Artificial intelligence (AI) is changing last-mile logistics through tools such as autonomous routing, predictive delivery management and unmanned delivery platforms (Dwivedi et al. 2021; Nwachukwu 2023). In sub-Saharan Africa, these technologies may help address persistent service-access gaps, although their value depends on the reliability and affordability of the surrounding digital and physical infrastructure (Aker & Mbiti 2010; World Bank 2024). Operational evidence from Zipline indicates that its network serves health facilities across several African countries and reaches more than 49 million people; these figures demonstrate the potential scale of drone-enabled logistics, but they do not by themselves establish consumer-acceptance (Seelos et al. 2024). In South Africa, deep spatial and socioeconomic inequalities make the possible benefits substantial but unevenly distributed (Statistics South Africa 2023).

The success of AI-driven delivery systems therefore depends not only on technical performance, but also on whether consumers regard them as useful, affordable, trustworthy and accessible. Infrastructure interruptions can disrupt mobile connectivity, cloud-based routing and logistics operations, while differences in income, digital literacy and location affect the ability to use technology-dependent services (Eskom 2023; Gillwald, Mothobi & Rademan 2018). These conditions make acceptance a contextual and information-management problem rather than a purely individual attitude.

Technology-acceptance research in emerging markets frequently uses the Technology Acceptance Model (TAM) and related individual-level constructs. Technology Acceptance Model remains useful for perceived usefulness and perceived ease of use, but it does not fully represent the value trade-offs, community diffusion processes, infrastructural constraints and regulatory conditions that accompany AI-enabled consumer services (Kim, Chan & Gupta 2007; Venkatesh, Thong & Xu 2012). The Value-based Adoption Model (VAM), Unified Theory of Acceptance and Use of Technology 2 (UTAUT2), Diffusion of Innovations (DOI) and Technology-Organisation-Environment (TOE) framework provide complementary constructs for examining those dimensions.

In South Africa, the Protection of Personal Information Act (POPIA), the regulatory framework for unmanned aircraft, infrastructure instability and persistent digital inequality create institutional and environmental conditions that are not adequately captured by TAM alone (Information Regulator South Africa 2021; South African Civil Aviation Authority n.d.). Evidence from Cape Town’s informal economy also shows that financial constraints, limited digital literacy and weak infrastructure interact in ways that may inhibit technology use (Ebrahim & Van den Berg 2024). A further limitation is the tendency to discuss South African consumers as a homogeneous adoption group despite substantial differences in income, skills and geographic location (Gillwald et al. 2018; Van Dijk 2020). The resulting problem is that decision-makers lack a consolidated, framework-explicit and critically appraised evidence base for designing AI-enabled delivery services in a structurally unequal market. This review addresses that problem by integrating consumer-level and structural perspectives while distinguishing evidence reported in the corpus from interpretations proposed for South Africa.

The aim of this review is to synthesise evidence on consumer-acceptance of AI-driven delivery systems in sub-Saharan Africa and to derive carefully qualified implications for information management in South Africa. The review is guided by the following central research question: Which contextual factors shape consumer-acceptance of AI-driven delivery systems in sub-Saharan Africa, and what do those factors imply for information management practice in South Africa?

Which contextual factors shape consumer-acceptance of AI-driven delivery systems in sub-Saharan Africa, and what are the implications of these factors for information management practice in South Africa?

The following secondary research questions support the central research question:

To synthesise and critically analyse the contextual factors shaping consumer-acceptance of AI-driven delivery systems in sub-Saharan Africa and determine their implications for information management practice in South Africa.

The following secondary objectives guide the review:

Five complementary theoretical lenses structure the review. Unified Theory of Acceptance and Use of Technology 2 (Venkatesh et al. 2012) and VAM (Kim et al. 2007) address consumer perceptions, facilitating conditions, price value and benefit-sacrifice trade-offs. Diffusion of Innovation (Rogers 2003) captures relative advantage and community-level diffusion, while TOE (Tornatzky & Fleischer 1990) accommodates technological, organisational and environmental conditions such as infrastructure, POPIA and aviation regulation. Technology Acceptance Model (Davis 1989) is retained as a historical comparator for perceived usefulness and perceived ease of use. The nine thematic categories were specified a priori from this combined architecture. Unified Theory of Acceptance and Use of Technology 2 and VAM primarily support RQ1 and RQ2; DOI and TOE extend the analysis to the community and institutional conditions needed for RQ3 and RQ4; and TAM provides the baseline against which broader model configurations are described. Table 4 presents the cross-framework mapping.

