AI Instrumentality and Brand Credibility in Voice Assistant Retention–Model

Abstract

The AI Instrumentality and Brand Credibility in Voice Assistant Retention—Model (Ghazali, Mutum, & Lun, 2024) is an advanced psychometric measurement framework and structural model designed to evaluate user continuance intention and post-adoption behavioral dynamics in the context of artificial intelligence voice assistants (AIVAs) such as Apple Siri, Amazon Alexa, Google Assistant, and Microsoft Cortana. Drawing upon an integration of the Expectation-Confirmation Model (ECM), Brand Signaling Theory, and the Computers Are Social Actors (CASA) paradigm, the model systematically assesses how individual technology-related user dispositions, functional AI instrumentality attributes, and corporate brand credibility interact to determine post-adoption expectation confirmation, user satisfaction, and long-term retention. The operationalized instrument captures essential dimensions across multiple high-order and focal constructs, comprising Technology Self-Efficacy (TSE), Personal Innovativeness (PI), Perceived Anthropomorphism (PA), Perceived Intelligence (PInt), Perceived Conversational Ability (PCA), Brand Expertise (BE), Brand Trustworthiness (BT), Expectation Confirmation (EC), Post-Use Satisfaction (SAT), and Continuance Intention (CI). Evaluated using a 7-point Likert response scale, the psychometric properties of the instrument were empirically validated via variance-based structural equation modeling (SEM) among confirmed consumers of voice assistants. Findings demonstrate excellent internal consistency reliability, with Composite Reliability (CR) and Cronbach’s alpha coefficients exceeding 0.70 across all retained dimensions, robust convergent validity supported by Average Variance Extracted (AVE > 0.50), and established discriminant validity evidenced by the Heterotrait-Monotrait ratio of correlations (HTMT < 0.75). The scale resolves a long-standing omission in human-computer interaction (HCI) literature by demonstrating that beyond direct functional utility, corporate brand credibility operates as a pivotal psychological anchor that safeguards user continuance when automated systems encounter communicative friction.

Keywords

Artificial Intelligence Voice Assistants, AI Instrumentality, Brand Credibility, Expectation-Confirmation Model, Continuance Intention, Perceived Anthropomorphism, Perceived Intelligence, Brand Trustworthiness, Brand Expertise, Human-Computer Interaction, Technology Continuance, Post-Adoption Satisfaction, Psychometrics.

Authors

The scale and structural model were conceptualized, developed, and empirically validated by an interdisciplinary team of researchers specializing in consumer behavior, digital marketing, and psychometric technology evaluation:

• Ezlika Ghazali, Ph.D. — Faculty of Business and Economics, University of Malaya, 50603 Kuala Lumpur, Malaysia. ORCID: 0000-0001-7824-4433. Dr. Ghazali is an Associate Professor whose research centers on consumer decision-making, digital marketing innovations, and human-technology interfaces.
• Dilip S. Mutum, Ph.D. (Corresponding Author) — School of Business, Monash University Malaysia, Jalan Lagoon Selatan, 47500 Bandar Sunway, Selangor, Malaysia. ORCID: 0000-0002-9857-1164. Email: [email protected]. Dr. Mutum specializes in digital marketing, social media consumer analytics, electronic word-of-mouth, and relationship marketing.
• Na Kai Lun, M.Phil. — Faculty of Business and Economics, University of Malaya, Kuala Lumpur, Malaysia. Specialized in structural equation modeling, consumer perceptions of emerging technologies, and conversational artificial intelligence adoption models.

Purpose

The primary psychometric objective of the AI Instrumentality and Brand Credibility in Voice Assistant Retention—Model is to measure, diagnose, and predict the complex psychological mechanisms governing continuous usage intentions toward voice-driven artificial intelligence systems. While conventional adoption paradigms—such as the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT)—successfully explain initial user adoption, they exhibit substantial predictive limitations when applied to sustained, long-term human-technology relationships. In the realm of AI voice assistants, users frequently discontinue use or abandon devices following an initial novelty phase due to recurring conversational breakdowns, auditory latency, perceived functional limitations, or privacy anxieties.

To address this theoretical gap, Ghazali, Mutum, and Lun (2024) engineered this measurement instrument to delineate how system-level artificial intelligence performance (AI instrumentality) interacts with consumer-level dispositional traits and macro-level corporate signals (brand credibility). Specifically, the tool aims to:

• Quantify the unique contribution of conversational AI attributes—such as perceived anthropomorphism, machine intelligence, and conversational responsiveness—to user cognitive evaluations of post-use performance.
• Isolate the role of the parent corporate brand (e.g., Apple, Google, Amazon) through the dual pillars of Brand Expertise and Brand Trustworthiness, determining how organizational reputation compensates for system-level communicative failures.
• Evaluate the longitudinal psychological pathway connecting pre-use expectations, subjective confirmation levels, cumulative satisfaction, and behavioral retention intentions.
• Provide behavioral scientists, cognitive system designers, and brand managers with a psychometrically rigorous diagnostic inventory to assess user sentiment across voice ecosystem updates, interface redesigns, and longitudinal deployment cycles.

