Why AI projects fail: The science experiment trap and how to avoid it

It’s easy to picture: A lone scientist locks themselves in a garage for weeks and then emerges with a discovery that changes the world. Or a mysterious, corporate skunkworks group who only show their faces every few months to present their latest ground-breaking invention.

Innovation in a vacuum—brilliant people working in isolation—has had its moments, and the idea certainly maintains a place in the public imagination

But I’ll tell you one situation where it does not work and in fact, can do more harm than good: enterprise AI adoption.

Attempting to implement enterprise AI transformation in a vacuum is guaranteed to fail. Excluding your strategic stakeholders, business unit leaders and collaborators ultimately means to neglect the perspectives and resources you need to succeed.

This approach might be why, according to the IBM Institute for Business Value’s (IBV) 2025 CEO Study, just 16% of AI initiatives have achieved scale at the enterprise level. A recent report from MIT’s NANDA initiative shared even grimmer findings—that 95% of generative AI pilots fail.

In too many cases, companies engage in multiple proofs of concept (POCs) that amount to little more than impractical science experiments. They might inspire awe at first (or FOMO, as IBV noted), but ultimately yield negligible value. As someone with a deep finance background, I know that organizations can do better—much better—than resigning themselves to such a meager return on investment.

AI initiatives that achieve scale can deliver impact beyond just one small part of the organization and achieve real return in the marketplace. But transcending silos doesn’t happen by accident. It requires alignment and support from management, including C-suite executives and even board members.

Leaders and managers can help orchestrate collaboration and systems that yield efficiency and increase impact. For example, when multiple departments work on their own distinct AI use cases, separate teams invest time in redundant efforts, from researching which AI models to use to developing governance programs. In contrast, when teams come together, they can combine resources and establish a unified approach that’s ripe for scaling, delivering more value to the greater organization.

One enterprise that’s excelled in this unified approach is PepsiCo. In recent years, PepsiCo has collaborated with IBM Consulting® to create a unified technology platform home to some 100 Generative AI use cases. We sat down with PepsiCo teams to map out their data and AI architecture, identify gaps, and create reusable services for their most important gen AI use cases.

The reusable services were made available through the unified technology platform, empowering teams across the enterprise with pre-approved models, tools and best practices. The platform also enabled centralized visibility, ensuring projects conformed with PepsiCo standards—something that ad hoc development through third-party vendors could not deliver.

With all these bases covered, PepsiCo teams accelerated experimentation, development and eventually, time to market. The results were AI-powered solutions ranging from hyper-personalized Gatorade bottle designs to optimized product positioning on retail shelves.

The right platforms and tools are critical to a scalable, unified approach to AI. But two other elements are arguably just as important: data and culture.

Data is the backbone of a successful AI program. It’s what AI models train on and it’s what enterprises use to determine whether a use case is worth pursuing. Yet experts estimate that less than 1% of enterprise data, so far, has been incorporated into AI models. That unused data represents a massive opportunity for enterprises.

Seizing it means to work on solidifying their data foundations—that is, ensuring their data is clean, organized and secure—as they pursue AI pilots. In the case of PepsiCo, the enterprise collaborated with IBM Consulting on a robust data strategy to manage its 60+ petabytes of data.

Closer to home, organizing data was a central component in IBM’s initiative to transform our own human resources with agentic AI. Training and fine-tuning models on domain-specific, high-quality datasets helped us develop a virtual agent, AskHR, that automated more than 80 HR tasks and now engages in 1.5 million employee conversations annually.

Employees can use AskHR to request employment verification letters, submit vacation requests and get important information on everything from sick leave to compensation. As a result, as of last year, support tickets and operational costs are down substantially, while our HR professionals have greater bandwidth to focus on strategic priorities.

Of course, AskHR wouldn’t have achieved its scale if our employees weren’t willing to use it. That’s true of just about every other enterprise AI initiative, too. Enterprise leaders must be careful to establish the right tone and culture around AI adoption.

That means being open about how AI augments employees’ work while also preparing them to reap those benefits. For the latter, education and skills-training programs are key; understanding how AI-powered tools work and how to best use them.

In many cases, managers learn right alongside their employees and that’s great—it reinforces the fact that successful enterprise AI adoption is a team sport. At the end of the day, we’re all learning, improving and pushing boundaries together. Skunkworks and silos need not apply.