Key points are not available for this paper at this time.
In 2024, a total of 361 billion emails were sent and received by businesses and consumers each day. Email remains the preferred method of communication for work-related matters, with knowledge workers spending two to five hours a day managing their inboxes. The advent of Large Language Models (LLMs) has introduced new possibilities for personalized email automation, offering context-aware and stylistically adaptive responses. However, achieving effective personalization introduces technical, ethical, and security challenges. This survey presents a systematic review of 32 papers published between 2021 and 2025, identified using the PRISMA methodology across Google Scholar, IEEE Xplore, and the ACM Digital Library. Our analysis reveals that state-of-the-art email assistants integrate RAG and PEFT with feedback-driven refinement. User-centric interfaces and privacy-aware architectures support these assistants. Nevertheless, these advances also expose systems to new risks such as Trojan plugins and adversarial prompt injections. This highlights the importance of integrated security frameworks. This review provides a structured approach to advancing personalized LLM-based email systems, identifying persistent research gaps in adaptive learning, benchmark development, and ethical design. This work is intended to guide researchers and developers who are looking to create secure, efficient, and human-aligned communication assistants.
Novelo et al. (Mon,) studied this question.