If GenAI Has $4.4 Trillion in Potential, Why Is Only 7% Scaled?

The number that stopped me was not the big one. It was the small one.

McKinsey's 2023 analysis of generative AI's economic potential put the annual value at somewhere between $2.6 trillion and $4.4 trillion across industries, on top of a broader AI economic impact that could reach $11 trillion to $17.7 trillion when you count all automation-adjacent AI. The report estimated that 60 to 70 percent of work activities could theoretically be automated with current or near-current technology. That is the number people quote in keynotes and LinkedIn posts. The $4.4 trillion. The 70 percent. The moonshot.

The number that caught my attention was from McKinsey's 2025 State of AI report. Of the 88 percent of organizations now using AI in some form, only 7 percent have fully scaled it. Seven. Percent.

So we have an estimated $4.4 trillion sitting there waiting, and most organizations are either still in pilot mode or doing something that counts as AI use only in the most generous interpretation of that word. Gartner's forecast that worldwide AI spending will total $2.5 trillion in 2026 makes the gap stranger, not smaller. Organizations are pouring money in. The scaling is not coming out the other end.

I want to try to think through why, because I think the standard answer ("change is hard, culture is slow") is technically true but not very useful.

The McKinsey potential figure is built on a task-level analysis. Researchers mapped work activities across industries, identified which of those activities involve skills that current AI systems can plausibly replicate, and added up the economic output associated with those activities. The result is a capability ceiling: if AI could do everything it is theoretically capable of doing, in every organization that could theoretically use it, here is what you get. That is a useful benchmark for thinking about the long run. It is not a description of what any given organization can actually do this year.

The problem is that tasks are not independent units floating free in space. They are embedded in workflows. Workflows are embedded in organizational processes. Organizational processes are governed by roles, incentive structures, approval chains, system dependencies, and human relationships. When McKinsey says 60 to 70 percent of activities are automatable, it means the cognitive or physical content of those activities is within reach of current AI capability. It does not mean any organization can actually automate them without first changing every organizational layer in which they are embedded.

This is a distinction that IS research has spent decades making, and I think it does useful work here. The distinction is between a technology's potential and an organization's capacity to realize that potential. Absorptive capacity, which Cohen and Levinthal (1990) defined as the organizational ability to recognize, assimilate, and exploit new external knowledge, is path-dependent. You need prior related knowledge to absorb new capabilities. An organization without data infrastructure, without ML engineering practices, without governance frameworks for AI outputs, and without leadership that understands what questions to ask about model quality cannot simply decide to fully scale AI because the potential is large. The capacity to absorb that potential has to be built first, and building it takes time, investment, and deliberate organizational change.

I wrote about this in the context of pilots in a post about why AI pilots rarely become products. The pilot-to-scale gap is exactly the absorptive capacity gap made visible. The pilot works under controlled conditions with dedicated teams. Full scaling means embedding AI into the messy, distributed, politically complex environment of actual organizational operations. Those are different problems, and success at the first tells you very little about readiness for the second.

The 7 percent figure makes more sense once you put it this way. Fully scaling AI is not a procurement decision or an API call. It is an organizational transformation that requires building complementary capabilities alongside the technology. Melville et al. (2004) modeled this explicitly: IT resources interact with complementary organizational resources to improve business process performance, which then drives organizational performance. The arrow runs through organizational capability, not around it. If the complementary organizational resources are not there, the IT resource sits underutilized. The potential stays potential.

There is also a measurement problem that cuts the other direction. The $4.4 trillion estimate assumes that automating a task captures all of the economic value that task currently represents. But many tasks produce value through human judgment, relationship context, and situational awareness that is not fully captured in the task description. When an account manager writes a proposal, the economic value is partly in the words on the page and partly in knowing which framing will land with this specific client at this specific moment in the relationship. AI can help with the former. It does less well with the latter, and "automatable task" does not distinguish between them.

None of this means the potential is fictional. It means the gap between potential and scale is an organizational and institutional problem, not a technical one. The technology, for a growing range of tasks, is capable enough. The organizations are not yet built to use it at scale. What IS researchers can contribute here is not better forecasts of the potential. It is better theory and evidence about what organizational and technical conditions allow AI to move from pilot to scale: what governance structures help, what prior knowledge is prerequisite, what kinds of workflow change are necessary before the model can do what the demo suggested it could do.

That strikes me as a more useful research question than the one most AI coverage is asking. The coverage keeps asking how large the potential is. The interesting question is what closes the gap between the potential and the 7 percent.