The role of artificial intelligence in food systems transformation: a review of opportunities, challenges, and future directions

1 Introduction

Global food systems face increasing pressure from climate change, population growth, resource scarcity, geopolitical instability, and supply-chain inefficiencies. These interconnected challenges threaten food security, environmental sustainability, and socio-economic resilience, particularly in low- and middle-income countries where infrastructural and institutional constraints can limit adaptive capacity (Béné et al., 2019; FAO, 2020; High Level Panel of Experts on Food Security and Nutrition, 2020; Ndhlovu, 2025; von Braun et al., 2006). Although agricultural productivity has improved over recent decades, these gains have not consistently translated into equitable food access, resilient supply chains, or sustainable resource use (Braimoh, 2020; Clark et al., 2020). Food-system transformation therefore requires more than production improvements; it requires integrated approaches that address efficiency, resilience, governance, and sustainability.

Artificial intelligence (AI) is increasingly recognized as a potential enabler of this transformation. Technologies such as machine learning, robotics, computer vision, and the Internet of Things (IoT) can enhance productivity, strengthen food safety, and improve supply-chain efficiency (Ayoub, 2026; Galali et al., 2026; Ikram et al., 2024). However, technological capability alone does not guarantee sustainable transformation. Weak institutions, inadequate digital infrastructure, governance constraints, and affordability barriers can limit AI adoption and effectiveness (İçöz, 2026; Munialo, 2025; Munyanyi and Pooe, 2024). AI’s contribution therefore depends on the conditions under which technologies are adopted, implemented, governed, and scaled.

Despite the rapid expansion of AI research, existing reviews largely synthesize applications within specific domains such as precision agriculture, food processing, and supply-chain management (Ayoub, 2026; Bayisa and Zewudu, 2026; Galali et al., 2026; Reitano et al., 2025). These reviews provide valuable insights into technological applications and their potential benefits but offer limited explanation of how AI contributes to transformation across the wider food system. In particular, less attention has been given to how technological capabilities interact with organizational readiness, innovation diffusion, governance, sustainability, and resilience to shape implementation and system-level outcomes. The principal gap is therefore not the absence of research on AI applications, but the lack of an integrated theoretical explanation of how and under what conditions AI capabilities translate into sustained food-system transformation.

Accordingly, this review addresses the following research question: How does AI enable food-system transformation, and what socio-technical, governance, and sustainability mechanisms shape its successful adoption and long-term impact? To address this question, the review synthesizes evidence on AI applications across the food system, critically examines opportunities and implementation challenges, and analyzes how socio-technical readiness, innovation diffusion, governance, sustainability, and resilience shape AI-enabled transformation.

Unlike previous reviews that primarily synthesize AI applications within individual food-system domains, this review integrates Socio-Technical Systems Theory (Trist and Bamforth, 1951; Trist, 1981) and Diffusion of Innovations Theory (Rogers, 2003) to develop the Integrated Dynamic Socio-Technical AI Food Systems Transformation Framework. The framework conceptualizes transformation as a dynamic process in which AI capabilities interact with socio-technical readiness, innovation diffusion, governance and ethics, and food-system outcomes. Its contribution lies in connecting technological capability with the conditions that shape implementation, adoption, scaling, and system-level transformation, providing a theoretical basis for future research, policy development, and responsible AI implementation.

The remainder of the paper is organised as follows: Section 2 reviews the literature and presents the theoretical framework underpinning the study. Section 3 describes the methodology. Section 4 presents and discusses the findings across five interrelated themes, followed by a cross-theme synthesis. Finally, Section 5 concludes the paper by summarising the key findings, theoretical and practical contributions, limitations and directions for future research.

2 Literature review

2.1 Overview of food systems

Food systems are increasingly understood as complex socio-technical systems rather than linear supply chains. They connect production, processing, storage, transportation, distribution, consumption, and waste management, while interacting with institutions, markets, people, and environmental systems that shape food security, nutrition, and sustainability (FAO, 2013; Béné et al., 2019; High Level Panel of Experts on Food Security and Nutrition, 2020; von Braun et al., 2006). Yet much of the literature still examines individual stages separately, offering useful operational insights but limited understanding of how interactions across the wider system shape outcomes (Clark et al., 2020; Stefanovic et al., 2020).

These interconnected components face different but related pressures. Production depends on land, soil, water, climate, management practices, and technology, while processing, storage, and transportation influence food quality, value addition, and losses (Braimoh, 2020; Mugabi, 2023; Ayoub, 2026; Galali et al., 2026). Distribution and markets shape food accessibility and affordability, while policies, consumer behavior, and market conditions influence consumption (FAO, 2020c; Hidalgo et al., 2021). Waste management affects resource efficiency and environmental sustainability (European Commission, 2016; Rohini et al., 2020). Weakness in one component can therefore affect outcomes elsewhere in the system.

Climate change further exposes these interdependencies, although vulnerability and adaptive capacity vary across contexts. Rain-fed agriculture in low-income settings is particularly vulnerable to climate shocks, while irrigation and precision technologies can strengthen adaptive capacity where the necessary resources and infrastructure are available (Ehui and Klytchnikova, 2020; Thomas et al., 2022). These differences suggest that food-system resilience depends not only on exposure to environmental stress but also on the capacity of institutions, infrastructure, and actors to respond.

