Understanding and Supporting Formal Email Exchange by Answering AI-Generated Questions

Research Background and Problem

• Identified Problems or Challenges:

  • Responding to formal emails often requires time and cognitive effort, especially in crafting polite expressions and accurately addressing the sender's needs.
  • Current automated reply tools based on large language models (LLMs) require users to provide detailed prompts, which increases user workload. Additionally, users may need to repeatedly modify prompts if the results do not meet expectations.

• Significance:

  • In formal contexts such as workplace environments and academic collaborations, high-quality and polite email replies are crucial for maintaining trust and relationships. Errors or delays in responses can lead to negative impressions, affecting interpersonal relationships and communication efficiency.

• Research Motivation and Related Work:

  • Previous studies have shown that reducing the burden of creating complex prompts and utilizing a question-and-answer (QA) format can more effectively convey user intent.
  • Building on this foundation, this study proposes a QA-driven approach to replace traditional prompt design, thereby optimizing the formal email reply process.

Solution

• Proposed Method or Approach:

  • This study developed a prototype system, ResQ, which uses a QA-driven approach to assist users in replying to formal emails.
  • The system automatically generates relevant questions from received emails and drafts replies based on the user's selected answers.

• Innovations:

  • Introduced AI-generated questions to break down the complex prompt design and input process into manageable steps.
  • Employed a QA mechanism as a cognitive scaffold to simplify tasks, enabling users to quickly understand email content and efficiently construct replies.

• Implementation Steps and Key Technologies:

  1. Question Generation: Based on the email content, the system generates questions with options, covering all key points in the email.
  2. User Answering Questions: Users provide relevant information by selecting or manually inputting answers.
  3. Draft Reply Generation: Using the user's answers and preferences, the system generates a draft email reply via LLM.
  4. Proofreading and Adjustments: Users can modify the draft before sending to ensure it aligns with their intent.
  5. Sending the Email: Finally, users confirm the reply and send it.

Research Outcomes

• Specific Outcomes:

  1. Increased Efficiency and Reduced Workload:

     • The QA model significantly reduced the time required to reply to emails and improved cognitive load management.
     • Compared to traditional prompt design, users no longer needed to repeatedly generate or modify prompts.

  2. Improved Email Quality:

     • Experiments showed that emails generated using the QA model outperformed manual replies in terms of politeness, coherence, and meeting requirements.

  3. Easier Task Initiation:

     • The QA model reduced cognitive barriers to starting tasks, minimizing procrastination in replying to emails.

  4. Impact on User Perception of Control and Agency:

     • While the QA-driven system improved efficiency, users reported a decrease in their perceived "sense of control" and "sense of participation" in content creation.

• Comparison with Existing Solutions:

  • Compared to traditional prompt design, the QA model reduced the number of characters users needed to input, eliminating the need for extensive language organization.
  • Compared to fully manual operations, it significantly improved the quality and speed of email replies.

• Experimental or Evaluation Results:

  • Experiment 1 (12-person controlled experiment): The QA model significantly outperformed manual and traditional prompt-based modes in terms of efficiency, cognitive load, email quality, and task satisfaction.
  • Experiment 2 (9-person field study): Participants reported higher work efficiency during actual use, though the impact on "psychological distance" and "agency" varied depending on the context.

• Limitations and Future Directions:

  1. Use is limited to specific cultural contexts (e.g., Japan), and results may not fully generalize to other cultures.
  2. Long-term reliance on AI may diminish users' ability to compose emails independently.
  3. Future research should explore how different QA designs can optimize user experience for various communication scenarios while reducing over-dependence on AI.
  4. Enhancing privacy management mechanisms is necessary to expand the application of this technology in diverse formal contexts.

Through this research, ResQ demonstrates the potential of a QA-driven email reply approach in improving efficiency and reducing workload, while also providing valuable insights for optimizing AI-mediated communication tools in the future.