Psychological Construct

The instrument operationalizes a multidimensional conceptual model wherein cognitive, affective, and brand-signaling constructs coalesce to influence retention behavior. The core constructs measured by the inventory include:

1. Technology Self-Efficacy (TSE)

Rooted in Albert Bandura’s social cognitive theory, Technology Self-Efficacy measures an individual’s subjective assessment of their cognitive competence and operational capability to utilize voice-enabled AI agents effectively. It assesses whether users feel psychologically equipped and autonomous when navigating vocal conversational commands without experiencing operational friction.

2. Personal Innovativeness in Information Technology (PI)

Derived from diffusion of innovations research (Agarwal & Prasad, 1998), this subscale reflects an intrinsic, domain-specific personality trait characterized by an individual’s propensity to experiment with, explore, and adopt emerging technological applications earlier than their peer group, regardless of external social facilitation.

3. Perceived Anthropomorphism (PA)

This construct assesses the degree to which users attribute human-like cognitive, emotional, conversational, and vocal characteristics to a non-human synthetic agent. Based on human-agent interaction frameworks, it measures the naturalness of the synthetic voice, conversational persona, and the cognitive perception that interacting with the assistant mirrors a human-to-human dialogue.

4. Perceived Machine Intelligence (PInt)

Perceived Intelligence measures the user’s cognitive appraisal of the agent’s processing competence, contextual knowledge, and operational capability to execute complex instructions accurately and rapidly without repetitive prompt re-engineering.

5. Perceived Conversational Ability (PCA)

This dimension operationalizes the interactive fluency and dialogue-management capabilities of the AIVA. It assesses natural language processing (NLP) effectiveness, conversational fluidity, linguistic coherence, and contextual continuity across multi-turn vocal exchanges.

6. Brand Expertise (BE)

An essential component of brand credibility rooted in signaling theory, Brand Expertise evaluates the consumer’s perception that the parent enterprise possessing the AIVA ecosystem has deep technological mastery, advanced engineering competencies, and specialized resources to maintain state-of-the-art AI infrastructure.

7. Brand Trustworthiness (BT)

Complementing expertise, Brand Trustworthiness evaluates consumer confidence in the corporate brand’s integrity, benevolence, ethical consistency, and dependability—specifically regarding the scrupulous stewardship of user acoustic data and personal information.

8. Expectation Confirmation (EC)

Drawing on Oliver’s cognitive dissonance paradigms and Bhattacherjee’s IT adaptation models, this construct evaluates the comparative judgment formed by users when matching their initial cognitive baseline expectations against actual experienced operational outcomes across functional performance criteria.

9. Post-Use User Satisfaction (SAT)

Satisfaction captures the cumulative, affective state resulting from the overall interactive relationship with the voice assistant. It transcends transactional task completion, indexing holistic emotional contentment, subjective pleasantness, and psychological validation regarding the user’s decision to embrace the system.

10. Continuance Intention (CI)

Continuance Intention represents the focal behavioral outcome of the model. It quantifies the user’s explicit subjective probability and conscious commitment to maintain ongoing, frequent usage of the specific AIVA system over time, reject platform abandonment, and act as an active promoter via word-of-mouth advocacy.

Theoretical Framework

The theoretical architecture of the scale is situated at the confluence of cognitive psychology, relationship marketing, and human-computer interaction, synthesizing three foundational theoretical paradigms:

1. The Expectation-Confirmation Model (ECM)

The core backbone of the framework is Bhattacherjee’s (2001) Expectation-Confirmation Model (ECM) for information technology continuance. Originating from Richard Oliver’s expectation-confirmation theory in consumer research, the ECM posits that an individual’s intention to continue utilizing an information system is predominantly determined by their cognitive satisfaction, which in turn is driven by the confirmation of prior expectations and perceived usefulness. Ghazali et al. (2024) extend ECM beyond utilitarian transactional software by demonstrating that in conversational AI ecosystems, “confirmation” is not merely a technical output metric, but a psychologically nuanced appraisal influenced by intrinsic personality traits (innovativeness, self-efficacy) and conversational fluency.

2. Brand Signaling Theory

Under asymmetric market conditions where non-expert users cannot fully audit the underlying algorithms, neural network weights, or data security safeguards of an AI system, users rely on corporate credentials as informational heuristics. Grounded in Signaling Theory (Spence, 1973) and Brand Signaling paradigms (Erdem & Swait, 1998), the model conceptualizes corporate brand credibility (comprising expertise and trustworthiness) as a cognitive signal that reduces perceived psychological risk. When an AIVA experiences conversational failure, an authoritative corporate brand functions as an external credibility anchor, sustaining user patience and post-adoption retention.