Economic, infrastructural, and socio-political pressures can intensify these vulnerabilities. Inadequate storage, fragmented supply chains, and limited access to finance contribute to post-harvest losses and market instability (Béné, 2020; Ndhlovu and Mhlanga, 2024). Conflict and political instability can further disrupt production and distribution, thereby deepening food insecurity (Eltahir and Siddig, 2024). The COVID-19 pandemic similarly demonstrated how labor shortages, market closures, and logistical disruptions can spread across interconnected food-system activities (Béné, 2020; Nchanjia and Lutomia, 2021a,b). Environmental degradation and rising food loss and waste further pressure resource use, nutrition, and sustainability (Braimoh, 2020; Mugabi, 2023; European Commission, 2016; Rohini et al., 2020).

Together, these challenges show that food-system vulnerabilities rarely occur in isolation. Environmental shocks interact with infrastructure, markets, institutions, governance, and social conditions, meaning that interventions focused on a single component may have limited system-wide effects. This interdependence strengthens the need for integrated approaches that improve resilience, sustainability, and food security across the wider food system.

2.2 Artificial intelligence and food systems

Artificial intelligence (AI) is increasingly used across agricultural production, food processing, supply chains, and food security. Technologies such as machine learning, deep learning, computer vision, robotics, predictive analytics, and generative AI support crop monitoring, disease detection, harvesting, quality control, process optimization, and supply-chain decision-making (Agrawal et al., 2025; Dutta et al., 2026; Javaid et al., 2023; Reitano et al., 2025; Zhang et al., 2025). Dutta et al. (2026) also points to a shift from stand-alone applications toward integrated agricultural systems that combine sensing, data analysis, automated control, and decision support.

Recent horticultural research illustrates this shift. In capsicum production, researchers have applied computer vision to fruit detection, growth-stage and maturity assessment, fruit counting, and robotic harvesting (Paul et al., 2024, 2026; Paul and Machavaram, 2024). Comparisons of specialized YOLO models with zero-shot approaches also show efforts to develop more flexible object-detection methods (Paul and Machavaram, 2025a). Similar advances across other crops combine deep learning with fruit recognition, 3D localization, ripeness assessment, and robotic harvesting under complex conditions (Huang et al., 2025; Song et al., 2025). Harvesting robots increasingly integrate machine vision, navigation, manipulators, and end-effectors (Hua et al., 2025). Together, these developments show a shift from isolated prediction and detection toward AI systems that integrate perception, decision-making, and physical automation.

This shift extends beyond primary production. In food processing and manufacturing, AI supports process control, quality assessment, real-time monitoring, waste reduction, predictive maintenance, and production optimization (Agrawal et al., 2025; Canatan et al., 2025; Zhang et al., 2025). Across agri-food supply chains, AI can improve logistics, planning, transparency, and resource efficiency (Tsolakis et al., 2023). Its integration with technologies such as blockchain can also strengthen trusted information flows among supply-chain actors (Reitano et al., 2025). Generative AI further extends these applications through knowledge support, communication, advisory services, and decision assistance (Krupitzer, 2024; Paul and Machavaram, 2025c). Overall, AI is evolving from task-specific applications toward an interconnected capability spanning multiple stages of the food system.

However, technological capability alone does not ensure successful or equitable AI adoption. Data quality, infrastructure, affordability, skills, interoperability, scalability, and institutional capacity remain key barriers (Benfenati et al., 2025; Dutta et al., 2026). These constraints are often greater in resource-constrained settings, particularly for smallholder farmers with limited access to infrastructure, finance, digital skills, and technical support (Nxumalo and Chauke, 2025). They are also interconnected: weak infrastructure can affect data quality and access, while limited technical capacity can hinder the use of AI outputs in decision-making. AI should therefore be viewed as a socio-technical capability whose contribution to food-system transformation depends on the interaction of technology, data, infrastructure, human capabilities, institutions, and governance.

2.3 Analysis of research patterns and gaps

A critical review of the 2020–2026 literature shows a shift from conventional precision agriculture toward more integrated applications of machine learning, deep learning, computer vision, robotics, generative AI, and intelligent automation. Recent studies demonstrate advances in fruit detection, maturity assessment, peduncle segmentation, object localization, and robotic harvesting across capsicum and other crops (Paul et al., 2024, 2026; Paul and Machavaram, 2024, 2025a,b; Verma et al., 2026). However, research remains concentrated mainly on agricultural production and food processing, with less attention to AI integration across post-harvest handling, distribution, logistics, and other connected food-system functions (Agrawal et al., 2025; Ayoub, 2026; Canatan et al., 2025; Munialo, 2025; İçöz, 2026). Technological development therefore appears to be advancing faster than understanding of how AI applications interact across the wider food system.

Methodological limitations also remain. Many studies evaluate AI at the model, task, pilot, or controlled-environment level, making it difficult to determine whether reported outcomes transfer across different agro-ecological, organizational, and socio-economic settings (Munyanyi and Pooe, 2024; İçöz, 2026). Differences in datasets, modelling approaches, evaluation metrics, and platform interoperability further limit comparison and replication (Laterza et al., 2025; Ayoub, 2026). High predictive accuracy or technical performance therefore does not necessarily demonstrate that an AI application can be effectively implemented, sustainably scaled, or produce similar outcomes across different food-system contexts.

The literature is also conceptually fragmented. Existing studies provide substantial evidence on AI applications and their potential operational benefits, but often examine technological capability separately from organizational readiness, infrastructure, human capabilities, governance, sustainability, and resilience (Yang et al., 2025; Munialo, 2025; Laterza et al., 2025; Bayisa and Zewudu, 2026). Yet these factors interact in shaping adoption and implementation. The literature therefore explains what AI can do more clearly than how and under what conditions its capabilities translate into sustained food-system transformation.

