I hear a lot from business leaders about how much they want and need AI to work for them. They’re facing pressures around urgency, investment, and rising expectations. What I see less often is an equally clear understanding of what it will take to move beyond the pilot phase and make AI work as part of the business in the long term.
There are multiple tactical challenges companies face when they try to grow AI from a small pilot to a full-blown program. McKinsey research shows that 62% of organizations are still experimenting with AI. Only 7% have managed to go from a successful pilot to a successful, full-scale program.
In our work, we see the same pattern. According to The Economist, only about 11% of U.S. workers are actively using AI in their day-to-day roles. The issue generally is not a lack of interest or activity. Progress is uneven, and in many cases, it happens in isolated pockets rather than across the enterprise.
Over the course of building a business through multiple phases of growth, I have seen this dynamic play out repeatedly. Early success creates momentum, but it can also create a false sense of readiness. Teams prove that something works under controlled conditions and assume the rest of the organization can absorb it. In reality, that is where the more difficult work begins. AI pilots are designed to validate a use case, not to test the durability of the systems, data, and processes required to support it.
When AI moves into production, it encounters the business as it actually operates. Data is fragmented across systems, definitions vary between teams, and information moves with a level of inconsistency that was manageable before but becomes a constraint under pressure. At that point, even well-performing models begin to produce outputs that feel unreliable, not because the technology has failed, but because the environment around it has not been built to support it. What follows is a gradual erosion of confidence, as employees begin to second-guess results and usage declines over time.
The first place this shows up is in the data itself. In a pilot, data is curated and intentionally prepared. At scale, it reflects the full complexity of the business, including duplicate records, outdated information, inconsistent formats, and disconnected sources that were never designed to work together.
None of these issues is individually severe enough to stop progress, but together they create enough friction that maintaining trust becomes difficult. In practice, what organizations experience extends beyond a system failure to a credibility problem, where outputs require constant validation, and the tool’s value begins to diminish.
Closely tied to that is how information moves across the organization. Many companies have invested in modern platforms and cloud environments, but the presence of those systems does not guarantee that data flows reliably between them. When information remains isolated within functions or tools, AI can only operate within those boundaries, limiting its impact to localized improvements. That is why so many initiatives show strong results in one area while failing to translate into broader operational value. The underlying issue is not capability, but connectivity.
As AI becomes more embedded in workflows, governance introduces another layer of complexity that cannot be addressed informally. AI’s output generation is only the beginning; organizations need clarity around how those outputs are validated, who owns them, and how risk is managed as usage expands. In many environments, those questions are still answered implicitly, based on individual judgment rather than defined processes. That approach can work in a small, controlled setting for a limited time, but it likely won’t scale consistently once implemented across multiple teams, functions, and geographies, where consistency and accountability are required.
There is also a timing element that is often overlooked. For AI to become operational, it has to move at the speed of the business. In many organizations, analytics still operate on cycles measured in days or weeks, which means AI insights arrive too late to influence real decisions. When that happens, AI becomes something adjacent to the work rather than embedded within it. That dynamic helps explain why adoption remains limited even in organizations that have already made significant investments.
These patterns are consistent across industries, even though they manifest differently. In energy, the data required to drive AI is abundant, but it is often split across upstream operations, downstream systems, and compliance environments that do not connect in real time. In manufacturing, inconsistent data quality across plants limits the ability to scale insights beyond a single facility. In healthcare, strong governance frameworks are in place, but data is distributed across systems with different standards and regulatory requirements. In each case, the challenge is not whether AI can deliver value, but whether the business has been structured to support that value.
In our Q1 Tech Trends Report, we examined this through the lens of data maturity and found that most organizations are not progressing linearly from experimentation to scale. Instead, they are operating across multiple levels of readiness at once, investing in advanced capabilities while foundational elements such as data quality, governance, and infrastructure are still evolving. That imbalance is what slows progress and creates the perception that AI is underperforming.
The organizations that are moving forward are approaching this differently. Rather than starting with where AI can be applied, they begin with a more disciplined understanding of how their business operates today, including where data is reliable, where it is not, how information flows, and where accountability sits. They recognize that the most visible components of AI are not what determine long-term success. The work that compounds over time is less visible and more foundational: standardizing data, defining ownership, enforcing quality, and building systems that allow information to move consistently across the organization.
There is a natural tendency to believe that once a pilot succeeds, scale will follow. In practice, scale follows preparation. The companies that will capture the most value from AI will only rarely be the ones moving the fastest, and more often the ones with the clearest understanding of their own readiness and the discipline to build the foundation required to support it.
Maruf Ahmed is the CEO of Dexian
Photo courtesy Hartono Creative Studio for Unsplash+