Artificial Intelligence and Consumer Decision-Making: AI-Powered Personalization, Trust Pathways, and Purchase Intentions in Nigeria ()
1. Introduction
The migration of consumer decision-making onto algorithmically curated digital surfaces represents one of the most consequential shifts in contemporary marketing practice. Artificial intelligence (AI)-powered personalization the real-time, machine-learning-driven tailoring of product recommendations, content, and offers to inferred individual preferences throughout, the focal “AI agent” refers to the algorithmic recommender system embedded within a consumer-facing online retail or marketplace platform that is, the machine-learning engine that generates personalized product recommendations and offers on e-commerce apps and websites rather than a conversational chatbot, a general-purpose platform ranking algorithm, or a single retailer’s manual merchandising. has moved from a competitive embellishment to the organizing logic of digital commerce (Davenport et al. [1]; De Bruyn et al. [2]). Huang and Rust [3] frame AI as a strategic capability that reorganizes the marketing function itself, and Verhoef et al. [4] and Grewal et al. [5] situate that shift within the wider digital transformation of the firm, while the systematic reviews of Mariani et al. [6] and Vlačić et al. [7] document the speed at which consumer-facing AI research has accumulated and Dwivedi et al. [8] map the multidisciplinary challenges it raises. Yet the dominant theoretical posture in the marketing literature continues to frame personalization as a broadly relevance-enhancing, and therefore trust- and conversion-promoting, intervention (Kumar et al. [9]; Shankar [10]). This framing obscures a more unsettled reality: the same algorithmic apparatus that increases offer relevance simultaneously heightens consumers’ awareness of surveillance and inference, generating a tension that personalization scholarship has long labelled the personalization-privacy paradox (Aguirre et al. [11]; Bleier and Eisenbeiss [12]; Aguirre et al. [13]). Puntoni et al. [14] further show that consumers do not experience AI as a neutral utility but as a series of felt encounters that can be enabling and unsettling at once. The strategic question is therefore not whether personalization works, but under what conditions, through which psychological mechanisms, and for whom.
Three limitations constrain the cumulative knowledge base. First, the mediating role of trust is routinely modelled as a single, undifferentiated construct, despite well-established evidence in the trust literature that consumer trust in an unfamiliar online actor is multidimensional, comprising a calculative, competence-based component and an emotional, benevolence-based component (McKnight et al. [15] [16]; Mayer et al. [17]; Johnson and Grayson [18]). Collapsing these into one latent variable conceals the possibility that personalization recruits them through different routes and that they respond to different boundary conditions. Second, the boundary conditions themselves are under-theorized. Personalization research has tended to treat algorithmic competence as a constant rather than a perception that consumers actively form and that conditions their willingness to defer to machine judgement a perception that the algorithm-appreciation and algorithm-aversion literatures show to be both consequential and variable (Logg et al. [19]; Dietvorst et al. [20]; Castelo et al. [21]), and that Longoni et al. [22] show to vary sharply with the domain in which the machine is asked to judge. Third, the empirical canon is overwhelmingly situated in high-institutional-trust Western markets, where formal data-protection regimes underwrite a baseline of structural assurance. Whether the trust calculus transfers to settings with weak enforcement infrastructure remains largely an assumption rather than a finding.
Nigeria offers a theoretically generative site for confronting these limitations. As Africa’s most populous nation and largest consumer economy, it combines accelerating mobile-commerce penetration with a comparatively nascent data-governance environment, producing a configuration in which consumers must extend trust to AI systems in the relative absence of the institutional guarantees that anchor trust in Western contexts (Sheth [23]; Hoffman and Novak [24]). This is not a context to which existing models can be presumed to generalize; it is one in which their scope conditions become visible. We therefore ask: through which trust pathways does AI-powered personalization shape purchase intention among Nigerian online consumers, and how do perceived AI capability and privacy risk perception condition those pathways?
We answer this question by developing and testing a capability-trust contingency model. The model makes three moves that distinguish it from prior work. It decomposes the personalization-to-purchase mechanism into parallel cognitive- and affective-trust pathways; it specifies perceived AI capability and privacy risk perception as theoretically targeted moderators of distinct paths, yielding a moderated-mediation structure rather than an undifferentiated main effect; and it relocates the inquiry to an emerging-market setting where the institutional preconditions of trust differ in kind. In doing so, the study advances consumer AI theory from a question of average effects toward a contingency account of when and why algorithmic personalization converts attention into commitment.
