Agentic AI in Latin America: characterization of the structural gap and proposal of a contextual adaptation framework

1 Introduction

The emergence of agentic AI, systems that autonomously plan, execute multi-step tasks, interact with external tools, and adapt their behavior based on environmental feedback, represents one of the most significant shifts in applied artificial intelligence since the introduction of transformer-based language models (Abou Ali et al., 2025). Frameworks such as ReAct, promoted by Yao et al. (2022), AutoGPT, and LangGraph have demonstrated that agents based on large language models (LLMs) can perform complex software engineering, research, and decision-making tasks with minimal human intervention. By late 2025, agentic AI systems had become widespread in enterprises, with significant productivity gains reported alongside new management challenges (Williams, 2026).

However, the global discourse on agentic AI has developed primarily within the constraints of a specific technological and socioeconomic context: high-bandwidth connectivity, predominantly English-speaking linguistic environments, consolidated regulatory ecosystems, and a large pool of technical talent. These conditions characterize the environments in which the most advanced agentic systems are designed, trained, and evaluated. They do not, however, characterize most of the world's operational environments, and are particularly misaligned with the structural reality of Latin America and the Caribbean (Bandi et al., 2025).

Latin America presents a paradox. On the one hand, the region demonstrates exceptionally high enthusiasm for AI and strong consumer adoption: according to ECLAC's Latin American Artificial Intelligence Index (ILIA 2025) CEPAL (2025), the region accounted for 14% of global visits to AI solutions in 2025, a figure disproportionate to its share of total internet users, while ranking third worldwide in generative AI application downloads. On the other hand, the structural conditions necessary for the effective deployment of agent systems remain significantly underdeveloped. The region captures only 1.12% of global AI investment, despite representing 6.6% of global GDP (CEPAL, 2025). More than half of LAC countries lack the critical digital infrastructure necessary to support advanced AI training or the sustained operation of agents. Ninety percent of regional supercomputing capacity is concentrated in a single country: Brazil. Furthermore, the linguistic foundations of current LLMs systematically underrepresent varieties of Spanish, Portuguese, and the hundreds of indigenous languages spoken across the region (Rivero, 2026).

This article argues that deploying standard agentic AI architectures in Latin America without deliberate contextual adaptation is not merely suboptimal—it is structurally deficient and risks exacerbating existing digital inequalities. It makes two core contributions: first, it offers a systematic characterization of the gap between dominant agentic AI architectures and the operational requirements of the Latin American context, organized around four structural dimensions—infrastructure, linguistics, regulation, and access equity; and second, it proposes the LATAM Contextual Adaptation Framework (LATAM-CAF), a conceptual three-layer architecture designed to guide the adaptation of agentic AI systems in emerging economy contexts.

The contributions of this work are structured around three research questions. The first examines the architecture of existing agent systems: What are the dominant design assumptions in current agentic AI frameworks, and how do these assumptions constrain their deployment in the Latin American context? The second analyzes the regional environment: What structural dimensions of the Latin American context constitute the most significant barriers to effective agentic AI deployment? The third moves from diagnosis to design: What framework components are necessary and sufficient to adapt agentic AI architectures to the multidimensional constraints identified in Latin America?

The research sequence moves from diagnostic to constructive: Section 2 reviews the background on agentic AI architectures and their inherent contextual assumptions; Section 3 presents the research methodology; Section 4 characterizes the four-dimensional structural gap; Section 5 proposes the LATAM-CAF framework; Section 6 validates the framework against regional evidence; Section 7 discusses implications, limitations, and future research directions; and Section 8 presents the conclusions.

2 Background: agentic AI architectures and their contextual assumptions

2.1 The architecture of agentic AI systems

Agentic AI represents a paradigm shift from passive, task-specific AI tools. Abou Ali et al. (2025) define agentic systems as those exhibiting: (1) proactive planning and goal decomposition, (2) contextual memory across interactions, (3) dynamic tool use and external API integration, and (4) adaptive behavior based on environmental feedback. These characteristics distinguish modern LLM-based agents from prior conceptions of AI agency grounded in symbolic Belief-Desire-Intention (BDI) models or Partially Observable Markov Decision Processes (POMDP) (Khan et al., 2022).

