The Cost of Convenience: How Delegation and Decision Horizons Erode Human Agency when overusing LLM Systems

Large Language Models (LLMs) are rapidly embedded in everyday interfaces as assistants, copilots, and autonomous agents that influence human decision-making. Prior research primarily evaluates these systems through accuracy, alignment, and task performance, assuming that agency is preserved as long as users retain formal control. This assumption cannot explain how agency erodes without obligation or manipulation. We present a conceptual framework that treats LLMs as delegated decision infrastructures to which users repeatedly offload decisions for convenience. We introduce decision horizons, defined as the set of options and framings made epistemically visible by LLMs, as a central analytic concept. Repeated delegation reshapes these horizons, increases opt-out friction, and raises the cost of reversibility. Rather than removing control, LLMs render agency latent by restructuring decision conditions. We conceptualize agency erosion as an emergent, cumulative interaction effect and outline design implications that foreground reversibility, horizon diversity, and long-term human agency.