Supporting Construction Worker Well-Being with a Multi-Agent Conversational AI System

1. INTRODUCTION

• Construction workers face high injury and mental health risks, while stigma and demanding work can discourage help-seeking.The reported suicide rate is 44% higher than the national average.Source: §1 · p. 1
• General-purpose AI may not provide the industry-specific safety information and experience-resonant support construction workers need.AI chatbots can offer accessible, immediate assistance and anonymity, but construction-tailored solutions remain limited.Source: §1 · p. 1
• The proposed system uses LLMs and retrieval-augmented generation to address construction workers’ support needs.Source: §1 · p. 1
• Specialized agents offer complementary perspectives through a group conversation where agents can choose whether and when to respond.Agents have distinct goals, personalities, and domain-specific knowledge.Source: §1 · pp. 1–2
• The controlled comparison evaluates usability, psychological needs, social presence, and trust against a single-agent baseline.Source: §1 · p. 2

2.1 Emotional Health Support for Construction Workers

• Construction mental health interventions include peer support, training, and digital services, but digital approaches can suffer from low engagement and limited industry tailoring.Source: §2.1 · p. 2

2.2 Conversational AI in Construction

• Conversational AI supports construction tasks, while generic chatbots may lack construction context; domain-specific well-being support remains an open research question.Conversational AI deployment in AEC is still relatively slow.Source: §2.2 · p. 2

2.3 Multi-agent System

• Multi-agent systems divide work among specialized agents, with cited strengths including parallel processing, robustness, verification, and adaptive control.The text attributes parallel processing to Shu et al. (2024), robustness to Chaaban and Müller-Schloer (2013), verification to Amirkhani and Barshooi (2022), and adaptive control to Rizk et al. (2018).Source: §2.3 · pp. 2–3
• Despite LLMs’ promise for multi-agent systems, human interaction with LLM-based conversational systems remains underexplored; this study investigates one for construction workers.The authors describe it as the first investigation of conversational multi-agent systems in construction.Source: §2.3 · p. 3

3.1 Overview

• The overview presents a RAG-based multi-agent system in which user messages are processed using a vector database and configurable external documentation.Source: §3.1 · p. 3, Figure 1 · p. 3
• Agents specialize in domains such as regulatory explanation or emotional support and combine their perspectives in multi-faceted responses.Source: §3.1 · p. 3
• The following sections cover interface design, agent configuration, orchestration, and evaluation against a baseline.Source: §3.1 · p. 3

3.2 User Interface Design

• The React-based interface simulates a group conversation between the user and multiple AI agents.Source: §3.2 · p. 3, Figure 2 · p. 4
• A dynamic management panel lets users add or delete active agents and configure their names, personalities, goals, and private knowledge.Each agent’s settings are represented by a customizable avatar.Source: §3.2 · pp. 3–4
• The conversation area visually distinguishes user messages from agent messages.Source: §3.2 · p. 4

3.3 Agent Configuration

• A two-tier prompt-chaining process automatically generates consistent agent descriptions from keywords such as occupation, personality, and conversation goals.This reduces the manual effort required for prompt engineering.Source: §3.3 · p. 4
• RAG retrieves relevant information from a knowledge base before an agent generates a response.Documents are chunked and indexed in a FAISS vector database using OpenAI semantic embeddings.Source: §3.3 · p. 4
• During conversation, agents dynamically retrieve domain-specific knowledge based on the user's message while maintaining natural dialogue flow.The system avoids explicit references to the knowledge source in its responses.Source: §3.3 · p. 4

3.4 Multi-Agent Orchestration

• The orchestration system determines both who should respond and the order of responses.Source: §3.4 · p. 4
• Agents respond when directly addressed or when their expertise is relevant, and the system ensures at least one agent answers each user message.A message addressed to Alice may receive only her response, while a general question may prompt all agents.Source: §3.4 · pp. 4–5
• Agents are randomly grouped for sequential or parallel responses; sequential turns can build on or disagree with earlier responses, while parallel replies offer simultaneous perspectives and can reduce latency.Agents share conversation history but keep their external knowledge private, allowing them to build on prior replies while retaining distinct perspectives.Source: §3.4 · p. 5

3.5.1 Personas of agents

• Three agents represented occupational safety and health, human resources, and worker-peer support, with distinct domain knowledge and complementary roles.The baseline used the same interface and GPT-4o but lacked predefined personas and external documentation.Source: §3.5.1 · p. 5

3.5.2 Scenarios settings

• Participants explored three structured scenarios: inadequate protective equipment, mental-health struggles after workplace injury, and burnout with limited job recognition.Participants retained control over the direction and content of their conversations within their assigned scenario.Source: §3.5.2 · pp. 5–6

3.5.3 User Study Procedure

• Twelve participants each used both systems in a randomly assigned scenario, with system order randomized to mitigate learning effects and surveys completed after each interaction.Surveys measured SUS, autonomy, competence, relatedness, perceived social presence, trust, adoption intent, and open-ended feedback.Source: §3.5.3 · p. 6

4.1 SUS Results and Analysis

• Normalized SUS scores were higher for the proposed system (Mean = 84.58, SD = 8.95) than the baseline (Mean = 71.88, SD = 16.46; p = .0083).The system fell in the SUS ‘A Grade’ range, and 11 of 12 participants preferred it.Source: §4.1 · pp. 6–7, Figure 4 · p. 7

4.2 Self-Determination Results and Analysis

• Autonomy, competence, and relatedness scores were higher for the multi-agent system; relatedness was 3.83 versus 2.08 (p < 0.001).Competence was 3.86 versus 3.25 (p = 0.048), while autonomy was 4.14 versus 3.50 (p = 0.066); participants also described greater flexibility, confidence, and connection.Source: §4.2 · pp. 7–8, Figure 5 · p. 8

4.3.1 Reliability analysis of the scale

• The three-item Perceived Social Presence and Trust scale showed excellent reliability for the proposed system (α = 0.911) and good reliability for the baseline (α = 0.817).Source: §4.3.1 · p. 8

4.3.2 Results analysis

• The proposed system scored higher than the baseline on all three social-presence and trust items: Q1 3.33 versus 1.42 (p < .001), Q2 4.00 versus 2.92 (p = .013), and Q3 4.00 versus 2.75 (p = .002).Participants associated the system's agent identities, professional profiles, and domain-specific expertise with higher ratings.Source: §4.3.2 · pp. 8–9, Figure 6 · p. 9

• The system showed exceptional usability and enhanced workers’ psychological needs, social presence, and trust compared with a generic baseline chatbot.Source: §5 · p. 9
• The system lets non-experts customize the LLM and integrate external knowledge.Source: §5 · p. 9
• The framework is presented as offering credible safety and mental-health guidance through more human-like, empowering, and socially engaging interaction.Source: §5 · p. 9
• The small participant group and predefined scenarios make this an initial exploration; future work will study more diverse users, including real construction workers, and individualized scenarios.Source: §5 · p. 9