Direction, execution and substitution in cognitive work with AI

Disclosure of AI use in research and education records which tasks AI performed and how much of the work it did. Two constructed researchers show what that measure misses. One sets every decision in advance and lets an agentic system run the analysis for most of a day. The other receives a single reply pointing out a confound and rebuilds the study around it. The first reports far more AI use, but the second answers a question that AI reshaped. Three terms are proposed, each applied to one stage of one piece of work. Direction asks where the path of the work came from, with assisted and guided as its two ends. Execution asks how much of the implementation AI carried out. Substitution asks whether AI did thinking the person was meant to do themselves, and applies only where such a target has been set. Five constructed cases show that knowing two of the terms leaves the third open. A record of AI use can count Execution within the exchanges it shows, while assessing Direction and Substitution requires evidence beyond the record. Each term has a precedent in earlier work on automation, attribution of AI-assisted text and cognitive offloading. Measuring the terms is the next step.