Facilitating Proactive and Reactive Guidance for Decision Making on the Web: A Design Probe with WebSeek

Paper Title

Facilitating Proactive and Reactive Guidance for Decision Making on the Web: A Design Probe with WebSeek

Publication Info

• Topic area: Human-AI collaboration for web-based data-driven decision-making.
• Keywords: Web agents, mixed-initiative systems, data-driven decision-making, proactive AI guidance, reactive AI guidance, human-AI collaboration, browser extension, data analysis, visualization, large language models.

Background and Problem

• Problem / challenge: Existing web AI agents like ChatGPT Agent and GenSpark rely on text-based input, lack proactive detection of user intent, and do not support interactive data analysis or decision-making workflows. These tools often present static, non-interactive outputs, limiting user control and transparency.
• Significance: Many web-based tasks, such as product research or fact-checking, require iterative data gathering, wrangling, and analysis. Current fragmented workflows impose high cognitive loads and inefficiencies, making it difficult for users to confidently make decisions.
• Motivation and related work: Prior tools focus on specific stages of data management or rely on chatbot-like UIs, which are insufficient for iterative, data-driven workflows. While some systems visualize decision-making processes, they often lack interactivity and user control. This paper addresses the gap by proposing a unified, data-centric approach to human-AI collaboration.

Solution

• Proposed approach: WebSeek, a mixed-initiative browser extension that enables users to interact with tangible data artifacts (e.g., tables, visualizations) on an interactive canvas, supported by proactive and reactive AI guidance.
• Novelty:
  1. Introduction of a data-centric paradigm where data artifacts are treated as first-class citizens for interaction.
  2. A principled design framework for proactive and reactive AI guidance tailored to web-based data tasks.
  3. Implementation of a browser extension integrating direct manipulation, AI suggestions, and chat-based assistance.
  4. Empirical insights from a user study on how users interact with AI guidance and tangible data artifacts.
• Procedure and key techniques:
  • Users can manually create and edit data instances or interact with AI guidance (proactive or reactive) for tasks like data extraction, wrangling, and visualization.
  • Proactive guidance includes micro (in-situ) and macro (peripheral) suggestions triggered by user actions.
  • Reactive guidance is provided through a chat interface, allowing users to issue commands or refine tasks.
  • AI guidance is generated using LLMs with context from HTML, user interactions, and data instances.

Results

• Concrete findings:
  • Technical evaluation showed high accuracy (97.2%) for AI guidance, with average latencies of 11.44–19.05 seconds for in-situ suggestions and 3.35–3.55 seconds for peripheral suggestions.
  • User study (N=15) reported high usability (SUS score: 73.11/100) and confidence (5.33–5.67/7) in task outcomes.
  • Participants preferred direct manipulation (67.1% of time) and in-situ guidance (82.8% sessions) over peripheral suggestions (13.8% sessions).
• Advantage over baselines:
  • Unified environment for data extraction, wrangling, and visualization reduced context-switching compared to traditional workflows.
  • Tangible data artifacts and transparent AI guidance enhanced user control and confidence.
• Experiments / evaluation:
  • Technical evaluation: 50 benchmark tasks across 3 difficulty levels (Easy, Medium, Hard) using simulated users.
  • User study: Two tasks (fact-checking and product comparison) involving diverse data analysis workflows.
• Limitations and future work:
  • Scalability issues with LLM context management for large datasets.
  • Canvas usability may degrade with a high number of data artifacts.
  • Limited participant diversity and scope of tasks in the user study.
  • Need for better integration with native web interactions and persistent data provenance.

Summary

WebSeek is a mixed-initiative browser extension that reimagines web-based data-driven decision-making by enabling users to interact with tangible data artifacts on an interactive canvas. It combines direct manipulation with proactive and reactive AI guidance, supported by a principled design framework. Technical evaluations and user studies demonstrated its effectiveness in improving task efficiency, user confidence, and control. The work advocates for a shift from conversational delegation to data-centric collaboration in web agents, opening new possibilities for human-AI interaction in data analysis and decision-making. Future work will address scalability, integration with native web interactions, and broader task applicability.