Contemporary agent-based frameworks can be organized around three dominant paradigms (Abou Ali et al., 2025; Khamis, 2025): symbolic architectures, which rely on explicit rule systems and excel in safety-critical domains such as healthcare; neural architectures, which leverage LLM-based reasoning for adaptive, data-rich environments; and hybrid neuro-symbolic architectures, which represent an emerging frontier combining deliberative planning with statistical inference. The ReAct framework, which interleaves reasoning traces and action calls, has become a foundational pattern for neural agent-based systems (Yao et al., 2022). LangGraph, AutoGPT, MetaGPT, and related frameworks extend this pattern to multi-agent orchestration, where specialized agents collaborate under a central coordinator (Guo et al., 2024).

Critically, existing reviews in this field identify significant research gaps around governance models and hybrid architectures, but do not address the geographic and socioeconomic context-specificity of existing designs (Piccialli et al., 2025). This omission reflects a broader tendency in generative AI research: the implicit assumption that globally deployed systems operate under conditions similar to those in which they were developed.

2.2 Implicit assumptions in dominant architectures

A critical analysis of leading agent-based frameworks reveals a set of structural assumptions that are rarely made explicit but are deeply embedded in design decisions. These include: high-bandwidth, low-latency connectivity (presupposed for real-time API calls and cloud-based reasoning chains); English as the primary language of interaction (reflected in training data distributions and benchmark design); mature regulatory environments with established AI accountability and data governance frameworks; and the availability of organizational contexts with sufficient resources to absorb AI integration costs (Guo et al., 2024).

Research on multilingual NLP model performance has documented consistent capability degradation in languages other than English: even for Spanish (one of the world's most widely spoken languages), financial NLP models specifically trained on English data exhibit significant performance gaps on Spanish financial NLP tasks (Hong et al., 2024). For lower-resource languages, the problem is more severe. LLaMA's official model documentation explicitly acknowledges that performance in unsupported languages may fall below safety and utility thresholds. The indigenous languages of Latin America, including Quechua, Guaraní, Nahuatl, Aymara, and hundreds of others, are virtually absent from the training corpora of commercial LLMs (Guo et al., 2024).

When agent-based systems are deployed under the assumption of conditions that do not hold (low connectivity, languages other than English, absence of regulatory frameworks, limited institutional capacity) the result is not simply poor performance. It is the systematic exclusion of marginalized populations from the benefits of AI and the risk of embedding new forms of algorithmic discrimination into fundamental social and economic processes.

2.3 Prior work on AI in emerging economy contexts

A modest but growing body of literature addresses AI deployment in low- and middle-income countries. This work has documented infrastructure barriers to AI adoption in Sub-Saharan Africa, language model failures in Indic languages, and governance challenges in Southeast Asian markets (Castaño Castaño, 2025; Cuskley et al., 2024). However, the Latin American context has received comparatively limited systematic attention, particularly with respect to agent-based AI. Existing regional AI research has focused primarily on predictive AI applications in agriculture and health, and on comparisons of national AI strategies (Aviles-Valenzuela et al., 2025).

2.4 Operational definitions of key terms

Four technical terms that will be used repeatedly throughout this article are defined below: *Agentic AI refers to systems capable of autonomous planning, multi-step task execution, and tool use with minimal human intervention, as opposed to single-turn predictive models. Multilingual NLP refers to the capacity of a language model to process and generate text with comparable accuracy across multiple languages, rather than exhibiting the performance degradation in non-English languages. Edge-hybrid execution refers to a deployment mode in which inference is distributed between local, low-resource computing and cloud-based resources, contingent on connectivity availability. Risk-based regulatory framework* refers to AI governance models, such as the EU AI Act, that calibrate compliance obligations to the potential harm of a given AI application rather than imposing uniform requirements across all use cases.

3 Research design and methodology

3.1 Conceptual framework development methodology

This article adopts a conceptual framework development methodology, recognized in the information systems and AI research communities as a rigorous approach to theoretical contribution when empirical data collection is premature (in the case of nascent phenomena) or structurally limited (Jaakkola, 2020). Conceptual frameworks establish the research domain, identify core constructs, specify their relationships, and generate testable propositions for subsequent empirical validation. This approach is particularly appropriate for the current research question: the gap between agentic AI architectures and Latin American requirements is an emerging phenomenon that lacks the empirical foundation necessary for quantitative modeling, but is sufficiently well-documented in secondary data to support rigorous conceptual analysis (Marc-Lim, 2026).