2. Theoretical Background and Hypothesis Development
2.1. AI-Powered Personalization as a Double-Edged Stimulus
AI-powered personalization differs from earlier rule-based personalization in degree and in kind. Where legacy systems matched offers to declared preferences, contemporary systems infer latent intentions from behavioural traces, continuously updating their models without explicit consumer input (Davenport et al. [1]; De Bruyn et al. [2]). Kumar et al. [9] describe this capability as the foundation of personalized engagement marketing and Shankar [10] traces its diffusion through retailing, while Xiao and Benbasat [25] provide the foundational account of how recommendation agents reshape consumer decision processes. This inferential opacity is the source of both the value and the threat that personalization carries. On one side, relevance reduces search costs and signals that the firm understands the consumer, mechanisms long argued to lift adoption and engagement (Aguirre et al. [11]; Chandra et al. [26]) and echoed in evidence that richer, more immersive digital presentation formats raise perceived usefulness and purchase intention (Yim et al. [27]). On the other, the covert character of behavioural inference can heighten perceived vulnerability, an affective state that depresses click-through and purchase responses even when relevance is objectively high (Aguirre et al. [11]; Bleier and Eisenbeiss [12]). The demonstration by Aguirre et al. [11] that covert data collection inverts the benefit of personalization established the empirical core of the personalization privacy paradox and, together with its extension in Aguirre et al. [13], remains the reference point against which subsequent boundary-condition research is evaluated. We retain a baseline expectation that, on average, personalization raises purchase intention, while treating that main effect as theoretically incomplete until its mediating and moderating structure is specified.
H1. AI-powered personalization is positively associated with consumers’ purchase intention.
2.2. Decomposing the Trust Mechanism: Cognitive and Affective Pathways
Initial trust theory holds that, when transacting with an unfamiliar online actor, consumers form trusting beliefs that allow them to overcome uncertainty and risk and to act by following advice, disclosing information, and purchasing (McKnight et al. [15] [16]; Gefen et al. [28]). Critically, these beliefs are not monolithic. Cognitive trust is a calculative judgement grounded in competence and reliability cues, whereas affective trust is an emotional bond grounded in perceived benevolence and goodwill (McKnight et al. [15] [16]; Mayer et al. [17]; Johnson and Grayson [18]). Although our respondents had recent personalized-purchase experience with a platform, initial trust theory remains the appropriate lens here because the object of trust is the AI recommender agent rather than the retail platform as a whole. Even experienced online shoppers encounter each algorithmic recommendation as an interaction with an opaque, non-human actor whose inference process is unobservable and whose “intentions” cannot be verified from prior familiarity; every new recommendation reactivates the same competence- and benevolence-based judgments that initial trust theory describes, because the machine unlike a human counterpart accrues limited relational history and is continually re-evaluated against fresh performance cues. In this sense the construct we model is best understood as the ongoing, repeatedly re-formed trust that consumers extend to AI-enabled recommendation systems, for which the cognitive and affective distinction of initial trust theory provides the operative structure; we therefore retain its two-component architecture while framing trust as recurrently negotiated across encounters rather than formed only at first contact. AI-powered personalization speaks to both, but through different channels. Accurate, well-targeted recommendations function as competence signals: they provide ongoing evidence that the system reliably models the consumer’s needs, supplying precisely the kind of performance information on which cognitive trust is calibrated (Dietvorst et al. [20]; Castelo et al. [21]), a mechanism Komiak and Benbasat [29] document directly for recommendation agents, whose personalization and familiarity raise both cognitive and emotional trust. Personalization also carries relational meaning being recognized and anticipated can be experienced as a form of attentiveness thereby feeding the affective component, though this channel is more fragile because the same recognition can be reinterpreted as intrusion (Bleier and Eisenbeiss [12]; Aguirre et al. [13]).
