Judgment Delegation Is (Almost) All You Need: A Theory of Why AI Tools Feel Intelligent

AI tools increasingly appear to perform intelligent work in software engineering and other creative domains, even though their core operation is plausible continuation of context. We argue that this impression arises from judgment delegation: tool support makes latent judgment capabilities in language models delegable by externalizing state, enabling feedback through action, and enforcing decision boundaries. When embedded in such scaffolding, AI systems can absorb much of the option-generation and pre-commitment work that dominates underspecified tasks, allowing users to interact primarily at commitment points. This redistribution of judgment work reduces cognitive burden and creates the experience that the system is “doing the thinking,” even when authority over final commitments nominally remains with the human. At the same time, fluent progress can make delegation difficult to see. Users may allow system proposals to function as decisions by default, not through explicit authorization but through convenience and momentum. This implicit commitment delegation helps explain why AI tools feel autonomous in practice and why costs often surface as late rework or reversals rather than immediate errors. To make these dynamics observable, we introduce Decision Accounting, a lightweight evaluation lens that tracks commitments (D), human option-generation burden (G), and mismatch penalties arising from late detection of incompatible assumptions. Through worked software planning examples, we show that tool support primarily reduces option-generation burden rather than the number of decisions required by a task, and that apparent autonomy becomes brittle when commitment authority shifts implicitly without governance. We argue that perceived intelligence in AI-assisted work is best understood as an outcome of judgment allocation and cost control, not model capability alone.