The methodology employed covers three phases. In Phase 1 (Architecture Analysis), a structured analysis of dominant agentic AI frameworks is conducted to identify their implicit contextual assumptions, drawing on published architectural specifications, benchmark documentation, and model cards. In Phase 2 (Gap Characterization), evidence from regional data sources is synthesized—including ECLAC's ILIA Index of CEPAL (2025), JP Morgan's regional infrastructure analysis (2025), the LatamGPT initiative reported by Rivero (2026) and the European Commission's AI governance documentation (2026); to operationalize four dimensions of structural misalignment. In Phase 3 (Framework Synthesis), the findings of Phases 1 and 2 are integrated into LATAM-CAF, a three-layer framework whose components are explicitly grounded in the identified gap dimensions.

3.2 Data sources and quality criteria

Secondary data sources were selected according to three criteria: (1) institutional credibility and methodological rigor; (2) recency, with preference for sources published or updated between 2024 and 2026; and (3) regional specificity, prioritizing sources that provide country- or region-level disaggregation for Latin America and the Caribbean. To ensure methodological consistency, the review followed a structured search and selection protocol across Scopus, Web of Science, IEEE Xplore, ACM Digital Library, and Google Scholar for institutional gray literature. Searches covered the period from January 2023 to March 2026 and employed a Boolean combination of (“agentic AI” OR “AI agents” OR “autonomous AI systems”) AND (“Latin America” OR “LATAM” OR specific countries) AND (infrastructure OR multilingual OR regulation OR governance OR adoption). Source selection was further guided by explicit inclusion and exclusion criteria, emphasizing peer-reviewed research, official regulatory documents, and institutional reports with publicly documented methodologies, while excluding non-institutional commentary and vendor-promotional materials. The quality and relevance of all institutional sources were assessed through an adapted CRAAP framework (Currency, Relevance, Authority, Accuracy, and Purpose) (Blakeslee, 2004).

Primary sources include: ECLAC's ILIA 2025, covering 19 LAC countries across more than 100 sub-indicators; JP Morgan Private Bank's regional digital infrastructure analysis (2025); TechPolicy.Press documentation and analysis of the LatamGPT project (2026); Peru's AI regulatory framework, Law No. 31814 (2023) and its Regulations; Brazil's AI regulatory framework, Bill No. 2338 (2023); and scientific literature on multilingual LLM performance from ACM, IEEE, and arXiv (Chang et al., 2024).

4 Characterization of the structural gap

This section analyzes the structural gap between dominant AI architectures and the Latin American operational context across four dimensions, each corresponding to a distinct category of architectural assumptions that are not met across much of the region; as outlined in Table 1.

Table 1. Structural Gap between agentic AI architectures and the Latin American context.

The infrastructure and linguistic dimensions present the highest implementation risk, as they operate at the architectural level and cannot be resolved solely through regulatory or policy measures.

4.1 Infrastructure heterogeneity

Standard agent-based architectures presuppose constant high-bandwidth connectivity as a prerequisite for real-time cloud API calls, multi-agent message exchange, and continuous model inference. This assumption is not met across much of Latin America. In 2025, fewer than 40% of Latin Americans regularly used 4G services, and approximately half had access to fixed broadband—a figure that conceals deep urban-rural divides, with only 4 in 10 rural Latin Americans having access to basic internet connectivity (JP Morgan, 2025).

The infrastructure challenge extends beyond connectivity to computational capacity. The ILIA CEPAL report (2025) documents that 90% of the region's supercomputing capacity is concentrated in Brazil, leaving the remaining 18 countries with very limited capacity for model training, fine-tuning, or large-scale inference. This concentration has direct implications for agent deployment: multi-agent workflows that rely on shared computational resources or high-frequency API calls become economically and technically unfeasible for most regional organizations.

Another dimension of infrastructure heterogeneity is cloud service fragmentation. While Brazil leads in fiber optic deployment and data center investment (with more than 87.8 million fiber optic connections as of Q2 2025), cloud service provision remains uneven and costly relative to local economic conditions across most of the region (Prescott, 2025).

