The Artificial Intelligence Economy

Artificial intelligence can make a draft, prediction, or proposed action cheap to produce. Delivering useful work still requires checking it, correcting mistakes, and taking responsibility for the result. This paper develops an economic model of how those remaining costs shape AI adoption, productivity, and the distribution of gains. Firms choose how to combine human work, AI, review, and recovery procedures to deliver a task at a specified quality standard. The analysis then traces how these choices affect costs and labor demand across tasks and connected industries. It establishes five connected results. If unavoidable costs of checking or error remain, better generation alone cannot make completed work arbitrarily cheap. Further savings can nevertheless come from better checking, cheaper computing, and more productive human work. For long workflows, reliability and the placement of checks determine how much work can economically be delegated; exposure to irreversible harm can require more frequent checks. When a task first switches to AI, the saving can be small even if the reduction in human work on that task is large. Whether this reduction translates into job losses depends on reassignment, demand, and the wider economy. Finally, faster generation of research ideas need not produce faster validated discoveries when the capacity to test them is limited. The paper connects these results to investment, wages, ownership, market structure, and policy. The policy analysis distinguishes support for testing and organizational investment from measures addressing worker adjustment, competition, and accountability. A reanalysis of public data on 246 software-development tasks involving 16 developers illustrates how observed working time enters a cost comparison, while a proposed study across organizations specifies the additional evidence needed. The results depend on stated assumptions and do not constitute a forecast of economy-wide productivity or employment.