systems.
This blog explores the major causes of failed AI adoption, and the strategies any organization can use to extract genuine business value from GenAI.
The "AI Experimentation Trap": Good Intentions, Misguided Execution
One of the most common reasons AI use cases flop is unfocused experimentation. Much like the digital transformation rush of the 2010s, many leaders today encourage teams to "experiment with AI" without providing direction, priorities, or resource structures.
The idea is simple—let’s try 10,000 small projects and hope one becomes the next unicorn. But in practice, this leads to:
A morass of pilots disconnected from value creation or operational needs.
Fragmented teams, that have scarced time, resources, and under-funded.
Dozens of proofs-of-concept that never evolve into full deployments.
Inevitably, these leaders conclude that "AI doesn’t work," when in reality, the problem was unfocused experimentation.
The Core Issue: No Link to Business Value
AI is often framed as disruptive and radical. But at its core, it’s simply one of many tools to solve problems for customers or to optimize business operations. The most successful AI projects are not flashy "cosmetic applications" for marketing and sales forecast—but back-end integrations that streamline workflows, financ cycles, enhance decision-making, or reduce costs.
How to fix this trap:
Treat AI as part of the larger digital transformation journey anchored to your strategy, not an isolated experiment.
Prioritise projects around specific customer or operational problems.
Narrow the focus based on ROI. Choose a handful of strategic initiatives instead of dozens of small experiments.
Scale aggressively when pilots show measurable value. Typically, the value is created in the 7th iteration and so setting expectations and upfront alignment on the definition of success and how to measure it are critical factors.
The "GenAI Divide": Why Friction Is Essential
Another overlooked factor in failed AI adoption is the belief that pilots should be seamless and frictionless. The value creation comes from friction and is necessary for long-term success.
What Is GenAI Friction?
Friction isn’t inefficiency for its own sake. Instead, it refers to the natural resistance that ensures systems, people, and workflows adapt properly. Without it, organizations end up with flashy demos that collapse during real-world deployment.
Examples of GenAI friction include:
Governance and compliance obligations.
Explicit discussions on ROI, trade offs, and risks
Redesigning workflows, not just layering AI tools onto old processes.
Conflicts across departments and incentives.
The challenge of unreliable outputs ("hallucinations") that must be verified.
Friction forces businesses to confront these realities before scaling.
The "Verification Tax" and the Humility Gap
One of the biggest friction points enterprises encounter is the verification tax—employees spend so much time correcting confidently wrong outputs that productivity gains evaporate.
This happens because most companies adopt generic tools like ChatGPT without customizing them for their operations. Without training on specific workflows, these tools don’t adapt, forcing humans to double-check everything.
The better approach? Humility-first design. Tools should:
Admit uncertainty when confidence is low.
Ask for human guidance instead of generating false answers.
Incorporate user corrections into a continuous improvement loop (abstain → correction → improvement). The general rule is the 7th iteration creates significant value
This “accuracy flywheel” builds organizational trust while steadily improving the system.
In short: Pilots that feel “too smooth” are probably too shallow. Real progress requires wrestling with friction.
Buying vs. Building: A Strategic Choice That Determines Outcomes
Purchasing AI tools versus building them in-house.
Purchased/vendor AI solutions: succeed ~67% of the time.
Internally built AI projects: succeed only ~33% of the time.
Why such a big gap?
The Case for Buying AI Tools
Vendors who specialize in AI can:
Provide pre-trained, benchmarked models that outperform open-source alternatives.
Handle compliance, monitoring, and scaling infrastructure.
Continuously update tools as the space evolves.
The Pitfalls of Building In-House
Enterprises often build in-house for reasons like privacy, compliance, or "control." While valid, these builds face major challenges:
Lack of deep AI expertise (too costly to staff in full).
Heavy reliance on open-source models, which may lag behind vendor products.
Slow, high-cost experimentation with little room for iteration.