4.2 Linguistic marginalization

The linguistic dimension of this gap is perhaps the most systematically underestimated in the technical literature. The dominant natural language processing models (NLPs), which form the basis of reasoning in contemporary agentic systems, are trained on corpora in which Spanish and Portuguese together account for approximately 2%–3% of total content (Rivero, 2026). This underrepresentation has measurable performance consequences: multiple studies have documented that even state-of-the-art models such as GPT-4 perform significantly worse on Spanish financial tasks compared to their English equivalents, and domain-optimized Spanish models outperform GPT-4 on Spanish financial NLP benchmarks (Djagba and Odinakachukwu, 2025).

For indigenous languages, the situation is more critical. The region is home to hundreds of indigenous languages, including Quechua (approximately 8 million speakers), Guaraní (between 6 and 12 million), Nahuatl and Aymara (1.7 million each, among others), that are virtually absent from LLM training data. The emerging LatamGPT initiative, led by Chile's National Center for Artificial Intelligence (CENIA), represents the first serious regional effort to address this gap, but is still in its early stages of development (Rivero, 2026).

4.3 Regulatory fragmentation

Los Agentic AI systems, operating with significant autonomy (executing multi-step tasks, interacting with third-party services, and making consequential decisions), require a regulatory framework that defines permitted actions, accountability mechanisms, and liability allocation. Existing architectures have been developed primarily with reference to the EU AI Act and emerging U.S. federal frameworks, which share key features: risk-based classification, mandatory human oversight for high-risk applications, and established data protection regimes.

The Latin American regulatory landscape is significantly more fragmented. As of mid-2025, only Peru had enacted a comprehensive AI-specific legal framework: Law 31814 and its September 2025 Regulations establish risk-based prohibited practices, human oversight requirements, regulatory sandbox environments, and differentiated implementation timelines for micro and small enterprises, making Peru one of the most advanced regulatory environments in the Global South. Brazil has advanced Bill No. 2338 (2023) on AI with a risk-based structure, but it had not yet been enacted as of early 2026. Most LAC countries lack binding AI legislation or a clear roadmap for its development.

This regulatory heterogeneity poses a specific challenge for agent-based architectures. An agent designed to operate within the permissive assumptions of an unregulated environment may exhibit behaviors that violate the constraints of more advanced regulatory frameworks, and vice versa. Moreover, the absence of regulatory frameworks across much of the region does not imply that agents can operate without constraints: informal institutional norms, general civil liability, and sector-specific regulations (e.g., financial services, healthcare) impose contextual requirements that standard agent-based architectures do not account for.

4.4 Asymmetric access and adoption patterns

The fourth dimension of the gap refers to the structure of AI adoption in Latin America, which exhibits a distinctive pattern that differs from high-income country contexts and has direct architectural implications. ECLAC's ILIA 2025 report characterizes regional AI adoption as broad but shallow: the region consumes significantly, but produces and integrates comparatively little. Some 65% of Latin American consumers already use AI tools, with adoption projected to grow 18% in 2024, reaching 40%, surpassing the global average in enthusiasm (Loaiza, 2025). However, this use is concentrated in consumer-oriented tools with low technical requirements, rather than the deep organizational integration that characterizes agent deployment.

The investment asymmetry is structurally significant. Latin America accounts for 6.6% of global GDP and 14% of global visits to AI solutions, but attracts only 1.12% of global AI investment (CEPAL, 2025). This investment deficit translates into a shortage of the technical talent needed to develop, adapt, and maintain AI systems: ILIA 2025 documents that while AI literacy levels are relatively high across the region, advanced technical training levels represent approximately one-quarter of literacy levels, and the emigration of specialized talent further reduces the regional talent pool.

The adoption pattern also reveals a structural dependency risk. Regional AI consumption is conducted primarily through external platforms, with the regional population dependent on global AI services rather than locally developed alternatives. This creates a systemic vulnerability: any disruption in the availability, pricing, or service terms of global platforms has disproportionate consequences for LAC users and organizations, in the absence of regional alternatives.

While the four dimensions identified above are presented sequentially for analytical clarity, they do not operate independently. Infrastructure constitutes a foundational constraint that conditions the feasibility of the remaining three dimensions: limited connectivity and compute capacity directly determine which deployment modes are viable for access and adoption initiatives, and shape whether local language models can be hosted and maintained at the edge. The linguistic dimension, in turn, exerts a direct influence on access and adoption patterns, since populations whose primary language is inadequately supported by available models are structurally excluded from the consumer-facing AI tools that currently drive regional adoption. The regulatory dimension interacts primarily with access and adoption through compliance cost: jurisdictions with more developed AI governance frameworks, such as Peru, impose differentiated implementation requirements that affect smaller organizations disproportionately. The coupling between the linguistic and regulatory dimensions is comparatively weak, as language model performance and regulatory frameworks have so far evolved largely independently of one another across the region.

