The 81% Gap Has a Name and It Is Absorptive Capacity

Eighty-one percent. That is the gap between organizations that adopted AI and organizations that scaled it, and the number has been stubbornly sitting there for two years while adoption climbed from roughly 55 percent to nearly 90. Recent industry reports suggest the scaling rate hovers around 7 percent. I find the spread more interesting than either number alone. Adoption is easy to measure and easy to report and easy to claim in a board meeting. Scaling is none of those things. And the theory that explains why they diverge has been sitting in the organizational learning literature since 1990.

Cohen and Levinthal (1990) defined absorptive capacity as a firm's ability to recognize the value of new external information, assimilate it, and apply it to commercial ends. The critical mechanism in that definition is not recognition. It is not assimilation. It is the word they bury inside the logic: prior. The capacity to absorb new knowledge is a function of prior related knowledge. Firms that have not invested in understanding a domain cannot absorb new knowledge in that domain, regardless of how much they spend acquiring it. This is not a technology problem. It is a learning structure problem, and it is self-reinforcing. The firms that know something learn faster because they can recognize what matters, interpret it quickly, and act on it. The firms that know nothing fall further behind because they cannot even tell what they are missing.

TAM and UTAUT can explain the 88 percent. They cannot explain the 7 percent. Davis (1989) built the Technology Acceptance Model on two constructs: perceived usefulness and perceived ease of use. The full chain runs from perceived ease of use to perceived usefulness to behavioral intention to actual use. Venkatesh et al. (2003) consolidated eight acceptance models into UTAUT, adding social influence and facilitating conditions to the mix. Both models predict whether an individual or organization will adopt a technology. They do this well, and decades of studies confirm it. But the dependent variable is adoption. The question TAM and UTAUT answer is: will they use it? The question the 81 percent gap raises is: will they get value from using it? That is a different question entirely, and TAM and UTAUT were never designed to answer it.

I think this distinction matters more than most IS researchers acknowledge, because the field has spent enormous effort refining adoption models while comparatively little effort has gone into understanding what happens after adoption succeeds. The adoption literature tells you who will show up. It does not tell you who will learn. I wrote about extending TAM instead of replacing it and the citational momentum that keeps the field circling the same question. That momentum is part of why we can explain adoption precisely and scaling poorly.

This is where Zahra and George (2002) become indispensable. They reconceptualized absorptive capacity into two components: potential and realized. Potential absorptive capacity is the ability to acquire and assimilate external knowledge. Realized absorptive capacity is the ability to transform and exploit that knowledge for organizational benefit. The distinction is precise and it maps directly onto the 81 percent. The organizations that adopted AI built potential absorptive capacity. They acquired tools, subscribed to platforms, hired talent, ran pilots. The knowledge came in the door. What the 81 percent lacks is realized capacity. They cannot convert absorbed knowledge into changed routines and improved performance. The data sits in the lake. The dashboard glows on the unused monitor. The pilot becomes a PowerPoint slide in a quarterly review.

The reason TAM and UTAUT cannot explain the gap is structural, not incidental. Both models locate the dependent variable at the point of adoption intention or use. Once the user crosses the threshold into actual use, the model stops. What happens inside use, whether use is effective or ineffective, whether it produces organizational value or just occupies time, is outside the model's boundary. Davis (1989) was explicit about this. TAM predicts intention to use and use. It does not predict the quality of use or the organizational outcomes that follow from use. Venkatesh et al. (2003) consolidated more constructs and explained more variance in intention, but the boundary stayed the same. The last node in the model is still use, not value.

Absorptive capacity starts where TAM stops. Its dependent variable is not adoption. It is the organizational ability to learn from external knowledge and apply it. The mechanism is cumulative and path-dependent. Prior knowledge enables recognition, recognition enables assimilation, assimilation enables application. Break any link and the chain breaks. An organization with no prior data analytics experience can buy the most advanced AI tool on the market and it will not matter, because the recognition link is broken. The tool produces insights. The organization cannot tell which insights matter. An organization with some experience but no decision routines for acting on insights has a broken transformation link. The insights are recognized and assimilated. They die in a meeting where nobody has authority to change the process. This is why tools alone fail. The tool is not the bottleneck. The organizational capacity to learn from the tool is.

The uncomfortable implication is that the 81 percent gap is not a technology gap at all. It is a prior knowledge gap. The organizations that scaled AI are, I think, disproportionately the ones that were already doing something like AI before it had that name. They had data teams running experiments. They had analyst communities asking hard questions of data. They had governance structures for evaluating model outputs. When large language models became accessible, they could recognize which applications had commercial potential, they could assimilate the new capabilities into existing workflows, and they could transform those workflows into something better. Their realized absorptive capacity was already high. The organizations that adopted but did not scale had potential capacity without realized capacity. They got the knowledge in the door and could not act on it. The gap is not between adopters and non-adopters. It is between organizations that can transform what they absorb and organizations that cannot.

Melville et al. (2004) showed that IT business value travels through complementary organizational resources and process change. The technology alone produces nothing. The value materializes when complementary resources, human expertise, organizational routines, decision authority, are in place. Absorptive capacity is part of the mechanism that determines whether those complementary resources get built. Cohen and Levinthal predicted this in 1990. The firm with high absorptive capacity builds the complementary resources around its technology investments. The firm with low absorptive capacity buys the same technology and changes nothing, which is exactly what the productivity paradox describes. Spending on IT rises. Productivity does not follow. The gap between the two is not a measurement error. It is a learning gap.

The research agenda that follows from this is not another adoption model. It is a scaling model. What organizational capabilities does moving from potential to realized absorptive capacity require, and how are those capabilities built over time? What is the minimum viable prior knowledge that lets an organization start building realized capacity for AI? How do firms with low absorptive capacity break out of the path-dependent trap Cohen and Levinthal described? These are questions that extend the scaling gap analysis into testable organizational theory, and they are questions the IS field can answer with the right study designs.

The 81 percent gap has a name. It is not adoption failure. It is not a technology problem. It is the distance between potential and realized absorptive capacity, and closing it requires something no vendor can sell.