The AI Scaling Gap Is the Most Underreported Story in Enterprise IT Right Now

We are in mid-2026. McKinsey's State of AI report says 88% of organizations report using AI in some form, up from 78% the year before and 55% the year before that. But only 7% have fully scaled AI across their organization. I find that spread more interesting than the adoption number, and I think it deserves more attention than it gets.

The 88% figure is the one that shows up in board presentations and industry conference decks. It is technically true and practically misleading. "Using AI" at the 88th percentile could mean a VP approved a pilot program that twenty engineers are running in a sandbox. It could mean the company pays for a Copilot license nobody uses consistently. It could mean someone in marketing uses ChatGPT for draft copy and calls it AI strategy. The question of whether an organization uses AI is a lot easier to answer yes to than the question of whether it has integrated AI into the business processes that determine performance. The 7% figure is the one that answers the harder question, and it tells a much more complicated story.

Cohen and Levinthal (1990) gave IS researchers a framework that explains the gap without flinching. Absorptive capacity is the organizational ability to recognize the value of new external knowledge, assimilate it, and apply it to commercial ends. The critical word in that definition is "prior." The capacity to absorb new knowledge is a function of prior related knowledge. Organizations that have not built data infrastructure, data literacy, and decision routines around evidence cannot absorb AI capabilities effectively, regardless of how much they spend on licenses or how prominently the CEO mentions AI in the annual report. You cannot learn from knowledge you are not equipped to recognize as relevant.

This is where the adoption-to-scaling gap comes from. The organizations that adopted AI fastest were not necessarily the ones best positioned to scale it. Many adopted under mimetic pressure, which is the same institutional force I have written about in the context of institutional isomorphism and AI adoption. Under uncertainty, organizations copy perceived leaders. They adopt. But absorptive capacity is cumulative and path-dependent, which means the adoption of AI tools does not automatically build the internal knowledge and routines needed to exploit those tools. The tool arrives. The organizational capacity to benefit from it is built slowly, through experience, investment, and failure, or not built at all.

What absorptive capacity predicts is that the organizations in the 7% are disproportionately the ones that had already been doing something like AI before it was called AI. They had data teams asking hard questions of data. They had decision routines that incorporated quantitative evidence. They had governance structures that could evaluate model outputs and decide when to trust them. When GenAI became available, they could recognize which applications mattered and build the organizational processes to act on what those applications produced. The prior knowledge was there. The remaining 81%, adoption without scaling, are organizations that recognized a signal and responded to it without the underlying capacity to do much with it.

Zahra and George's 2002 reconceptualization of absorptive capacity is useful here. They split the concept into potential and realized. Potential absorptive capacity covers the ability to acquire and assimilate new knowledge. Realized covers the ability to transform and exploit it. An organization can be high on potential and low on realized. It gets the knowledge in the door. It cannot convert it into changed routines and performance gains. That is, I think, exactly what the 88%-to-7% ratio describes at the level of an industry. The mass adoption phase produced lots of potential absorptive capacity. The scaling phase requires realized capacity, and realized capacity is harder, slower, and less visible.

From what I see in IS research, we spend a lot of time studying adoption and comparatively little studying what happens after the pilot succeeds and the license is signed. I understand why. Adoption studies are cleaner. You can survey people about their intent to use a system or measure when a system was deployed across an organization. The scaling phase is messier. It unfolds over years, not quarters. It is entangled with organizational politics, IT governance, and business process change. The measurement problems are harder. But if only 7% of organizations that claim to use AI have actually scaled it, and if the gap has been sitting at roughly this level for the past two years despite massive increases in adoption, then the scaling phase is where the real action is. The rest of the adoption literature is describing how organizations arrive at the starting line, not how they run the race.

The data on GenAI specifically sharpens this picture. McKinsey reports that GenAI use went from 33% of organizations in 2023 to 71% in 2024 to 79% in 2025. That is one of the fastest adoption curves ever recorded for an enterprise technology category. And the scaling rate sits at 7%. If the scaling bottleneck were just a function of time, we would expect to see it climb as adoption matures. But Gartner's projection that 40% of agentic AI projects will be canceled by end of 2027 suggests the bottleneck is not primarily about time. It is about organizational capability. Projects are not failing because organizations waited too long. They are failing because the organizations lacked the realized absorptive capacity to convert what they adopted into something that performs.

The IS research agenda that follows from this observation is less about why organizations adopt AI, which is now well-understood at a general level, and more about what determines whether organizations move from adoption to scaling. What organizational capabilities does scaling require, and how are those capabilities built? What role does prior IT governance experience play? How do firms with low absorptive capacity break out of the path-dependent trap Cohen and Levinthal described? These are questions the field can answer with the right study designs, and they matter more right now than one more adoption model.

The 88% headline will keep getting used to demonstrate that enterprise AI has reached mainstream status. It is true, and it is useful context. But the 7% scaling rate is the number that tells you what the era actually looks like from inside the organizations living through it. It tells you that most of the work is still ahead.