The key research gap is therefore not whether AI can improve specific food-system tasks, but how technological capabilities interact with organizational, institutional, infrastructural, human, governance, sustainability, and resilience factors to shape adoption, scaling, and broader transformation. Existing research provides limited explanation of how these conditions jointly enable or constrain the transition from technical capability to sustained food-system transformation. Addressing this gap requires a theoretical perspective that connects technological and social systems with the processes through which innovations are adopted and diffused. The following section therefore integrates Socio-Technical Systems Theory and Diffusion of Innovation Theory as the foundation for the Integrated Dynamic Socio-Technical AI Food Systems Transformation Framework.

2.4 Theoretical framework

2.4.1 Socio-technical systems (STS) theory

Socio-Technical Systems (STS) Theory explains how technological innovations interact with social, organizational, and institutional systems to shape performance and sustainability (Trist and Bamforth, 1951; Trist, 1981). It views technical and social subsystems as interdependent, meaning that technology cannot be optimized in isolation. (Appelbaum, 1997) Effective performance therefore depends on alignment between technology, human capabilities, organizational processes, governance, and the wider operating environment (Mumford, 1983; Abbas and Michael, 2026). As shown in Figure 1, socio-technical systems comprise interconnected elements such as people, tasks, technologies, organizational structures, and the physical environment (Carayon et al., 2012). Changes in one element can affect the others. Technology can reshape organizational routines and human behavior, while culture, institutional capacity, governance, and user capabilities influence how technologies are adopted, used, and sustained (Malatji et al., 2019; Mumford, 2003). Technological performance therefore emerges from the wider system rather than from the technology alone. (Gorejena et al., 2016).

In food systems, STS Theory helps explain why technically similar AI applications can produce different outcomes across contexts. Precision agriculture, automated quality control, predictive maintenance, and intelligent supply-chain management depend not only on technical performance but also on infrastructure, data quality, organizational readiness, user competence, institutional support, and regulatory capacity. STS Theory therefore provides a useful basis for examining how these interacting conditions shape AI implementation and its contribution to food-system transformation.

However, STS Theory provides less insight into how AI innovations are adopted and spread among actors over time. A complementary perspective is therefore needed to explain the processes that influence adoption and diffusion across organizations and food systems. Diffusion of Innovations Theory addresses this dimension and complements the systemic explanation provided by STS Theory.

2.4.2 Diffusion of innovations (DOI) theory

Diffusion of Innovations (DOI) Theory explains how new technologies, practices, and ideas spread and are adopted within social systems over time (Rogers, 2003). While STS Theory focuses on the socio-technical conditions shaping implementation, DOI explains how innovations spread across individuals, organizations, and communities. Five key attributes influence adoption: relative advantage, compatibility, complexity, trialability, and observability. DOI also distinguishes between innovators, early adopters, the early majority, the late majority, and laggards, reflecting differences in adoption timing.

In AI-enabled food systems, DOI helps explain why farmers, processors, distributors, policymakers, and other actors adopt technologies such as machine learning, predictive analytics, computer vision, robotics, and AI-enabled supply-chain platforms at different rates. Technologies perceived as beneficial, compatible with existing practices, and relatively easy to use are more likely to be adopted than those viewed as costly, complex, or poorly aligned with organizational needs (Garcia-Aviles, 2020; Rogers, 2003). DOI therefore provides insight into the behavioral, organizational, and communication processes that shape AI adoption and diffusion across food systems.

However, adoption alone does not guarantee successful implementation or sustained transformation. AI technologies may diffuse but fail to deliver expected benefits where infrastructure, governance, digital capabilities, or institutional support are inadequate. DOI therefore explains how and why AI innovations spread, while STS explains how socio-technical conditions shape their implementation and outcomes. Neither theory alone fully explains AI-enabled food-system transformation. Integrating them recognizes that sustained transformation depends on both the diffusion of AI innovations and the socio-technical conditions that enable their effective use. This complementarity provides the theoretical foundation for the integrated framework proposed in this study (see Figure 2).

2.5 Integrated socio-technical AI food systems transformation framework

Building on Socio-Technical Systems (STS) Theory and Diffusion of Innovations (DOI) Theory, this study proposes the Integrated Dynamic Socio-Technical AI Food Systems Transformation Framework (Figure 3). STS explains the interdependence between technology and social systems, including people, organizations, infrastructure, and institutions, while DOI explains how innovations are adopted and diffuse through social systems (Trist and Bamforth, 1951; Trist, 1981; Rogers, 2003). Although complementary, neither theory alone fully explains how AI capabilities become embedded, scaled, governed, and translated into sustained food-system transformation.

The framework also draws on digital-agriculture, responsible-AI, and food-system perspectives. Digital-agriculture frameworks emphasize the application of digital technologies, data, automation, and decision support; responsible-AI frameworks focus on transparency, accountability, privacy, fairness, and human oversight; and food-system frameworks connect production, processing, distribution, consumption, sustainability, resilience, and food security. While valuable, these perspectives provide largely partial explanations of how AI capability interacts with socio-technical readiness, innovation diffusion, and governance to shape system-level outcomes.

The proposed framework addresses this fragmentation through five interconnected components: AI capability, socio-technical readiness, innovation diffusion, governance and ethics, and food-system transformation outcomes. AI capability includes machine learning (Jordan and Mitchell, 2015), deep learning, computer vision, predictive analytics, robotics, generative AI, digital twins, and IoT-enabled systems. Socio-technical readiness captures infrastructure, data quality, affordability, human capabilities, organizational processes and culture, and institutional capacity. Innovation diffusion represents the processes through which AI is accepted, adopted, institutionalized, and scaled among food-system actors. Governance and ethics encompass data governance, transparency, accountability, privacy, equity, human oversight, and regulation. Transformation outcomes include productivity and operational efficiency, food safety, reduced food loss, environmental sustainability, supply-chain resilience, and food security.