Because cognitive and affective trust are themselves established antecedents of behavioural intention in online exchange (McKnight et al. [15] [16]; Pavlou [30]; Bart et al. [31]), we model them as parallel mediators linking personalization to purchase intention. Distinguishing them is not a psychometric nicety: it is the analytical precondition for detecting the divergent boundary conditions developed below, which an undifferentiated trust construct would average away.
H2a. AI-powered personalization is positively associated with cognitive trust in the AI agent.
H2b. AI-powered personalization is positively associated with affective trust in the AI agent.
H3a. Cognitive trust mediates the relationship between AI-powered personalization and purchase intention.
H3b. Affective trust mediates the relationship between AI-powered personalization and purchase intention.
2.3. Perceived AI Capability as a Contingency on the Cognitive Pathway
If cognitive trust is calibrated on competence evidence, then the diagnostic value of any given personalization cue depends on the consumer’s prior belief about whether the system is capable of competent inference at all. The algorithm-appreciation literature shows that, absent disconfirming experience, individuals often prefer algorithmic to human judgement, but that this appreciation is conditional and collapses when competence is in doubt (Logg et al. [19]; Dietvorst et al. [20]). Conversely, the algorithm-aversion literature documents a steep penalty for perceived machine error and a reluctance to rely on systems judged incompetent (Dietvorst et al. [20]; Castelo et al. [21]), an aversion that Longoni et al. [22] find to be sharpest where consumers doubt that a machine can accommodate their individuality. Evidence from service settings converges on the same conclusion: acceptance of AI-enabled and robotic frontline agents is governed principally by beliefs about what the technology can competently do (Wirtz et al. [32]; Ostrom et al. [33]; Belanche et al. [34]), and comparable capability assessments drive AI adoption at the organizational level (Chatterjee et al. [35]), while Hermann [36] argues that the visible ethical conduct of firms deploying AI further shapes the credibility consumers assign to it. Perceived AI capability the consumer’s general belief in the system’s competence to understand and serve them should therefore act as an interpretive lens. When capability beliefs are high, personalization cues are read as confirmation of competence and translate efficiently into cognitive trust; when they are low, the same cues are discounted as noise or coincidence. The personalization-to-cognitive-trust relationship is thus expected to strengthen as perceived AI capability rises.
H4. Perceived AI capability moderates the relationship between AI-powered personalization and cognitive trust, such that the relationship is stronger when perceived AI capability is high.
2.4. Privacy Risk Perception as a Contingency on the Affective Pathway
The affective pathway is governed by a different boundary condition. The personalization–privacy paradox locates the threat of personalization not in competence but in vulnerability the felt exposure that arises when a firm appears to know more than the consumer disclosed (Aguirre et al. [11]; Bleier and Eisenbeiss [12]; Aguirre et al. [13]). Privacy risk perception, the consumer’s expectation that personal data may be misused, sits at the centre of the privacy calculus through which consumers weigh disclosure against benefit (Dinev and Hart [37]), and firm-level evidence confirms that mishandled data carries measurable performance costs (Martin et al. [38]); it should therefore condition the emotional, benevolence-based response to personalization rather than the competence-based one. In emerging-market settings where institutional data protection is weak, this perception is unlikely to be offset by structural assurance, intensifying its relevance (Sheth [23]; McKnight et al. [16]). When privacy risk perception is high, the relational warmth that personalization might otherwise generate is overwritten by suspicion of opportunism, weakening the personalization-to-affective-trust link; when it is low, personalization is freer to register as attentiveness.
H5. Privacy risk perception moderates the relationship between AI-powered personalization and affective trust, such that the relationship is weaker when privacy risk perception is high.
Taken together, H1 - H5 constitute a moderated-mediation system in which the indirect effect of personalization on purchase intention is conditional on perceived AI capability (through cognitive trust) and on privacy risk perception (through affective trust). Figure 1 presents the model.
Figure 1. Capability trust contingency model of AI personalization and purchase intention. Note: Solid paths denote hypothesized direct and mediating relationships; dotted paths denote moderation. APP = AI-Powered Personalization; PI = Purchase Intention.
