Occupational exposure indices measure what generative AI could do, not what it is used for. I build a behavioral measure of realized AI adoption from the Anthropic Economic Index and compare it with the standard exposure index across 47 occupational sectors. Adoption and exposure rank occupations alike but diverge in level, because adoption is far more concentrated. In weekly Indeed job postings after ChatGPT, labor demand tracks adoption rather than exposure: a one standard deviation increase in adoption predicts a decline of about 10 log points, while exposure adds nothing. A second adoption measure, built from OpenAI usage, reproduces the pattern, so the result is not specific to one model family. Realized adoption captures AI’s early labor market footprint.