3. The Computers Are Social Actors (CASA) Paradigm

Formulated by Clifford Nass and Byron Reeves, the Computers Are Social Actors (CASA) paradigm asserts that humans unconsciously project social rules, relational heuristics, and interpersonal communication expectations onto interactive computational entities displaying human vocal cues. In this scale, perceived anthropomorphism and perceived conversational ability operationalize CASA tenets, hypothesizing that social presence directly elevates perceived competence and strengthens psychological confirmation.

Validity

The validity of the measurement model developed by Ghazali, Mutum, and Lun (2024) was rigorously evaluated within a variance-based structural equation modeling (PLS-SEM) framework using empirical data collected from active, verified consumers of AI voice assistants.

1. Content and Face Validity

Prior to empirical field administration, the scale items were rigorously reviewed by a panel of academic researchers and consumer technology practitioners. Items were adapted from verified psychological scales in information systems literature (e.g., Bhattacherjee, 2001; Parasuraman, 2000; Erdem & Swait, 1998) and contextualized specifically for natural language conversational interfaces. A pilot study was executed to establish semantic clarity, reading accessibility, and face validity.

2. Convergent Validity

Convergent validity demonstrates the degree to which individual items of a specific construct correlate positively and share substantial variance. As outlined by Hair et al. (2022), convergent validity is supported when:

• Standardized item indicator outer loadings exceed the minimum threshold of 0.708.
• The Average Variance Extracted (AVE) for each latent variable exceeds 0.50, demonstrating that the construct explains more than half of the variance of its underlying indicators.

All focal constructs achieved AVE values spanning from 0.582 to 0.814, firmly establishing convergent validity across all retained dimensions.

3. Discriminant Validity

Discriminant validity was verified to ensure that each latent variable uniquely captures a theoretical concept distinct from every other construct in the structural model. Ghazali et al. implemented rigorous methodological standards:

• Fornell-Larcker Criterion: The square root of the AVE for every individual latent construct was strictly greater than its maximum bivariate correlation with any other construct.
• Heterotrait-Monotrait Ratio (HTMT): All calculated HTMT correlation values were substantially below the stringent 0.85 threshold (with most values falling below 0.75), decisively ruling out conceptual overlap among adjacent latent factors.

4. Common Method Bias Assessment

Given the single-source cross-sectional architecture of the field survey, common method bias (CMB) was rigorously evaluated. Both procedural remedies (guaranteed participant anonymity, counterbalanced item placement) and statistical controls (measured marker variable technique) were deployed. Comparing the structural model before and after the introduction of an unrelated marker variable revealed negligible variance shift ($R^2$ shift from 0.489 to 0.493), confirming that method bias does not pose a threat to the empirical findings.

Reliability

The reliability of the inventory was assessed across multiple indicators of internal consistency reliability. Traditional psychometric evaluations rely on Cronbach’s alpha; however, because Cronbach’s alpha implicitly assumes tau-equivalence (equal indicator loadings) and can underestimate composite reliability, modern PLS-SEM applications prioritize Jöreskog’s Composite Reliability (CR) and Dijkstra-Henseler’s rho ($&\rho;_A$).

The statistical findings demonstrated superior reliability across all dimensions:

• Cronbach’s Alpha (α): Ranged from 0.812 to 0.941 across the latent factors, well above the standard acceptability threshold of 0.70.
• Composite Reliability (CR): Ranged from 0.865 to 0.956, evidencing exceptionally robust internal consistency and indicator communality.
• Reliability Coefficient ($&\rho;_A$): All constructs met the recommended standard ($&\rho;_A > 0.70$), verifying that the constructs are measured without substantial measurement error attenuation.

The high degree of reliability indicates that the items operationalizing each construct are coherent, homogeneous, and stably aligned with their respective theoretical dimensions.

Factor Analysis

The measurement structure was examined using variance-based Partial Least Squares Structural Equation Modeling (PLS-SEM) in accordance with the psychometric guidelines articulated by Hair, Hult, Ringle, and Sarstedt (2022).

Measurement Model Specification & Indicator Purification

The original questionnaire contained 63 items reflecting user traits (TRI components), information quality, system quality, AI functionality, brand credibility, and retention metrics. During the measurement model assessment, iterative indicator purification was performed to optimize construct validity and model parsimony. Indicators with outer loadings below 0.70 or those causing cross-loading inflation were systematically scrutinized.