3. Method
3.1. Sample and Procedure
Data were collected from active online consumers in Nigeria through a structured self-administered questionnaire distributed across major commercial centres and online shopping communities. Respondents were recruited through non-probability purposive and snowball sampling between March and May 2024. The questionnaire, administered online via Google Forms, was circulated through three channels: 1) e-commerce and online-shopping interest groups on WhatsApp, Facebook, and Telegram; 2) targeted social-media posts on Instagram and X (formerly Twitter) using shopping-related hashtags; and 3) intercept referrals in high-footfall commercial areas of Lagos, Abuja, and Port Harcourt. Participation was voluntary; on completion, respondents could enter a raffle for one of twenty ₦2000 mobile-airtime vouchers as a modest incentive. To ensure a common referent, each respondent was asked before answering the focal items to recall one specific, recent personalized online-shopping experience, prompted as follows: Please think of the most recent occasion, within the past three months, on which an online shopping app or website showed you personalized product recommendations chosen for you by its system. Keep that specific experience and that platform in mind as you answer the following questions. Respondents also named the platform they had in mind; the responses spanned Jumia, Konga, Amazon, AliExpress, Temu, and Instagram and Facebook Shops, with no single platform exceeding 40% of cases. Eligibility required at least one personalized online purchase in the preceding three months, ensuring that respondents could meaningfully evaluate AI-driven recommendation experiences. Of 561 questionnaires initiated, 512 were submitted 49 abandoned before submission. Screening then removed cases sequentially: 38 for failing the eligibility no personalized purchase in the past three months, 21 for incomplete responses on focal items, 25 for failing at least one of two embedded attention checks, and 16 for straight-line or near-zero-variance response patterns, leaving an analytic sample of 412 (73.4%) of submitted questionnaires. After screening for incomplete responses, attentiveness-check failures, and straight-line patterns, the analytic sample comprised 412 respondents. The sample skewed toward younger, mobile-first consumers consistent with the demographic profile of Nigerian e-commerce, with balanced gender representation and variation across income and education strata. Table 1 summarizes respondent characteristics.
Table 1. Sample characteristics (N = 412).
Characteristic |
Category |
n |
% |
Gender |
Female |
201 |
48.8 |
Male |
211 |
51.2 |
Age |
18 - 24 |
148 |
35.9 |
25 - 34 |
169 |
41.0 |
35 - 44 |
71 |
17.2 |
45+ |
24 |
5.8 |
Monthly online spend |
Low |
139 |
33.7 |
Medium |
186 |
45.1 |
High |
87 |
21.1 |
Primary device |
Smartphone |
357 |
86.7 |
Desktop/Other |
55 |
13.3 |
3.2. Measures
Respondents evaluated an AI-powered recommender system the personalized product-recommendation engine of a consumer e-commerce platform or online marketplace (e.g., recommendation feeds, recommended for you and because you viewed modules on shopping apps and websites) and not a customer-service chatbot, a social-media feed algorithm, or a retailer’s human-curated promotions. All focal constructs were measured with multi-item, seven-point Likert scales adapted from validated instruments and contextualized to AI-powered shopping in Nigeria. AI-powered personalization was operationalized through perceived recommendation relevance and adaptivity, building on the personalization-effectiveness scales of Aguirre et al. [11] and Chandra et al. [26] and on the recommendation-agent measures of Komiak and Benbasat [29]. Cognitive and affective trust were measured using items reflecting the competence/reliability and benevolence and goodwill distinction central to initial trust theory (McKnight et al. [15] [16]), with wording informed by the credibility-based trust scales of Lou and Yuan [39]. Perceived AI capability captured beliefs about the system’s competence to understand consumer needs (Logg et al. [19]; Dietvorst et al. [20]). Privacy risk perception was assessed with items reflecting expected likelihood and severity of data misuse (Bleier and Eisenbeiss [12]; Aguirre et al. [13]; Dinev and Hart [37]). Purchase intention was measured with established behavioural-intention items (McKnight et al. [15]; Venkatesh et al. [40]). Age, gender, income, prior online-purchase frequency, and product category were retained as covariates. Table 2 reports the measurement model.
Table 2. Measurement model: reliability and convergent validity.
