Generative Intelligence for Personalized Digital Assistants

Abstract

Personalized digital assistants -- AI systems that adapt their communication style, knowledge focus, task management, and proactive suggestions to individual user characteristics, preferences, and goals -- represent the most ambitious deployment context for large generative AI models: they must maintain rich, accurate user models over long interaction histories, generate contextually appropriate responses across highly diverse task types, and adapt in real time to changing user needs. This paper proposes the Generative Personalized Assistant (GPA) architecture, a comprehensive framework for building generative AI-powered digital assistants that achieve genuine personalisation through five integrated layers: a deep user model (DUM) that represents user preferences, knowledge states, goals, and interaction patterns across multiple dimensions; a personalised response generator (PRG) that conditions LLM generation on the DUM to produce responses adapted to user expertise, communication style, and context; a proactive task anticipator (PTA) that predicts user needs before explicit requests using temporal pattern analysis; a multi-domain knowledge integrator (MDKI) that maintains user- specific knowledge bases across professional, personal, and educational domains; and a personalisation quality monitor (PQM) that tracks personalisation effectiveness and triggers model updates when preference drift is detected. GPA is evaluated through a 120-day deployment study with 96 participants across three assistant application contexts: professional task management, personal health coaching, and educational tutoring. GPA achieves 4.6/5.0 user satisfaction (SD = 0.3) versus 3.2/5.0 for non-personalised baseline (Cohen d = 5.88); task completion rate improves from 64.8% to 88.4%; and proactive suggestion acceptance rate reaches 72.4% (SD = 6.8%). The study contributes the GPA specification, a Personalisation Effectiveness Index (PEI), and empirical evidence that deep user modelling with multi-layer personalisation substantially outperforms shallow personalisation approaches.