AI seems to be everywhere these days, treated as unavoidable and inevitable. And of course, the technology is remarkable in many ways. With that said, all of its power and capability has yet to materialize into the promised revolution, as evinced in the recent MIT study that found a truly dismal 95% failure rate for AI pilot projects in the organizations it sampled.
AI is not a panacea. There are limits to what it can do for your organization
Fueled by vendor hype and media sensationalism, leaders often expect AI to deliver near-instant, transformative ROI. Leadership expects unrealistic cost reduction within unrealistically short time frames, or a flawless system from day one. Users in general, not just leaders, seem to have a poor grasp on the fact that AI is probabilistic (deals in likelihoods, not certainties) and almost certainly requires iterative testing and refinement, even in the most well-strategized and well-executed implementations.
When the first pilot doesn’t meet these inflated expectations, executive sponsorship evaporates, funding is cut, and the entire program gets labeled a failure. AI excels in particular use-cases and is arguably mediocre (compared to human professionals) in many more. Yes, it’s still advancing, but is not yet a replacement for a lot of the work that human beings do.
“Nail, Meet Hammer.”
In 1966, the renowned American psychologist Abraham Maslow wrote, “it is tempting, if the only tool you have is a hammer, to treat everything as if it were a nail.” While Maslow was speaking of physical tools, his Law of the Instrument applies just as readily to computer science and IT, where it is known as the “Golden Hammer.” If you have a large amount of effort and money invested in a tool, every problem starts to look like a good job for it.
The overwhelming enthusiasm about AI, and the speed at which it has gone from a research curiosity to 75% of the gains, 80% of the profits, and 90% of capex for the American stock market in 2025, has created a huge amount of urgency around adoption. Organizations act out of fear of being left behind. Assuming this frenzied posture primes an organization to implement a solution in search of a problem.
A C-level executive declares, “We need an AI strategy!” without specifying why. Teams are then forced to find applications for a pre-selected technology, rather than selecting the right tool for a known job. Pilots and PoCs are created that are technically impressive but have no measurable impact on revenue, cost, or customer satisfaction. Into the project graveyard it goes.
Don’t risk treating AI as purely an IT project
AI is not simply a new software package to be installed and distributed to employees or customers. Thoughtful integration requires business-level planning and strategy across the entire organization, and thus mandates careful stakeholder engagement. This makes it impossible for even the best-resourced projects to succeed without heavy cross-departmental collaboration.
The least useful AI solution is one built in a vacuum. These projects don’t integrate smoothly into real-world workflows, fail to gain user adoption (or actively turn users hostile), and will consistently fail to deliver the intended business value.
Orgs consistently underestimate challenges with data debt and infrastructure
AI models are built on data. Many organizations have vast amounts of data, but it’s often a Frankenstein-like combination of siloed, inconsistent, poorly labeled, and low-quality data. This is the aforementioned “data debt.” As with financial debt, it can be accumulated with ease despite the best of intentions, and dealing with it is not as simple as acquiring it.
Leadership assumes that because they have “Big Data,” they are ready for AI. They are shocked by the time, cost, and effort required to consolidate, clean, and properly label their data to make it usable and performant for training models. The result? Prolonged development phases focused on trying to prepare the data for training. Budgets are blown, momentum is lost, and the problem is entirely preventable when organizations understand the value of data management best practices.
AI, like most software, isn’t a “one-size-fits-all” solution
Cost-conscious organizations often believe they can simply purchase an off-the-shelf AI solution. However, the most valuable AI applications are always going to be tailored to a company’s unique processes and data. This is an example of short-term vs. long-term value creation. An off-the-shelf solution will create short-term value by generating immediate insights, but those insights will be much shallower, and far fewer of them will be actionable compared to a custom solution.
The cost associated with developing and implementing a custom, cutting-edge AI model is naturally going to be steeper than handing the company card to OpenAI, Anthropic, et al. However, the end result is far more likely to improve an organization’s competitive advantage.
Don’t neglect the human element, with employees or customers
AI changes job roles, creates fear of job displacement—particularly when paired with aggressive cost-cutting, and requires new skills to master. A top-down mandate that ignores this human impact is doomed to face resistance, either from employees, customers, or both.
Employee buy-in for these initiatives is incredibly important. Projects need to be announced with a clear communication plan, training for affected employees, and a vision for how AI will augment, not just replace, current human work. Leaders often focus on AI’s transformative potential while underestimating the genuine disruption it creates in employees’ work lives.
Production and scaling can’t be afterthoughts
Many organizations are good at creating a one-off Proof of Concept (PoC), but have no plan for moving it into a live, operational environment where it must perform reliably at scale. So, the PoC ends up on a shelf, never delivering value. Even if deployed, the model’s performance will inevitably degrade over time without the means to maintain it, leading to slow and silent failure.
AI, for the right organizations with specific, high-value use cases, has the potential to offer transformative solutions. But those are going to be transformative solutions. They cannot be treated as off-the-shelf or one-size-fits-all. Thoughtful and successful integration must scale from the bottom up, not be applied from the top down.
About the Author: Brad Ausrotas
Brad Ausrotas works at Ivey Publishing and first began experimenting with neural networks and LLM-based AI after the release of Google’s DeepDream in 2015.