The framework’s novelty lies not in introducing new constructs, but in integrating them and defining their roles within a dynamic transformation process. AI capability provides the technological potential; socio-technical readiness enables or constrains implementation; innovation diffusion shapes adoption and scaling; and governance and ethics influence implementation and the distribution of benefits and risks. Together, these interactions shape food-system transformation outcomes.

Importantly, these relationships are dynamic rather than linear. Transformation outcomes can influence organizational learning, institutional adaptation, governance, and subsequent adoption and scaling, creating feedback across the framework. The proposed relationships should therefore be understood as conditional and mutually reinforcing rather than as statistically established mediation or moderation effects. They remain conceptual propositions requiring empirical testing.

The framework thus challenges the assumption that technological capability or adoption alone leads to food-system transformation. Similar AI technologies may produce different outcomes because food systems vary in infrastructure, data, human capabilities, organizational and institutional readiness, governance, and diffusion conditions. Its main theoretical contribution is an integrated pathway linking AI capability, readiness, adoption and diffusion, implementation, governance, and scaling to system-level transformation, while recognizing feedback from outcomes to subsequent adaptation and diffusion. This provides a conceptual basis for examining how, why, and under what conditions AI can contribute to sustainable, resilient, equitable, and food-secure transformation across different food-system contexts. Table 1 compares the proposed framework with the main theoretical and analytical approaches relevant to AI-enabled food-system transformation, highlighting how existing approaches provide important but partial explanations of this process.

Table 1. Comparison of the proposed framework with existing approaches.

The proposed framework is conceptual and has not yet been empirically validated. Its relationships should therefore be treated as theoretically informed propositions drawn from STS, DOI, and patterns in the reviewed literature, rather than as established causal relationships. Future research should test these propositions across diverse food-system and institutional contexts using expert assessment, comparative case studies, longitudinal research, surveys, or structural equation modelling.

3 Methodology

This study employed a systematic literature review (SLR) to critically examine AI’s contribution to food-system transformation, focusing on applications, opportunities, implementation challenges, governance, and future research directions. An SLR provides a structured, transparent, and reproducible approach to identifying, evaluating, and synthesizing multidisciplinary evidence while reducing potential selection bias (Xiao and Watson, 2019). The review followed the PRISMA 2020 guidelines for literature identification, screening, eligibility assessment, and inclusion (Page et al., 2021). The completed PRISMA 2020 checklist is provided in Supplementary Table S2.

3.1 Information source and search strategy

We conducted a comprehensive literature search across multiple electronic sources, including Google Scholar, Scopus, Web of Science, EBSCOhost, ScienceDirect, Emerald Insight, JSTOR, ProQuest, and SpringerLink. The search combined three concept blocks covering AI technologies, food-system domains, and transformation outcomes. The core Boolean search string was: (“artificial intelligence” OR “machine learning” OR “deep learning” OR “computer vision” OR robotic OR automat OR “generative AI” OR “predictive analytics”) AND (“food system” OR “agri-food system” OR agriculture OR farming OR crop OR “food processing” OR “food supply chain” OR “food security”) AND (transform OR sustainable OR resilience OR innovation OR adoption OR governance OR application OR challenge OR barrier).

We adapted the search syntax to the indexing structure and functionality of each source. Where supported, searches were restricted to titles, abstracts, and keywords and limited to English-language publications from 2020 to 2026. Supplementary searches included reports from recognized international organizations, including the Food and Agriculture Organization (FAO). We also considered reports from the African Technology Policy Studies Network (ATPS) to capture relevant policy and institutional evidence. Supplementary Table S1 provides the complete source-specific search strings, field restrictions, filters, and search dates.

We conducted the initial search between January and March 2026 and up to the final updated search updated it in September 2026 to capture recent publications. The 2020–2026 publication window reflects recent advances in AI and growing attention to the governance, sustainability, resilience, infrastructure, and organizational conditions shaping its use across food systems. This restriction applied only to publications included in the systematic evidence synthesis. Seminal pre-2020 literature was retained where necessary to establish the review’s theoretical and conceptual foundations, particularly Socio-Technical Systems Theory and Diffusion of Innovations Theory.

3.2 Eligibility criteria and selection

Publications were initially screened using eligibility criteria defined before study selection. These covered publication period, language, food-system scope, relevance to AI, study design, publication type, and full-text availability. During manuscript revision, the publication-type criterion was broadened to better reflect the multidisciplinary evidence on AI-enabled food-system transformation. The revised criterion allowed empirical qualitative, quantitative, and mixed-methods studies; systematic and narrative reviews; conceptual studies; scholarly editorials, viewpoints, and perspectives; book chapters; conference proceedings; and reports from recognized institutional or international organizations. This amended criterion was applied consistently to all potentially eligible publications during the final evidence-base audit. All other eligibility criteria remained unchanged.

Under the final criteria, eligible publications were English-language sources published between 2020 and 2026, up to the final updated search, that contributed substantively to understanding AI-enabled food-system transformation or the technological, socio-technical, governance, sustainability, resilience, and institutional conditions shaping it. Publication within the specified period and an eligible publication type alone were insufficient; each source also had to contribute substantively to the review objectives.

Publications were excluded if they fell outside the relevant food-system, AI, or socio-technical scope; lacked substantive evidence or analysis relevant to the review objectives; fell outside the specified publication period; were not available in English; or lacked accessible full text. Blogs, news articles, promotional materials, general web content, other non-scholarly sources, and duplicate records were excluded. Seminal pre-2020 publications used to establish the theoretical and methodological foundations, including STS, DOI, thematic analysis, and systematic-review methodology, were cited in the manuscript but excluded from the 2020–2026 systematic evidence synthesis. Table 2 summarizes the inclusion and exclusion criteria.

