Forecasting AI Adoption Across U.S. States

AI adoption is expanding across sectors, but its diffusion remains uneven across firms, industries, and geographic areas. These differences have important implications for productivity, labor demand, skill needs, and regional economic development. Yet, while prior work has largely focused on measuring AI adoption and identifying its drivers and barriers, relatively little research has examined how adoption evolves over time or how it can be forecast across geographic regions. One key challenge has been the limited availability of timely, geographically detailed data. Since 2023, the U.S. Census Bureau’s Business Trends and Outlook Survey (BTOS) has provided repeated measures of business AI use at the national, state, and industry levels, creating an opportunity to study the temporal dynamics of adoption at the state level. Using state-level current-use AI rates from BTOS, we formulate AI adoption as a temporal forecasting problem and examine whether adoption patterns in other states provide predictive information beyond a state’s own historical trend. We construct three graph-derived sources of cross-state information: a national graph connecting each state to all other states, a geographic graph connecting neighboring states, and a workforce-composition graph connecting states with similar industry employment and occupational skill profiles. Adoption rates from connected states are aggregated to construct lagged cross-state features, which are evaluated individually and in combination with state-level adoption history using multiple forecasting models. Our results show that cross-state adoption features generally improve forecasting performance beyond a state’s own historical adoption trend. Combining national, geographic, and workforce-composition features achieves the strongest overall performance, while national adoption patterns provide the strongest predictive signal among individual cross-state feature groups. We also find substantial variation across states: cross-state information improves forecasts for many states but provides smaller gains or worsens performance for others. These findings demonstrate that AI adoption dynamics extend beyond state-specific historical trends and that incorporating relational information across states can improve geographically granular forecasts of emerging technology adoption.