This study employed a systematic literature review guided by Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 (Page et al. 2021). The review combined thematic synthesis with quantitative content analysis organised around UTAUT2, VAM, DOI, TOE and TAM. It did not conduct a formal meta-analysis because the included studies reported heterogeneous designs, constructs and statistical measures.

The review covered consumer and user acceptance of AI-enabled, autonomous or technology-driven delivery systems in sub-Saharan Africa, with a deliberate focus on South Africa. Evidence from Nigeria, Kenya, Ghana, Rwanda, Tanzania, Zimbabwe and wider regional sources was used where South African-specific studies were limited. Transfer to South Africa was treated as an interpretive step and not assumed automatically.

A systematic search was conducted across six academic databases: Scopus (n = 98), Web of Science (n = 74), Google Scholar (n = 89), IEEE Xplore (n = 31), EBSCOhost (n = 52) and African Journals Online (AJOL) (n = 34), yielding 378 records before deduplication. African Journals Online was included to improve coverage of African scholarship. The core search combined three conceptual clusters and was adapted to the syntax of each database:

The eligibility window was 2013–2025, which is consistent with the final corpus and the inclusion of Badu-Baiden et al. (2025). Supplementary terms included ‘performance expectancy’, ‘price value’, ‘relative advantage’, ‘TOE framework’ and ‘facilitating conditions’. The exact final search date and the database-specific search logs were not retained in the source manuscript; this reduces reproducibility and is acknowledged in the limitations.

Table 1 presents the pre-defined inclusion and exclusion criteria applied during screening. The inclusion criteria required engagement with at least one construct from UTAUT2, VAM, DOI, TOE or TAM, broadening prior TAM-only requirements to capture the full breadth of the technology acceptance literature.

Screening was conducted in three PRISMA-consistent stages. From 378 records, 64 duplicates were removed, leaving 314 titles and abstracts for screening. A further 226 records were excluded for failing to meet the geographic, thematic or quality criteria. Eighty-eight full texts were assessed, 46 were excluded, and 42 studies were retained. Figure 1 presents the screening flow. The reported counts reflect the primary reason for exclusion where more than one reason applied.

A structured extraction template was applied to the 42 included studies and captured author and year, country or region, design, sample characteristics, framework, constructs, reported statistical evidence, key findings and contextual factors. Table 2 provides the summary extract available in the manuscript. The complete 42-study extraction matrix was not included in the submitted file; consequently, study-level calculations that cannot be checked from the available record are interpreted cautiously. The complete matrix should accompany the article as supplementary material to permit independent verification.

Thematic synthesis followed Thomas and Harden (2008) in three phases: line-by-line coding against relevant framework constructs, development of descriptive themes across nine a priori categories and development of analytical themes for the South African context. Quantitative content analysis was conducted in parallel following Krippendorff (2018) and Neuendorf (2017). The quantitative component was descriptive and did not pool unlike statistical measures.

Four analytical layers were retained. Firstly, construct frequency and a review-specific Valence Consistency Index (VCI) were calculated as VCI = |n+ - n-|/(n+ + n-), where n+ and n- are the numbers of directionally positive and negative findings. Valence Consistency Index = 1.00 denotes complete directional agreement only when at least one positive or negative finding is available; it is reported as not applicable for constructs coded only as moderate, indirect or non-directional. The term VCI is used in this review for this simple descriptive calculation and is not presented as an established index from the content-analysis literature. Secondly, a construct co-occurrence matrix recorded how often pairs of constructs were investigated in the same study. Thirdly, the Herfindahl–Hirschman Index (HHI = sum of squared country shares) was adapted from its standard concentration formulation to describe geographic concentration in the review corpus (Hirschman 1964; Rhoades 1993). Fourthly, inter-rater reliability was assessed using Cohen’s kappa (Cohen 1960). The originally reported compound annual growth rate (CAGR) analysis was removed because the source manuscript did not provide annual base counts and used unequal publication intervals, preventing independent verification.