Table 2. Inclusion and exclusion criteria.

Source: Authors eligibility criteria.

The search and publication-selection process followed the PRISMA 2020 reporting guidelines (Page et al., 2021). Database and supplementary searches identified 181 records. Before title and abstract screening, 50 records were removed, comprising 20 duplicates, 13 records that did not meet the publication-type eligibility criteria, and 17 records published outside the predefined eligibility period. The remaining 131 records underwent title and abstract screening, of which 24 were excluded because they did not meet the review objectives or eligibility criteria.

Full texts were successfully retrieved for all 107 remaining publications and assessed against the eligibility criteria. Following full-text assessment and the final evidence-base audit, 67 publications were excluded because they did not satisfy the final eligibility criteria. The systematic evidence synthesis therefore comprised 40 publications. The complete identification, screening, eligibility, and inclusion process is presented in the PRISMA flow diagram (Figure 4).

The review was not prospectively registered, and we did not publish a separate formal review protocol. Accordingly, there were no amendments to a registered or published protocol. Supplementary Table S2 provides the completed PRISMA 2020 checklist, and Supplementary Table S3 presents the methodological quality appraisal of the included studies. Given the substantial heterogeneity in study designs, geographical contexts, AI applications, and reported outcomes, statistical pooling was considered inappropriate; the evidence was therefore synthesised thematically and narratively. Formal assessment of publication bias was not undertaken because no quantitative meta-analysis or pooled effect estimates were produced. We also did not conduct certainty-of-evidence grading because the review synthesised heterogeneous evidence narratively rather than estimating a common intervention effect.

One reviewer conducted the initial study selection using the eligibility criteria established at the start of the review. During manuscript revision, we broadened the publication-type criterion as described above. The resulting final eligibility criteria were then applied during a comprehensive evidence-base audit. A second reviewer independently verified the final evidence base, including the application of the revised publication-type criterion and the full-text inclusion and exclusion decisions. Disagreements were resolved through discussion and consensus, with provision to consult a third reviewer where consensus could not be reached. Although a single reviewer conducted the initial screening, subsequent independent verification and reapplication of the final eligibility criteria provided an additional check on the consistency of the study-selection process.

3.3 Data collection and extraction

We used a structured data-extraction framework to ensure consistency across the included studies. Extracted data included authors and publication year, geographical context, evidence type and study design, AI technology, food-system application/domain, dataset or sample where applicable, methodology, principal outcomes and findings, reported or design-related limitations, implementation challenges, sustainability and governance considerations, and analytical contribution to the review. One reviewer conducted the initial extraction. During manuscript revision, a second reviewer independently verified the included publications against the final eligibility criteria, including the revised publication-type criterion, and checked the full-text inclusion and exclusion decisions. The team resolved disagreements through discussion and consensus, with consultation with a third reviewer when necessary.

The extracted data were synthesized using thematic analysis informed by Braun and Clarke’s (2006) six-phase approach. Following familiarization with the studies, one reviewer generated codes primarily inductively while remaining guided by the review objectives. The reviewer grouped related codes into preliminary themes and iteratively reviewed them against the evidence for coherence and distinctiveness. A second reviewer independently reviewed the coding structure and preliminary themes, and the team resolved interpretive differences through consensus. The data-extraction framework, coding categories, and theme-development stages were retained as an audit trail. The authors synthesized final themes narratively to identify recurring patterns, contextual differences, relationships among technological and socio-technical factors, and gaps in the literature.

After developing the themes, we mapped publications that explicitly contributed evidence to each of the five thematic narratives. Theme assignment was not mutually exclusive because individual publications could contribute to more than one theme. This mapping identified 34 theme-level evidence contributions from 22 unique publications. The remaining included publications contributed to the broader descriptive, contextual, conceptual, or evidence-mapping components of the review but were not explicitly cited as direct evidence within the five thematic narratives. Supplementary Table S4 maps the analytical contribution of all 40 included publications. We cross-checked the thematic mapping against the coding structure and extracted evidence to ensure consistency between the reported themes and their evidentiary basis.

3.4 Quality and risk of bias assessment

We assessed the methodological quality of eligible empirical studies using the Mixed Methods Appraisal Tool (MMAT) (Hong et al., 2018). We applied the relevant criteria according to study design. In line with MMAT guidance, we did not assign an overall numerical score to studies. Instead, we considered each criterion separately to identify methodological strengths and limitations. Methodological quality was not used as a threshold for excluding studies that met the eligibility criteria. Rather, it informed the interpretation and weighting of evidence during the synthesis. We gave greater weight to findings from methodologically stronger empirical studies, while interpreting findings from studies with methodological limitations or insufficient reporting more cautiously. As the review included different types of evidence, MMAT was applied only to eligible empirical studies. Review articles were assessed for methodological rigor and transparency, while scholarly editorials, viewpoints, perspective articles, book chapters, and institutional reports were assessed according to their analytical relevance, transparency of argument, source credibility, and contribution to the review objectives. These sources were not treated as equivalent to empirical evidence and were used primarily to support conceptual, policy, governance, and contextual interpretation (see Table 3).

Table 3. Thematic characteristics of evidence contributing to the five review themes.

