Not with the bot! The relevance of trust to explain the acceptance of chatbots by insurance customers

The paper investigates policyholders’ acceptance of chatbots in the insurance sector, a key area affected by Industry 4.0 and insurtech. As insurers digitize processes and expand digital channels, conversational bots are increasingly used for policy management and claims communication. The study focuses on the critical post-purchase stage of the policy lifecycle, where policyholders interact with insurers to manage active policies, especially for claims. Grounded in the Technology Acceptance Model (TAM) and recognizing the heightened importance of trust in AI-mediated services, the research addresses two questions: (RQ1) What are customers’ average intention to use and attitude toward using chatbots for managing existing policies? (RQ2) What factors drive intention to use and attitude toward chatbot-assisted interactions?

The study builds on TAM (Davis, 1989) and related models (TRA, UTAUT) widely used to explain technology acceptance in fintech, insurance, and chatbot contexts. Prior research shows attitude mediates effects of perceived usefulness (PU) and perceived ease of use (PEOU) on behavioral intention (BI). In chatbot and service automation settings, PU and PEOU generally improve attitude and intention. However, chatbots’ current limitations (e.g., handling complex requests, lack of empathy, conversational constraints) can dampen perceived benefits and ease. Trust emerges as a pivotal construct for AI-mediated services, especially in insurance where moral hazard and adverse selection make trust foundational. Trust is conceptualized with cognitive (technology effectiveness) and relational (confidence in the insurer using chatbots to enhance service) dimensions; emotional trust is noted but considered less central for routine policy interactions. Prior studies report trust can directly affect attitude and BI and also serve as an antecedent to PU and PEOU.

Design: Cross-sectional survey using a structured questionnaire in Spanish. The instrument was pretested with 15 Spanish insurance professionals, refined, and then piloted with 12 additional volunteers before full deployment. Sample: Policyholders in Spain. Demographics indicate approximately 226 respondents (about 53% male, 44% female), skewed toward higher education (87% graduate or beyond). Many responses were collected via professional social networks. Measures: Five latent constructs measured with multi-item Likert scales: - Behavioral intention (BI): 3 items (e.g., likelihood to use chatbots for insurance procedures). - Attitude (ATT): 4 items (favorability toward managing policies with chatbots). - Perceived usefulness (PU): 4 items (e.g., improved service, lower costs, higher coverage). - Perceived ease of use (PEOU): 4 items (effort, clarity, accessibility, ease of communication). - Trust (TRUST): 3 items (trustworthiness, keeping promises, policyholder interests). Analysis: RQ1 assessed by comparing means/medians of BI and ATT items to a neutral value of 5 (Student’s t-tests for means; Wilcoxon tests for medians). RQ2 tested via PLS-SEM (SmartPLS 4). Reliability and validity checks included Cronbach’s alpha, composite reliability (CR), rho_A, average variance extracted (AVE), item loadings (>0.7), Fornell–Larcker criterion, and HTMT (<0.90). Structural paths estimated with bootstrapping (5,000 subsamples). Model fit assessed via R² and predictive relevance (Stone–Geisser’s Q²) and cross-validated predictive ability (RMSE, MAE, ALD).

RQ1: Acceptance levels are low. Across BI and ATT items, means and medians are significantly below the neutral value of 5 (p < 0.01), indicating reluctance to use chatbots for policy procedures. Measurement model: All scales exhibit strong reliability and validity (Cronbach’s alpha, CR, rho_A > 0.7; AVE > 0.5; loadings > 0.7; Fornell–Larcker and HTMT criteria satisfied). Structural model: R² values indicate substantial-to-moderate explanatory power—ATT (R² = 0.717), BI (R² = 0.604), PEOU (R² = 0.626), and PU (R² ≈ 0.629). Key path coefficients (β) and significance: - ATT → BI: β = 0.675, t = 24.597, p < 0.01 (accepted). - PU → ATT: β = 0.166, t = 2.558, p < 0.05 (accepted). - PEOU → ATT: β = 0.413, t = 4.927, p < 0.01 (accepted). - PEOU → PU: β = 0.138, t = 1.664 (not significant). - TRUST → ATT: β = 0.335, t = 4.038, p < 0.01 (accepted). - TRUST → BI: β = 0.675, t = 24.446, p < 0.01 (accepted). - TRUST → PEOU: β = 0.812, t = 35.236, p < 0.01 (accepted). Mediation effects on BI (indirect): PU (coef 0.129, t = 2.691, p < 0.05), PEOU (coef 0.339, t = 5.942, p < 0.01), TRUST (coef 0.623, t = 16.537, p < 0.01). Predictive assessment shows positive Q² values and significant cross-validated predictive ability across constructs (p < 0.01). Overall, trust is the most influential factor for BI through direct and mediated paths.

The study confirms widespread reluctance among policyholders to use chatbots for managing existing policies, aligning with prior reports of customer hesitancy in insurance contexts. Despite low acceptance, a TAM-based framework augmented with trust explains substantial variance in attitude and behavioral intention. Attitude robustly drives intention, while PU and PEOU significantly enhance attitude; however, PEOU does not significantly increase PU in this context. Trust is pivotal: beyond its direct effect on attitude and intention, it strongly improves perceived ease of use and contributes to perceived usefulness, underscoring both cognitive and relational dimensions of trust in insurer–policyholder interactions mediated by AI. These results emphasize that improving chatbot usability and demonstrable utility must be coupled with strategies that build customer trust to increase acceptance, particularly given chatbots’ current limitations with complex or sensitive issues.

The paper demonstrates that policyholders’ low acceptance of chatbots for managing active insurance policies can be effectively explained by TAM constructs—perceived usefulness and ease of use—augmented with trust, which emerges as the most critical driver of attitude and intention. The model exhibits strong explanatory and predictive performance. Practically, insurers should prioritize trust-building, alongside enhancing chatbot utility and usability, to mitigate reluctance. Future research should: (1) incorporate the emotional dimension of trust for contexts involving high emotional stakes (e.g., life and health insurance claims), (2) apply and test the model across other phases of the insurance lifecycle (advisory, pricing) and with different stakeholder groups (e.g., professionals), and (3) extend to multi-country, diverse samples to assess generalizability and track longitudinal changes as conversational AI capabilities evolve.

Single-country (Spain) and sampling via professional networks likely skew the sample toward higher education and socioeconomic status, limiting generalizability. The cross-sectional design precludes assessment of changes over time in a fast-evolving technology landscape. The study operationalizes trust via cognitive and relational dimensions and does not model emotional trust, which may matter in high-sensitivity scenarios. Measurement and reporting include minor inconsistencies in descriptive tables, though reliability and validity criteria are met.