The takeaway? Unless AI is core to your business model (e.g., being an AI company), partnering with vendors is both cheaper and more effective.
Experimentation Without a Scaling Path
Even when companies run AI pilots properly, another pitfall emerges: no scaling strategy.
Failure comes in two forms:
Leaders charge ahead with pilots, without asking how they’d scale across departments.
Or worse—they obsess about enterprise-readiness from Day 1, which creates bottlenecks and delays.
The "Goldilocks Zone" of AI Pilots
The best experiments are:
Rooted in real value creation (customer, cost, or workflow).
Low-cost enough to allow multiple test cycles.
Built with a clear roadmap for expansion if proven successful.
Frameworks like IFD (Intensity, Frequency, Density) help prioritize problems worthy of AI adoption:
Intensity: How painful is this issue for customers or internal stakeholders?
Frequency: How often does it occur?
Density: How many people struggle with this same issue?
Once a problem checks all three boxes, it’s a strong candidate for AI-enabled scaling.
Measurement Mistakes: Defining Success Too Narrowly
One critique of MIT’s study is its narrow success criteria:
Deployment beyond pilot.
Measurable ROI within six months.
This definition misses longer-term iterations or indirect value, such as:
Efficiency gains (saving employee hours).
Reduced customer churn.
Increased decision quality on insights conversion.
The Reality: ROI Often Appears Beyond Six Months
AI initiatives—like ERP systems, CRMs, or cloud migrations before them—take longer than six months to pay off. If companies cut funding prematurely because ROI isn’t immediate, they kill projects before the benefits compound.
The fix is clear: use multi-dimensional metrics of success. Track not just profit-and-loss (P&L), but operational efficiency, employee time savings, quality improvements, and customer experience enhancements.
Shadow AI: The Hidden Success Story
Interestingly, while official AI pilots often fail, employees themselves are already successfully adopting AI tools. Data shows widespread reliance on “shadow AI”—unsanctioned but highly useful tools like ChatGPT.
Employees use these tools to:
Draft reports and presentations faster.
Automate repetitive financial data cleanup tasks and analysis.
Reduce reliance on expensive external agencies.
This unsanctioned adoption is already saving enterprises millions annually. Yet most companies ignore or even discourage it.
Forward-looking organizations instead formalize "shadow AI" by:
Understanding where employees are already seeing gains.
Adopting governance to make these practices safe.
Scaling grassroots experiments into sanctioned initiatives.
The Path Forward: Building the Next-Generation AI Enterprise
The bigger picture here is that every company is becoming a technology company. AI isn’t an optional add-on—it’s part of a long shift toward digitally core businesses, where workflows, data, compliance, and customer service are reshaped at the foundation.
To thrive in this new era, leaders must:
Anchor AI strategy in value creation. Focus first on solving real problems customers care about.
Embrace friction. Treat resistance not as failure, but as a signal that adaptation and learning are underway.
Favor partnerships over solo builds. Leverage vendor expertise wherever possible.
Design scalable pilots. Ensure experiments are cheap enough to iterate yet tied to eventual rollouts.
Measure holistically. Look beyond six-month ROI reports—focus on long-term transformation metrics.
Empower employees. Channel shadow AI into safe, productive, and officially recognized workflows.
Conclusion: From AI Tryouts to Business Transformation
The ‘95%’ failure rate statistic sounds alarming, but it’s not a condemnation of AI. Instead, it’s a wake-up call to rethink how companies approach adoption.
AI projects fail not because the technology lacks potential, but because of:
Weak links to business value.
Aversion to friction.
Costly in-house builds.
Poor scaling strategies.
Narrow, short-term success metrics.
Successful AI finance projects treat AI not as a spectacle, but as part a strategic transformation tool as in the case of the finance function transformation.
Over the next decade, the organizations that design for friction, prioritize customer value, and build scalable AI strategies will rise far above competitors stuck in experimentation purgatory.