Theme assignments are non-mutually exclusive because individual publications may contribute evidence to more than one theme. The table identifies publications explicitly contributing to the five thematic narratives; other included publications provide descriptive, contextual, application-mapping, or research-gap evidence. The complete analytical contribution of the 33 included publications should be reported in Supplementary Table S4. Source: Authors’ synthesis based on the systematic evidence review.

As shown in Table 4, the reviewed evidence was unevenly distributed geographically. Of the 40 publications included in the systematic evidence synthesis, 25 (62.5%) were global, international, or not confined to a specific national context. Country-specific evidence was concentrated primarily in India (n = 7; 17.5%) and South Africa (n = 3; 7.5%), while Africa-focused regional or multi-country evidence (n = 2; 5.0%) and United States-specific evidence (n = 2; 5.0%) were comparatively limited. One publication (2.5%) combined global and India-specific evidence. These findings demonstrate that the literature does not provide a geographically balanced basis for a direct comparison between technologically advanced and developing-country food systems. Accordingly, geographical differences identified in this review are interpreted as context-specific patterns within the available evidence rather than as generalisable differences between these broad country groups.

Table 4. Geographical distribution of publications included in the review.

Source: Authors’ synthesis based on the systematic evidence review.

4 Results and discussion

This section presents and discusses the findings from systematic evidence synthesis. The section begins by examining the geographical distribution of the evidence base to establish the contexts represented in the reviewed literature, then discusses the five interrelated themes identified through the thematic synthesis. The findings are discussed in relation to the study’s research question and theoretical framework, with particular attention to the socio-technical, governance, and sustainability conditions that shape AI-enabled food-system transformation.

The thematic synthesis subsequently identified five interrelated themes explaining how AI contributes to food-systems transformation. To improve transparency about the evidentiary basis for the thematic findings, Table 5 presents the distribution of evidence contributions across the five themes. Theme assignments were non-mutually exclusive because individual publications could contribute evidence to more than one thematic category.

Table 5. Distribution of evidence contributions across the five thematic findings.

Theme assignments are non-mutually exclusive; therefore, an individual publication may contribute evidence to more than one theme. The 33 entries represent theme-level evidence contributions, not 33 unique publications. These contributions derive from 21 unique publications explicitly cited in support of the five thematic findings. The remaining included publications contribute to descriptive mapping, contextual interpretation, application characterization, or identification of research gaps. Percentages total 99.99% because of rounding. Source: Authors’ synthesis based on the systematic evidence review.

4.1 AI as an enabler of operational efficiency and productivity

The reviewed evidence identifies operational efficiency and productivity as prominent areas of AI application across food systems. Machine learning, computer vision, robotics, IoT, and predictive analytics are increasingly applied to crop monitoring, maturity assessment, harvesting, processing, and post-harvest handling. In capsicum production, AI has been used for fruit detection and segmentation, growth-stage classification, counting, peduncle detection, and maturity assessment (Paul et al., 2024, 2026; Paul and Machavaram, 2024). Robotic harvesting and handling systems extend these capabilities towards physical automation, while graph-augmented computer vision demonstrates their potential under complex field conditions (Paul and Machavaram, 2025b; Verma et al., 2026). Beyond application-specific advances in computer vision and robotic harvesting, AI supports broader agricultural functions including crop and soil monitoring, disease and pest detection, weather and yield prediction, irrigation management, and agricultural automation (Javaid et al., 2023). Across the wider food system, AI is also associated with crop monitoring, predictive maintenance, quality control, demand forecasting, and resource optimization (Ayoub, 2026; Canatan et al., 2025; Galali et al., 2026; Laterza et al., 2025).

However, the reviewed studies assess AI outcomes in different ways. Technical studies mainly report gains in detection, prediction, automation, and efficiency, while implementation-focused studies question whether these gains can be sustained or scaled across different settings. This is especially relevant in resource-constrained contexts, where weak infrastructure, fragmented data systems, high costs, limited technical expertise, and institutional capacity can hinder adoption (Munialo, 2025; Munyanyi and Pooe, 2024; İçöz, 2026). Thus, technical studies tend to demonstrate task-specific performance, while broader food-system studies highlight the conditions that shape successful implementation and scaling.

Within the proposed framework, this distinction separates AI capability from food-system transformation. Improvements in accuracy, productivity, quality control, and automation show technological capability, but system-level transformation requires these gains to extend beyond isolated applications and deliver sustained operational outcomes. Operational efficiency and productivity should therefore be interpreted as potential transformation outcomes rather than automatic consequences of AI adoption.

4.2 AI, resilience, sustainability, and food security

The reviewed evidence positions AI as a potentially important enabler of food-system resilience, sustainability, and food security through predictive analytics, climate-smart agriculture, intelligent resource management, and early-warning systems (Bayisa and Zewudu, 2026; Galali et al., 2026; Ndhlovu, 2025). These applications can support improved forecasting, resource allocation, production planning, and responses to environmental risks. Their relevance is heightened by increasing climate-related pressures. Vegetable production, for example, is increasingly exposed to heat stress, water scarcity, altered rainfall patterns, and related production risks, strengthening the potential value of precision monitoring, predictive analytics, and adaptive crop management (Paul and Machavaram, 2026).

However, the evidence is less consistent regarding whether these technological capabilities translate into greater resilience or food security across different implementation contexts. While application-oriented studies demonstrate the potential of AI for forecasting, resource allocation, and adaptive production, studies examining resource-constrained settings identify limitations in digital infrastructure, institutional capacity, finance, and technical capabilities that can restrict adoption and scaling (Munialo, 2025; Ozor et al., 2023). These findings suggest that differences in reported outcomes partly reflect the contexts in which AI is implemented. Given the uneven geographical distribution of the reviewed evidence, however, these patterns should not be interpreted as systematic differences between technologically advanced and developing-country food system.