Construct |
Items |
α |
CR |
AVE |
AI-powered personalization (APP) |
4 |
0.89 |
0.91 |
0.72 |
Cognitive trust |
4 |
0.91 |
0.93 |
0.77 |
Affective trust |
4 |
0.88 |
0.91 |
0.71 |
Perceived AI capability |
3 |
0.86 |
0.91 |
0.78 |
Privacy risk perception |
4 |
0.90 |
0.92 |
0.74 |
Purchase intention (PI) |
3 |
0.92 |
0.95 |
0.86 |
All standardized item loadings were significant (p < 0.001) and ranged from 0.71 to 0.92, exceeding the 0.70 threshold. The measurement model demonstrated good fit: χ2(174) = 312.6, χ2/df = 1.80, CFI = 0.968, TLI = 0.961, RMSEA = 0.044 (90% CI [0.037, 0.051]), and SRMR = 0.038. Internal consistency and composite reliability exceeded conventional thresholds, and average variance extracted (AVE) exceeded 0.50 for every construct, supporting convergent validity. Discriminant validity was satisfied under the Fornell-Larcker criterion (Fornell and Larcker [41]), with the square root of each construct’s AVE exceeding its inter-construct correlations, and corroborated by heterotrait-monotrait ratios below the 0.85 benchmark recommended by Henseler et al. [42]. Because all data were self-reported, common-method bias was assessed following the guidance of Podsakoff et al. [43]: Harman’s single-factor test returned a first factor accounting for less than half of total variance, and a marker-variable adjustment left the structural estimates substantively unchanged.
3.3. Analytic Strategy
Hypotheses were tested using covariance-based structural equation modelling for the measurement and direct-path estimates, complemented by bias-corrected bootstrap moderated-mediation analysis (5000 resamples) to estimate conditional indirect effects and indices of moderated mediation following Hayes [44]. Predictors entering interaction terms were mean-centred. This combined strategy directly tests whether the indirect effects of personalization on purchase intention vary across levels of perceived AI capability and privacy risk perception, which is the empirical crux of the contingency argument.
4. Results
4.1. Descriptive Statistics and Correlations
Table 3 presents means, standard deviations, and bivariate correlations. Personalization correlated positively with cognitive trust, affective trust, and purchase intention; privacy risk perception correlated negatively with affective trust and purchase intention. The pattern is consistent with the proposed dual-pathway structure and provides preliminary support for the moderating roles developed in the model.
Table 3. Means, standard deviations, and correlations.
Variable |
M |
SD |
1 |
2 |
3 |
4 |
5 |
6 |
1) APP |
4.92 |
1.18 |
- |
|
|
|
|
|
2) Cognitive trust |
4.71 |
1.24 |
0.58 |
- |
|
|
|
|
3) Affective trust |
4.38 |
1.31 |
0.46 |
0.61 |
- |
|
|
|
4) Perceived AI capability |
4.85 |
1.22 |
0.49 |
0.55 |
0.41 |
- |
|
|
5) Privacy risk perception |
4.63 |
1.40 |
0.12 |
−0.18 |
−0.34 |
−0.09 |
- |
|
6) Purchase intention |
4.80 |
1.29 |
0.55 |
0.64 |
0.52 |
0.48 |
−0.27 |
- |
Correlations |r| ≥ 0.12 are significant at p < 0.05; |r| ≥ 0.16 at p < 0.01. APP = AI-powered personalization.
4.2. Direct Effects and Mediation
The structural model fit the data well by conventional indices (χ2(179) = 331.4, χ2/df = 1.85, CFI = 0.965, TLI = 0.959, RMSEA = 0.045 [90% CI 0.038 - 0.052], SRMR = 0.041). The model explained substantial variance in the endogenous constructs: R2 = 0.34 for cognitive trust, R2 = 0.19 for affective trust, and R2 = 0.52 for purchase intention. Supporting H1, personalization exhibited a positive total effect (β = 0.55) on purchase intention, of which the direct effect (β = 0.21) net of the cognitive- and affective-trust mediators was smaller; H1 is thus supported as a total-effect hypothesis, and the direct path in Table 4 is labelled accordingly so that its interpretation matches the partially mediated model. Supporting H2a and H2b, personalization significantly predicted both cognitive trust and affective trust, with the standardized coefficient larger on the cognitive path. Both trust dimensions in turn predicted purchase intention, with cognitive trust the stronger proximal driver. Bootstrap mediation analysis confirmed significant indirect effects through both pathways, supporting H3a and H3b; the cognitive-trust pathway carried the larger share of the total indirect effect, indicating that, in this market, personalization persuades primarily by demonstrating competence rather than by cultivating warmth. Table 4 reports the structural estimates.
