Decomposing the Capability Overhang: Access, Agency, and the Geography of AI Adoption

The gap between what AI systems can do and what people actually do with them, recently termed the “capability overhang” by OpenAI, has become a prominent policy concern at the national and international level. Yet the concept, as currently used, treats this gap as a single undifferentiated phenomenon and proposes correspondingly uniform remedies: more access, more education, more infrastructure. This paper argues that capability overhang is structurally two-dimensional. It simultaneously reflects failures of access: the inability of populations to reach AI tools, and failures of agency: the inability to use those tools in ways that generate productive, iterative engagement. These two failures are distinct in origin, operate at different speeds, and require fundamentally different policy responses.

We introduce the AI Matrix (Simpson, 2025a) as a framework for decomposing capability overhang along these two dimensions, and operationalize it using observed interaction data from the Anthropic Economic Index (AEI), a privacy-preserving analysis of 818,492 Claude.ai Free and Pro consumer conversations across 173 countries during November 13–20, 2025. Using use case composition, collaboration pattern distributions, and task success rates as proxies for the agency dimension, combined with conversation volume as a proxy for effective AI reach within the same dataset, we show that countries cluster into patterns broadly consistent with the AI Matrix’s four quadrants. The resulting typology describes three observable stages of AI adoption that represent a progression in structural complexity: an education-first stage characterized by single-domain use concentrated in academic coursework and high rates of passive delegation; a professional adoption stage characterized by single-domain concentration in work tasks among technically skilled populations; and a consumer diffusion stage characterized by broad use across work, personal, and creative purposes with declining rates of passive delegation.

We argue that this typology is not merely descriptive. By mapping it onto the AI Matrix’s transition logic, we offer a theoretically grounded interpretation of why the adoption curve takes the shape that Anthropic and OpenAI have independently documented, and why global gaps between countries have remained stable even as within-country convergence accelerates. This interpretation is consistent with the observed evidence rather than established by it. The paper is explicit about the limitations of a cross-sectional, single-provider dataset: the AEI data captures only countries above the access threshold, cannot establish causal direction, and reflects one week of interactions on one platform. Within these constraints, we argue that the patterns are consistent with the AI Matrix’s predictions and provide an empirical foundation for a policy framework that moves beyond access alone.

Keywords: AI adoption; capability overhang; AI Matrix; digital divide; human agency; technology diffusion; global inequality; Anthropic Economic Index.