The environmental implications of AI also reveal an important tension. While AI can improve resource efficiency, its deployment requires energy, computing infrastructure, data storage, water, and hardware. In the United States, expansion of AI server infrastructure has been projected to generate an additional 24–44 Mt. CO₂-equivalent annually and require 731–1,125 million m³ of water per year between 2024 and 2030 (Xiao et al., 2025). Conversely, AI-enabled lighting and climate-control systems in plant factories can reduce energy use relative to conventional approaches (Decardi-Nelson and You, 2024). Sustainability should therefore be assessed across the AI lifecycle rather than inferred from application-level efficiency gains alone, particularly because rebound effects may offset some resource-efficiency improvements as adoption expands (Haruna et al., 2026; United Nations Environment Programme, 2024).

Taken together, the evidence shows that AI can improve forecasting, resource management, and adaptive decision-making, but its broader resilience and sustainability benefits remain uncertain. While application-level studies emphasize efficiency and adaptation, system- and lifecycle-focused studies highlight infrastructure demands, environmental costs, and possible rebound effects (Decardi-Nelson and You, 2024; Haruna et al., 2026; Xiao et al., 2025; United Nations Environment Programme, 2024). These outcomes should therefore be assessed across the wider food system rather than inferred from individual AI applications.

4.3 Data as a strategic resource for AI transformation

Across the reviewed literature, data emerge not merely as an input into AI systems but as a strategic resource shaping whether AI applications can be implemented and scaled. Data quality, accessibility, interoperability, and governance influence applications ranging from predictive modelling and precision agriculture to quality assurance and supply-chain management (Munyanyi and Pooe, 2024; Polcyn, 2025). Evidence from the Thai tuna supply chain further demonstrates the importance of mapping material, data, and information flows and identifying key data elements required to support AI-enabled digital supply-chain operations (Tsolakis et al., 2023). Conversely, fragmented datasets, incompatible information systems, restricted data access, and weak governance arrangements can limit the reliability and operational usefulness of AI-generated outputs. İçöz (2026) and Munialo (2025) similarly identify limited data availability, poor interoperability, and fragmented information systems as potential barriers to implementation in resource-constrained contexts.

Data-related challenges also extend to the reliability of AI models in operational environments. Unrepresentative or biased datasets can reduce model performance when systems are transferred across crops, regions, seasons, or production conditions, creating domain-shift and generalization problems (Pai et al., 2025; Fu et al., 2025). Explainability and uncertainty assessment are therefore important for interpreting predictions and supporting appropriate human oversight. Operational deployment must also account for computational requirements and edge-versus-cloud trade-offs, particularly where connectivity and computing capacity are limited (Pai et al., 2025). Cybersecurity, secure data exchange, and continued monitoring for model degradation or drift further require AI systems to be maintained, updated, and, where necessary, retrained rather than treated as one-time technological deployments.

Overall, the evidence from the review highlights the importance of data quality, accessibility, and interoperability, but also shows that more data do not necessarily lead to reliable AI. Fragmented, unrepresentative, poorly governed, or incompatible data can limit performance, especially as operating conditions change. Effective data capability therefore requires not only access to data, but also sound governance, representativeness, validation, security, monitoring, and ongoing model management.

4.4 Ethical, governance, and socio-economic considerations

The reviewed evidence demonstrates that AI-enabled food-systems transformation raises governance and ethical questions alongside its potential operational benefits. AI can strengthen decision-making, traceability, food safety, and regulatory compliance, but these capabilities are accompanied by concerns relating to algorithmic bias, data privacy, transparency, accountability, unequal digital access, and regulatory preparedness (Ikram et al., 2024; Laterza et al., 2025; Polcyn, 2025). Governance is therefore not simply a regulatory response to technological adoption; it forms part of the institutional environment through which AI implementation and its outcomes are shaped. Studies examining resource-constrained settings identify fragmented institutions, limited regulatory capacity, inadequate digital infrastructure, affordability barriers, and insufficient institutional support as potential constraints on responsible implementation (Munialo, 2025; Munyanyi and Pooe, 2024).

However, the geographical evidence is insufficiently balanced to establish that these challenges are systematically greater in developing countries than in technologically advanced settings. They should therefore be interpreted as context-specific constraints within the available literature rather than as generalizable characteristics of broad country groupings. The socio-economic implications are similarly conditional. Although AI may improve productivity, transparency, food safety, and decision-making, unequal access to finance, infrastructure, data, and technical expertise may influence which actors capture these benefits. Smaller producers and resource-constrained organizations may face higher relative adoption barriers, creating a risk that technological diffusion could reproduce existing inequalities. Responsible implementation therefore requires consideration not only of whether AI improves performance but also of who can access it, whose data inform it, who participates in its deployment, and how benefits and risks are distributed.

Relatively, the reviewed studies concur that responsible AI requires attention to bias, privacy, transparency, accountability, and equitable access, but offer limited evidence on which governance approaches work best. This is important because effective governance may differ across technologies, actors, regulations, and institutional settings. Future research should therefore move beyond identifying ethical concerns to examining which governance mechanisms work, for whom, and under what conditions.

4.5 Human-AI collaboration and capacity building

The reviewed evidence indicates that human capability remains central to AI implementation despite increasing technological sophistication. AI can extend analytical capacity, automate selected activities, and support decision-making, but the available evidence more strongly supports complementarity between human expertise and AI than wholesale substitution of human decision-making (Thomas et al., 2022; Ayoub, 2026). Effective implementation therefore depends not only on access to AI technologies but also on the ability of users to interpret, evaluate, and integrate AI-generated outputs into organizational practice.

