Decision Debt in AI-Generated Work: Preserving Observability, Traceability, Modularity, and Changeability

Artificial intelligence can produce work that is functionally correct at delivery yet difficult to inspect, reconstruct, or modify later. This paper develops the concept of decision debt for AI-generated work: the future cost created when an acceptable output is retained without preserving the assumptions, evidence, dependencies, and interfaces needed to understand and safely change it. The concept extends but does not replace technical debt. Whereas technical debt traditionally concerns implementation and architectural compromises that increase future maintenance cost, decision debt focuses on lost decision structure across software, analytical work, and agentic workflows. The paper proposes four preservation properties - observability, traceability, modularity, and changeability - and a future-changeability test for evaluating whether AI-generated artifacts remain governable after delivery. It also offers a qualitative decision-debt checklist and a risk-adjusted preservation heuristic for deciding how much reconstruction evidence to retain. It also distinguishes external reconstruction evidence from private model reasoning: maintainability does not require access to an AI system's hidden chain of thought, but it does require sufficient records of sources, actions, tests, assumptions, and artifacts to reconstruct consequential decisions. The contribution is integrative rather than a claim that provenance, maintainability, or technical debt are individually new. The central claim is that correctness is a present-tense property, while changeability is a future-tense property; reliable AI-generated work requires both.