Two independent coders assessed a random sample of 10 studies (24% of the corpus). Observed agreement was 93%, yielding Cohen’s kappa = 0.89, conventionally interpreted as almost perfect agreement (Landis & Koch 1977). Discrepancies were resolved through discussion before full-corpus coding. NVivo 14 was used for data management.

The source manuscript did not contain a completed study-level risk-of-bias assessment or the full extraction information needed to reconstruct one reliably. No Joanna Briggs Institute (JBI) ratings or sensitivity analysis are therefore claimed in this version. The absence of a formal design-appropriate appraisal limits the certainty that can be attached to the descriptive summaries, especially where findings come from heterogeneous surveys, qualitative studies and operational reports. A future update should apply appropriate critical-appraisal tools at study level and report the decisions in full.

This study is a systematic literature review involving the secondary analysis of publicly available, previously published literature. It does not involve collecting primary data, recruiting human participants or using animal subjects. This article does not contain any studies involving human participants performed by any of the authors. Ethics approval was not required for this study. All sources consulted were cited in accordance with Harvard referencing conventions. No personal or identifiable information was collected or processed; accordingly, no data confidentiality measures beyond standard academic citation practice were required.

Following PRISMA screening, 42 studies were retained. The reported distribution comprised Nigeria (n = 12; 29%), Kenya (n = 9; 21%), South Africa (n = 8; 19%), Ghana (n = 7; 17%), Rwanda (n = 3; 7%), Tanzania (n = 2; 5%) and one study from an additional sub-Saharan African setting (n = 1; 2%). This accounts for six named countries plus one additional setting, rather than the nine countries stated in the earlier version. The resulting HHI is 0.200, indicating moderate concentration. Study designs were predominantly quantitative (n = 28; 67%), followed by qualitative (n = 8; 19%) and mixed-method or conceptual studies (n = 6; 14%). For quantitative studies, the reported sample-size range was n = 103 to n = 1231, with a median of n = 401; qualitative participant counts were not combined with this range because they are not directly comparable. Publications spanned 2013–2025, and 31 studies (74%) were published from 2019 onward. Framework frequencies were overlapping because one study could contribute to more than one category: TAM constructs appeared in 34 studies (81%), UTAUT or UTAUT2 in 22 (52%), DOI-aligned constructs in 18 (43%), VAM-aligned constructs in 12 (29%) and TOE-aligned constructs in 11 (26%). Fourteen studies (33%) used two or more frameworks.

Table 4 maps the nine thematic categories across the five-framework architecture and identifies the statistical type used in the source studies. In the subset reporting model-level R-squared values, TAM-only applications had a median R-squared of 0.49, while studies using three or more frameworks reported medians of 0.73–0.78. The absolute difference is 0.24–0.29, equivalent to 24–29 percentage points of explained variance; relative to 0.49, the upper comparison is approximately 59% higher. This comparison is descriptive. The subgroups differ in country, sample, measurement model and number of predictors, and R-squared can increase mechanically as predictors are added. The results therefore do not establish that theoretical breadth caused the difference.

Table 5 summarises the thematic evidence relevant to RQ4. For each theme, it identifies the principal framework, the settings represented in the available evidence, the statistical form reported and the carefully qualified relevance to South Africa.

The review answers RQ1 and RQ2 by showing that consumer-level perceptions and structural conditions recur together across the corpus. Usefulness and relative advantage were consistently positive, while cost, inadequate infrastructure and absent trust were recurring inhibitors. Unified Theory of Acceptance and Use of Technology 2 and VAM help explain the consumer-level benefit, effort and sacrifice calculations; DOI and TOE extend the analysis to community diffusion, infrastructure and regulation. The evidence supports using these frameworks in combination as an interpretive architecture, but it does not demonstrate that every multi-framework model is superior or that one framework configuration will transfer unchanged across countries.