Table 4. Structural path estimates.
Path |
β |
SE |
t |
Result |
APP → Purchase intention (H1, direct effect; total effect β = 0.55) |
0.21 |
0.05 |
4.20 |
Supported |
APP → Cognitive trust (H2a) |
0.54 |
0.04 |
13.50 |
Supported |
APP → Affective trust (H2b) |
0.39 |
0.05 |
7.80 |
Supported |
Cognitive trust → Purchase intention |
0.42 |
0.05 |
8.40 |
Supported |
Affective trust → Purchase intention |
0.24 |
0.05 |
4.80 |
Supported |
APP → Cognitive trust → PI (H3a) |
0.23 |
0.03 |
- |
Supported |
APP → Affective trust → PI (H3b) |
0.09 |
0.02 |
- |
Supported |
All coefficients significant at p < 0.01. Indirect effects estimated via 5000 bias-corrected bootstrap resamples; 95% CIs excluded zero.
4.3. Moderation and Conditional Indirect Effects
Supporting H4, the interaction of personalization and perceived AI capability on cognitive trust was positive and significant (β = 0.19, SE = 0.05, t = 3.80, p < 0.001, 95% CI [0.09, 0.29]): the slope of personalization on cognitive trust was substantially steeper for consumers high in capability beliefs than for those low in them, as depicted in the simple-slopes plot in Figure 2. The diagnostic value of a personalization cue, in other words, depends on whether the consumer already credits the system with competence.
Figure 2. Interaction of personalization and perceived AI capability on cognitive trust. Note: Simple slopes of AI-powered personalization on cognitive trust at high (+1 SD) and low (−1 SD) levels of perceived AI capability.
Supporting H5, the interaction of personalization and privacy risk perception on affective trust was negative and significant (β = −0.16, SE = 0.05, t = −3.20, p = 0.001, 95% CI [−0.26, −0.06]): under high privacy risk perception, the affective benefit of personalization was attenuated, consistent with the vulnerability logic of the personalization-privacy paradox. The index of moderated mediation was significant for the cognitive pathway moderated by perceived AI capability, confirming that the indirect effect of personalization on purchase intention through cognitive trust is conditional on capability beliefs. Figure 3 plots the conditional indirect effects, which rise monotonically from a non-significant effect at low capability to a strong, significant effect at high capability.
![]()
Figure 3. Conditional indirect effects at levels of perceived AI capability 95% (Bias corrected bootstrap CIs). Note: Conditional indirect effects of AI-powered personalization on purchase intention via cognitive trust, across levels of perceived AI capability, with 95% bias-corrected bootstrap confidence intervals.
Table 5 summarizes the conditional indirect effects and the indices of moderated mediation, providing the formal test of the contingency structure.
Table 5. Conditional indirect effects and indices of moderated mediation.
Pathway/moderator level |
Indirect effect |
95% CI |
Significant |
Cognitive trust @ low capability (−1 SD) |
0.06 |
[−0.02, 0.15] |
No |
Cognitive trust @ mean capability |
0.21 |
[0.12, 0.31] |
Yes |
Cognitive trust @ high capability (+1 SD) |
0.37 |
[0.24, 0.51] |
Yes |
Index of moderated mediation (capability) |
0.15 |
[0.07, 0.24] |
Yes |
Affective trust @ low privacy risk (−1 SD) |
0.14 |
[0.07, 0.22] |
Yes |
Affective trust @ high privacy risk (+1 SD) |
0.04 |
[−0.01, 0.10] |
No |
Index of moderated mediation (privacy risk) |
−0.05 |
[−0.10, −0.01] |
Yes |
5. Discussion
5.1. Theoretical Contributions
This study reframes AI-powered personalization from a uniform persuasion lever into a contingent, mechanism-specific process, and in doing so makes three contributions. First, by decomposing the trust mediator into cognitive and affective components, it resolves an aggregation problem that has masked how personalization actually operates. The finding that the cognitive pathway dominates the affective pathway in this market indicates that, where institutional assurance is scarce, consumers lean on demonstrated competence rather than inferred goodwill when deciding to act on algorithmic recommendations an insight unavailable to models that treat trust as a single construct (McKnight et al. [15] [16]; Johnson and Grayson [18]). Second, by specifying perceived AI capability and privacy risk perception as moderators of distinct paths, the study converts the personalization–privacy paradox from a static tension into a structured contingency: capability beliefs govern when competence signals translate into trust, while privacy risk governs when relational warmth survives (Aguirre et al. [11]; Aguirre et al. [13]; Logg et al. [19]). The significant index of moderated mediation for the capability pathway is, to our knowledge, a novel demonstration that the persuasive power of personalization is conditional on the consumer’s prior theory of the machine’s competence. Third, by situating the model in Nigeria, the study extends consumer AI theory beyond the high-institutional-trust settings that dominate the literature and shows that the scope conditions of trust formation differ where structural assurance is weak (Sheth [23]; McKnight et al. [16]; Hoffman and Novak [24]).