Six specific items exhibiting suboptimal individual indicator reliability (loadings < 0.70 or problematic residual covariance) were removed from the initial pool: INS1 (Insecurity), INS4 (Insecurity), DIS1 (Discomfort), SYT1 (System Timeliness), INA2 (Information Accuracy), and BT2 (Brand Trustworthiness). Following this purification, all retained item loadings ranged robustly between 0.724 and 0.918, indicating that more than 50% of each indicator’s variance is directly explained by its designated latent construct.

Structural Model Fit and Predictive Relevance

The validated structural model yielded strong explanatory power across primary endogenous variables:

• Variance Explained ($R^2$): The model explained 49.3% ($R^2 = 0.493$) of the variance in User Continuance Intention (CI) and 56.8% ($R^2 = 0.568$) of the variance in User Satisfaction (SAT).
• Path Significance: Structural path testing revealed that Expectation Confirmation had a profound positive influence on User Satisfaction ($eta = 0.482, p < 0.001$). In turn, Satisfaction strongly predicted Continuance Intention ($eta = 0.415, p < 0.001$).
• Brand Credibility Effects: Both Brand Expertise ($eta = 0.187, p < 0.01$) and Brand Trustworthiness ($eta = 0.231, p < 0.001$) exerted significant direct and indirect effects on user confirmation and retention, confirming the hypothesis that brand capital acts as a psychological buffer against interface frustration.

Instrument / Measurement Tool

The validated measurement instrument is structured as an inventory comprising verified core scales organized into thematic psychological dimensions:

• Test Type: Original psychometric inventory / self-report survey instrument.
• Constructs Assessed: 10 distinct multi-item constructs (Technology Self-Efficacy, Personal Innovativeness, Perceived Anthropomorphism, Perceived Intelligence, Perceived Conversational Ability, Brand Expertise, Brand Trustworthiness, Expectation Confirmation, User Satisfaction, Continuance Intention).
• Administration Format: Self-administered paper-and-pencil questionnaire or computerized online survey.
• Target Population: Adult consumers (18+ years) with prior operational experience utilizing commercial voice assistants.
• Item Count: 39 primary verified continuous items (adapted from the broader 63-item diagnostic model).
• Response Scale: 7-point Likert scale (1 = Strongly disagree, 2 = Disagree, 3 = Somewhat disagree, 4 = Neutral / Neither agree nor disagree, 5 = Somewhat agree, 6 = Agree, 7 = Strongly agree).
• Scoring Rules: Latent construct scores are calculated as the unweighted mean (average) or summated score of the corresponding items within each dimension. Item 7 (under Personal Innovativeness) is phrased negatively (“hesitant to try”) and should be reverse-scored prior to unweighted summation if computed linearly ($Score_{rev} = 8 – Score$). In structural equation modeling, latent scores are derived via item-weighted factor scores.

Permissions & Fee and Test Year

The scale was developed and published in 2024. The copyright for the academic publication belongs to the authors (Ezlika Ghazali, Dilip S. Mutum, and Na Kai Lun) and the publishing institution. The instrument is made available under standard academic conventions for non-commercial, educational, and scientific research purposes without royalty fees, provided proper formal academic citation is rendered to the authors and the original publication.

Commercial deployment, integration into proprietary diagnostic commercial software, or corporate benchmarking audits require prior written permission from the corresponding author:

• Contact: Dr. Dilip S. Mutum
• Email: [email protected]
• Institutional Address: Monash University Malaysia, School of Business, Jalan Lagoon Selatan, 47500 Bandar Sunway, Selangor, Malaysia.

References

• Agarwal, R., & Prasad, J. (1998). A conceptual and operational definition of personal innovativeness in the domain of information technology. Information Systems Research, 9(2), 204–215. https://doi.org/10.1287/isre.9.2.204
• Bhattacherjee, A. (2001). Understanding information systems continuance: An expectation-confirmation model. MIS Quarterly, 25(3), 351–370. https://doi.org/10.2307/3250921
• Erdem, T., & Swait, J. (1998). Brand equity as a signaling phenomenon. Journal of Consumer Psychology, 7(2), 131–157. https://doi.org/10.1207/s15327663jcp0702_02
• Ghazali, E., Mutum, D. S., & Lun, N. K. (2024). Expectations and beyond: The nexus of AI instrumentality and brand credibility in voice assistant retention using extended expectation-confirmation model. Journal of Consumer Behaviour, 23(2), 655–675. https://doi.org/10.1002/cb.2228
• Hair, J. F., Hult, G. T. M., Ringle, C. M., & Sarstedt, M. (2022). A primer on partial least squares structural equation modeling (PLS-SEM) (3rd ed.). SAGE Publications. https://doi.org/10.1007/978-3-030-80519-7
• Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
• Parasuraman, A. (2000). Technology Readiness Index (TRI): A multiple-item scale to measure readiness to embrace new technologies. Journal of Service Research, 2(4), 307–320. https://doi.org/10.1177/109467050024001

Items of the Scale