The reported model-level comparison is descriptive: a median R-squared of 0.49 for TAM-only applications and 0.78 for five-framework applications represents a 29-percentage-point difference. The five-framework median is approximately 59% higher when 0.49 is the denominator, while the TAM-only median is approximately 37% lower when 0.78 is the denominator. Neither expression establishes causation. The HHI of 0.200 indicates moderate geographic concentration and cautions against treating the regional evidence as evenly distributed.

South Africa’s load-shedding experience illustrates how a technologically advanced market can still face recurring operational instability. Artificial intelligence-enabled delivery depends on electricity, connectivity, cloud services, routing data and charging infrastructure; disruption at any of these points may weaken reliability and trust (Eskom 2023; Gillwald et al. 2018). I use ‘infrastructure paradox’ to describe this tension between capability and continuity. The term is an interpretation generated by the synthesis. It should be tested through South African consumer and operator data rather than treated as a fact proven by the VCI or frequency counts.

Protection of Personal Information Act and South Africa’s unmanned-aircraft regulatory framework can operate as both safeguards and constraints. Clear privacy protections may reduce perceived risk, while complex or restrictive operational requirements may slow the scaling of drone services into underserved areas (Information Regulator South Africa 2021; Republic of South Africa 2013; South African Civil Aviation Authority n.d.). I interpret this as regulatory duality: regulation may build legitimacy while also increasing implementation burden. The mixed VCI for institutional and policy support is consistent with that interpretation, but it does not quantify the effect of POPIA or aviation rules on South African consumers directly.

South African townships present practical challenges, including informal addressing, dense settlement patterns, security concerns and unequal access to conventional delivery services. Diffusion of Innovation suggests that relative advantage may be especially salient where the alternative is limited service, while VAM highlights whether that benefit outweighs cost, privacy and reliability concerns (Kim et al. 2007; Rogers 2003). Operators may therefore consider framing AI-enabled delivery around access and service reliability rather than convenience alone. This remains an evidence-informed proposition that requires local validation.

Unified Theory of Acceptance and Use of Technology 2’s facilitating-conditions construct and digital-divide scholarship indicate that South African consumers should not be treated as a homogeneous adoption group (Gillwald et al. 2018; GSMA 2024; Van Dijk 2020). Differences in connectivity, skills, language, income and location are likely to shape onboarding and continued use. The low co-occurrence of institutional support and digital literacy in the corpus suggests that regulatory and individual-level research have often proceeded separately. South African primary research should examine how those levels interact rather than presuming a fixed segmentation model.

The review’s principal contribution is not the claim that one acceptance framework should replace another. It is the argument that AI-driven delivery in sub-Saharan Africa must be analysed across linked levels: consumer perceptions, value sacrifices, community diffusion, infrastructure and institutions. For South Africa, this shifts information management from a narrow question of whether users intend to adopt towards a broader question of whether services remain accessible, trustworthy and operable under unequal conditions. The infrastructure paradox and regulatory duality are proposed as organising interpretations for that broader agenda. They are supported by the synthesis but remain open to empirical challenge.

Strengths of the review include the use of five complementary frameworks, PRISMA-guided screening, inclusion of African-focused sources, explicit construct mapping and high inter-rater agreement for the sampled coding (kappa = 0.89). Important limitations remain. The protocol was not prospectively registered; the exact final search date and database-specific search logs were not retained; the complete 42-study extraction matrix was not included in the supplied manuscript; and no formal study-level risk-of-bias assessment was available. The evidence is geographically concentrated, dominated by quantitative surveys and based on heterogeneous instruments and statistical measures. The R-squared subgroup comparison is descriptive and may partly reflect model complexity. Operational reports and policy documents provide contextual insight but should not be treated as equivalent to peer-reviewed consumer-acceptance studies. These limitations restrict causal inference and transferability to South Africa.

Table 6 presents evidence-informed implications by stakeholder group. The actions are framed as options to be evaluated and tested in context rather than as outcomes guaranteed by the review evidence.