5.2. What Is New
Prior work has established that personalization can both attract and repel, that trust mediates digital-commerce outcomes, and that algorithm appreciation is conditional. What has been missing is an integrated account specifying which trust mechanism each force operates through and which consumer belief switches it on or off. The capability-trust contingency model supplies that account and tests it as a moderated-mediation system rather than as a set of disconnected main effects. The relocation to an emerging African market is not incidental: it is the condition under which the dominance of the cognitive pathway and the salience of privacy risk become observable, because the institutional buffers that would otherwise obscure them are absent.
5.3. Managerial Implications
For firms personalizing at scale in emerging markets, the results counsel against treating personalization intensity as a monotonic good. Because the cognitive pathway dominates and is gated by capability beliefs, returns to personalization are largest among consumers who already credit the system with competence; for sceptical segments, investments in visible competence signalling explainable recommendations, accuracy cues, and transparent rationale are prerequisites rather than embellishments (Castelo et al. [21]). Because the affective pathway is suppressed by privacy risk, firms operating where data-protection enforcement is weak cannot rely on relational warmth to carry conversion and should instead reduce perceived vulnerability through overt, consent-based data practices that substitute for absent structural assurance (Aguirre et al. [11]; Aguirre et al. [13]; Martin et al. [38]). Strategically, this implies adaptive personalization: calibrating intensity and transparency to inferred capability and privacy-risk profiles rather than applying a uniform policy.
5.4. Limitations and Future Research
The cross-sectional design limits causal inference; longitudinal and experimental replications would strengthen the temporal ordering of personalization, trust, and intention. The focus on a single national market, chosen for its theoretical leverage, invites comparative replication across African and other emerging economies to test the generalizability of the cognitive-pathway dominance. Future work might also incorporate behavioural outcomes beyond stated intention, examine the dynamics of trust repair following algorithmic error, and test whether explainable-AI interventions can shift capability beliefs and thereby unlock the conditional indirect effect among currently sceptical consumers.
6. Conclusion
AI-powered personalization does not persuade uniformly; it persuades conditionally, through separable trust mechanisms whose activation depends on what consumers believe about machine competence and data exposure. By decomposing the trust mediator, specifying targeted moderators, and testing the resulting moderated-mediation system in an under-studied emerging market, this study replaces the prevailing average-effect view with a contingency account of when algorithmic personalization converts attention into commitment. For theory, it clarifies the architecture of trust in consumer-AI interaction; for practice, it reframes personalization strategy as an exercise in trust calibration rather than relevance maximization.
Author Contributions
Conceptualization, N.U.G.O.; methodology, N.U.G.O.; software, N.U.G.O.; validation, N.U.G.O. and Y.Y.; formal analysis, N.U.G.O.; investigation, N.U.G.O.; resources, N.U.G.O. and Y.Y.; data curation, N.U.G.O.; writing original draft preparation, N.U.G.O.; writing review and editing, N.U.G.O. and Y.Y.; visualization, N.U.G.O.; supervision, Y.Y.; project administration, N.U.G.O.; funding acquisition, N.U.G.O. All authors have read and agreed to the published version of the manuscript.