Munialo (2025) and İçöz (2026) identify shortages of digital skills and technical expertise, along with limited organizational readiness and institutional support, as barriers to adopting and diffusing AI in resource-constrained settings. Across the contexts represented in the review, variations in human and organizational capacity may therefore contribute to differences in implementation outcomes. Organizations with stronger digital capabilities, relevant training, effective change-management processes, and institutional support may be better positioned to integrate AI into established workflows. Importantly, digital capability extends beyond operating AI systems to understanding their limitations, evaluating uncertainty and potential bias, and determining when human judgment should supplement algorithmic recommendations.

These findings reinforce the socio-technical character of AI-enabled transformation. Socio-Technical Systems Theory emphasizes the interdependence of technological systems, human actors, organizational routines, and institutional structures. Diffusion of Innovation Theory, in turn, helps explain how knowledge, compatibility, perceived usefulness, communication, and organizational readiness influence adoption. Capacity building should therefore be treated as an integral component of AI implementation. Investment in digital literacy, professional training, interdisciplinary collaboration, organizational learning, and institutional capability can strengthen the capacity of food-system actors to adopt and use AI responsibly and sustainably.

4.6 Cross-theme synthesis: towards an integrated understanding of AI-enabled food systems transformation

Across the five themes, the review consistently shows that AI can improve specific food-system functions. These include monitoring, prediction, quality control, resource management, automation, and decision support. However, it is less clear whether these gains lead to sustained system-level transformation. Technical studies mainly report performance improvements, while implementation, governance, and socio-technical studies highlight challenges with integration, scaling, institutionalization, and equitable access. These differences reflect variations in analytical scale, study design, technology, and implementation context.

The review also reveals important trade-offs. Automation and predictive tools can improve efficiency but increase dependence on reliable data, computing infrastructure, and technical expertise. AI can support resource efficiency and climate adaptation while creating additional energy, water, carbon, and material demands (Decardi-Nelson and You, 2024; Haruna et al., 2026; Xiao et al., 2025; United Nations Environment Programme, 2024). More data can also improve prediction and coordination while raising concerns about bias, privacy, cybersecurity, interoperability, and model reliability. AI benefits and risks can therefore coexist.

Context further shapes these outcomes. Benefits are more evident when AI addresses clearly defined tasks supported by suitable data and infrastructure. Greater uncertainty arises when technologies are applied across different locations, organizations, production conditions, or user groups, especially where domain shift, fragmented data, limited connectivity, institutional constraints, or skills gaps affect implementation (Fu et al., 2025; İçöz, 2026; Munialo, 2025; Pai et al., 2025). Given the uneven geographical coverage of the evidence, treat these patterns as context-specific rather than generalized differences between technologically advanced and developing-country food systems.

Finally, these findings support the Integrated Dynamic Socio-Technical AI Food Systems Transformation Framework by distinguishing AI capability from system-level transformation. AI can improve performance, but broader transformation depends on how these capabilities are integrated into organizational processes and interact with data, human expertise, institutions, and governance. The evidence therefore points to a conditional and dynamic transformation pathway rather than a universal link between AI adoption and positive outcomes. These patterns are informed by the reviewed evidence rather than established causal relationships and require further testing across different food-system and institutional contexts (see Table 6).

Table 6. Summary of key findings and policy implications of AI applications in food systems.

Source: Authors’ synthesis based on the systematic evidence review.

5 Conclusion and recommendations

This review shows that AI can contribute to food-systems transformation through improved operational efficiency, resilience and sustainability, data-driven decision-making, and human–AI collaboration. However, these benefits depend on infrastructure, data readiness, organizational and human capabilities, institutional support, and effective governance. AI is therefore best understood as an enabler of transformation rather than an independent driver of change.

The paper contributes the Integrated Dynamic Socio-Technical AI Food Systems Transformation Framework, combining Socio-Technical Systems Theory and Diffusion of Innovation Theory. The framework explains how AI capability, socio-technical readiness, innovation diffusion, and governance interact to shape adoption and transformation outcomes. These relationships are conceptual rather than empirically established and provide propositions for future testing.

The findings suggest that effective AI adoption requires investment not only in technology but also in digital infrastructure, interoperable data systems, skills, organizational capacity, and governance. Policy and implementation strategies should also address affordability, inclusion, accountability, privacy, and environmental impacts while remaining responsive to different institutional and food-system contexts.

This review has several limitations. Restricting the search to English-language publications from 2020 to 2026 with accessible full texts may have excluded relevant earlier, non-English, or inaccessible studies. Despite searching multiple databases and supplementary sources, we may have missed relevant evidence indexed elsewhere. Initial screening by one reviewer introduced potential selection bias, although a second reviewer independently verified the final evidence base and full-text eligibility decisions. Considerable variation in study designs, geographical contexts, AI applications, and outcomes limited direct comparison and made meta-analysis inappropriate.

The synthesis therefore identifies patterns and contextual differences rather than pooled estimates of AI effectiveness. Uneven geographical coverage further limits the findings’ generalizability. Finally, the Integrated Dynamic Socio-Technical AI Food Systems Transformation Framework is conceptual and has not yet been empirically validated. Its proposed relationships should therefore be treated as theoretically informed propositions rather than established causal mechanisms. Overall, these limitations call for cautious interpretation and further empirical testing across diverse food-system and geographical contexts.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fsufs.2026.1892956/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Lochan Singh, Indian Institute of Science (IISc), India

Reviewed by: Chandrabhan Seniya, Manipal University Jaipur, India

Ayan Paul, Indian Institute of Technology Kharagpur, India