The Paradox at the Heart of AI Adoption: 88% Explore, 7% Scale

Eighty-eight percent of organizations say they use AI. Only seven percent have scaled it. I keep coming back to that pair of numbers because the gap between them is not just a lag. It is a paradox.

Organizations must explore new AI capabilities and exploit the ones they already have. March (1991) made this tension famous. Exploration is search, discovery, and risk. Exploitation is refinement, efficiency, and execution. The two require different mindsets, different resource allocations, and different organizational structures. They also compete for the same limited attention and budget. March's argument was that adaptive systems that engage exclusively in exploitation will become competent at the wrong thing as their environment changes, while systems that engage exclusively in exploration will never reap the returns from what they discover. Success requires both, but the two are fundamentally incompatible in their demands.

Smith and Lewis (2011) took this further by treating paradox not as a problem to be solved but as a permanent feature of organizing. Their dynamic equilibrium model says that contradictions like exploration and exploitation are interdependent and persistent. Managers do not resolve them once and for all. They manage them through temporal cycling, structural separation, or acceptance of the tension itself. The goal is not to eliminate the paradox but to sustain productive instability.

When I look at the 88/7 gap through this lens, it stops looking like a failure of implementation and starts looking like a structural symptom. The 88 percent who "use" AI are exploring. They are piloting generative models, running Copilot licenses, testing customer service chatbots, and building proof-of-concept dashboards. I wrote about the scaling gap before as a measurement problem, but this time I want to treat it as something deeper. The vast majority are in exploration mode: collecting information, learning what the technology can do, and staying current with what competitors appear to be doing. The 7 percent who have scaled are exploiting. They have embedded AI into workflows, restructured roles around it, built governance for model outputs, and measured its impact on business outcomes. Exploitation is slower, harder, and less glamorous. It also generates the actual returns.

The gap is paradoxical because the activities that drive the 88 percent upward are exactly the activities that prevent the 7 percent from growing. Exploration absorbs resources, attention, and political capital that could go toward exploitation. Every pilot that gets funded is a process redesign that does not. Every workshop on AI awareness is an hour not spent fixing the data pipeline that would make an existing model reliable enough to scale. Organizations celebrate exploration because it is visible, fast, and defensible. Exploitation is invisible until it works, slow to show results, and politically risky because it requires changing how people work. I have seen this pattern before in why AI pilots rarely become products. The pilot is exciting. The scaling is where you hit the data problems and the resistance all at once.

This is not a new pattern in IS research. Fichman and Kemerer (1997) identified the assimilation gap, the distance between technology acquisition and full productive deployment. In their study of object-oriented programming adoption, only 4.6 percent of organizations reached limited deployment and 1.0 percent reached general deployment. The gap was driven by knowledge barriers, complementary asset requirements, and network effects. AI today is replaying the same gap at a larger scale. Organizations acquire AI but fail to deploy it productively because they lack the complementary organizational capabilities that would allow exploitation to succeed. You cannot exploit what you have not learned to stabilize.

Cohen and Levinthal (1990) explain why some organizations close this gap and others do not. Absorptive capacity is path-dependent. Prior related knowledge determines whether a firm can recognize the value of new information, assimilate it, and apply it commercially. An organization that spends all its energy exploring the newest large language model release is not building the internal knowledge base needed to exploit the model it already licensed six months ago. The exploration itself crowds out the exploitation that would make prior exploration pay off.

The IS literature on dynamic capabilities adds another layer. Pavlou and El Sawy described dynamic capabilities as exploration capabilities and operational capabilities as exploitation capabilities. The distinction matters for AI because merely using an AI system is an operational capability, while transforming organizational processes to capture value from that system is a dynamic capability. As Teece et al. (1997) argued, sustained advantage comes from sensing, seizing, and transforming. Most organizations are stuck in the sensing phase. They sense the opportunity in AI, but they do not seize it by reallocating resources, and they do not transform their structures to support it. Sensing without seizing is just expensive exploration.

Benbya, Pachidi, and Jarvenpaa (2021) explicitly map this tension onto AI-enabled innovation. They note that machine learning requires abundant training data, which biases AI toward domains where past data is plentiful. Novelty and creativity, by contrast, depend on small but rich data and tacit knowledge that is costly to digitize. Organizations pursuing AI-enabled innovation therefore face a tension between exploiting existing data and exploring beyond it. The paradox is built into the technology's architecture. AI systems are excellent at exploitation, which makes exploration feel unnecessary until the environment shifts and the exploited model becomes obsolete.

I think this is why the scaling problem is so hard to fix with better technology. The bottleneck is not model capability. It is organizational capability. You do not close the 88/7 gap by giving organizations a better API. You close it by building the structures that allow exploitation to compete with exploration for resources and status. That means recognizing that the paradox is permanent, that Smith and Lewis (2011) were right that dynamic equilibrium is the best we can hope for, and that the organizations succeeding at scale are not the ones that eliminated the tension but the ones that learned to manage it.

What strikes me about the current moment is how unevenly the field talks about this. AI conferences are full of exploration. New models, new benchmarks, new possibilities. The exploitation conversation, governance, integration, process redesign, measurement, happens in quieter rooms with fewer attendees. Until that imbalance changes, the 88/7 gap is exactly what we should expect.