Token Leverage: A Framework for AI Inside the Firm

Abstract

Studies of generative AI at work find large gains in some settings, no effect in others, and losses in others. The pattern is less puzzling if the treatment is not access or adoption, but priced inference chosen for a worker-task pair. We call this margin token leverage, billed AI inference spend per dollar of task-allocated labor cost. Inference creates value only when the task lies within the model's frontier, the worker can orchestrate and audit output, the workflow supplies useful context, governance catches errors, and workers know when to abstain. Two measurement results follow. Under voluntary firm choice, positive billed spend, effective-cost dominance, and correct firm expectations, billed token leverage is a revealed willingness-to-pay lower bound on expected AI-created operating value per task labor dollar. Access and adoption estimates instead mix chosen use, optimal abstention, forced overuse, and constrained underuse, so their sign can be positive, zero, or negative even when chosen use is valuable. AI productivity research should therefore measure worker-task inference spend rather than stopping at whether workers had or used AI.

Keywords: AI, generative AI, artificial intelligence, large language models, AI productivity, enterprise AI, AI adoption, worker productivity

JEL Classification: M15, O33, D24, J24, L23