Toward LLM-aware software effort estimation: a conceptual framework

Software effort estimation has traditionally been grounded in the assumption that development cost is primarily driven by human labor, approximated through proxies such as code size, functional complexity, or perceived task difficulty. The increasing adoption of large language models (LLMs) as software development assistants challenges this assumption by automating substantial portions of reasoning, coding, and refactoring. In LLM-assisted workflows, effort increasingly shifts toward interaction management, validation, correction, and integration, leading to growing misalignment between established estimation techniques-such as COCOMO, Function Points, and Story Points-and actual development cost. This paper argues that the limitations of existing estimation models in LLM-mediated development are structural rather than parametric. When core development activities are delegated to automated reasoning systems. Through conceptual analysis supported by exploratory observations, we illustrate systematic mismatches between traditional effort estimates and LLM-assisted task execution, particularly in agile environments that rely on Story Points. To address this gap, we introduce a unified conceptual foundation for LLM-aware software effort estimation. We reconceptualize effort as Hybrid Intelligence Effort, emerging from the interaction between LLM cognitive complexity and human oversight effort. We further identify five core dimensions governing effort in LLM-assisted development: LLM reasoning complexity, context and information completeness, code transformation impact, iterative reasoning cycles, and human oversight effort. These dimensions capture cost drivers that are largely absent from conventional estimation theory. Rather than proposing a parametric estimation model, this work establishes a theoretical foundation for future empirical calibration and data-driven approaches. By redefining what constitutes effort in the presence of LLMs, the paper contributes a conceptual basis for estimation models aligned with contemporary AI-augmented software engineering practices.