Position: Assistive AI requires Personalized Specialists, not Generalists

The AI community is rapidly converging on generalist foundation models trained on web-scale data. While this paradigm has yielded impressive gains, we argue that reliable assistive AI requires a complementary objective: systems that become increasingly well matched to a particular individual, local environment, and interaction history. The most valuable assistants will not merely be those that can attempt many tasks for many users, but those that can do the right things for a specific person over a long period of time. We take the position that deployment-adaptive specialists are a central target for high-impact assistive AI: systems whose specialization is defined not by narrow task taxonomies, but by sustained coupling to an individual user, their local environment, and their evolving interaction history. This is not a rejection of generalist pretraining: broad foundation models provide increasingly strong priors. The unresolved scientific problem is converting those priors into stable local competence without repeated costly mistakes. We substantiate this argument through three case studies: (i) AI agents that help humans automate web activities; (ii) wearable assistants that predict actions in-context from continuous egocentric streams; and (iii) home robots that assist in daily tasks under safety and compliance constraints. In these settings, the most critical data is often generated after deployment as a streaming, privacy-sensitive, on-policy interaction stream. We outline research directions for building locally adaptive assistants that learn from organic observational data, avoid self-reinforcing errors, and improve safely over long horizons.