Figure 1 summarizes these directional dependencies, which the LATAM-CAF explicitly incorporates through its Context Profiling layer, where the four dimensions are not assessed in isolation but as a jointly conditioning environmental profile.

5 LATAM contextual adaptation framework (LATAM-CAF)

5.1 Design principles

The LATAM-CAF is designed according to three guiding principles derived from the gap characterization in Section 4. The principle of contextual primacy establishes that the operational context of deployment must be assessed and incorporated before task execution begins—not as an afterthought, but as a first-order architectural input. The principle of graceful degradation establishes that agentic systems deployed in Latin America must maintain functional utility across a range of infrastructure quality levels, rather than failing entirely when connectivity or computation assumptions are not met. The principle of regulatory modularity establishes that agent autonomy levels and permitted action spaces must be configurable to comply with the heterogeneous and rapidly evolving regulatory landscape across more than 19 jurisdictions.

5.2 Framework architecture

The LATAM-CAF consists of three hierarchically organized layers, each addressing a distinct subset of the identified gap dimensions; as shown in Table 2.

Table 2. LATAM-CAF: three-layer framework specification.

The three layers are designed to operate sequentially; omitting the L1 context profiling invalidates the adaptive logic of L2 and L3.

5.3 Layer 1: context profiling (L1)

The first layer of LATAM-CAF performs an environmental assessment before initiating any agent task. The Context Profiler classifies the deployment environment across four dimensions corresponding to the four gap dimensions identified in Section 4: connectivity class (ranging from high-bandwidth cloud to intermittent mobile and offline-only); linguistic context (identifying the primary language of interaction and the availability of adequate LLM support for that language); regulatory class (mapping the deployment jurisdiction to one of three regulatory maturity tiers based on the presence and scope of AI-specific legislation); and institutional capacity class (assessing the organizational and technical resources available for agent oversight and error correction).

The output of L1 is a context profile object that governs the configuration of all subsequent components. This approach builds on the concept of environment-aware agent design developed in recent multi-agent systems research, but extends it specifically to the socioeconomic and linguistic dimensions of deployment in emerging economies (Yang et al., 2024).

5.4 Layer 2: adaptive execution (L2)

The second layer implements the core adaptive mechanisms of LATAM-CAF. It consists of two main components; the Degradation Engine and the Language Router. The Degradation Engine receives the connectivity and capacity classifications from L1 and dynamically configures the execution architecture accordingly. In high-connectivity, high-capacity contexts, standard cloud-based agent architectures can operate without modification. In intermediate contexts, the engine switches to a hybrid-edge mode, distributing inference between lightweight local models and cloud resources according to task complexity thresholds. In severely constrained contexts, the engine configures an offline execution mode using locally cached models and task queues for deferred execution.

The Language Router complements the Degradation Engine by addressing the linguistic dimension of the gap. When the L1 context profile indicates that the deployment language is Spanish, Portuguese, or an identified indigenous language, the router evaluates the confidence of the primary LLM's output on a sample of domain-relevant prompts. If confidence falls below a calibrated threshold, the router redirects reasoning tasks to language-specific fine-tuned modules; such as domain-adapted Spanish models for financial contexts, or the LatamGPT model for general Spanish and Portuguese tasks when available. This component implements a model routing architecture analogous to mixture-of-experts at the system level, but targeted specifically at linguistic competence rather than domain knowledge.

5.5 Layer 3: contextual evaluation (L3)

The third layer of LATAM-CAF provides a context-adapted evaluation suite for assessing the performance of agentic systems in Latin American deployment environments. Standard agentic benchmarks—including MMLU, HumanEval, SWE-Bench, and WebArena—were designed in and for high-income, English-speaking, and high-connectivity contexts, and fail to capture the most relevant performance dimensions for Latin American deployment. The L3 Contextual Evaluation Suite supplements these benchmarks with four Latin America-specific evaluation dimensions: linguistic accuracy in Spanish and Portuguese (measured with regional corpora, not English-translated benchmarks); task completion under latency constraints (simulating the connectivity conditions documented in Section 4.1); regulatory compliance (evaluating agent behavior against the risk-based prohibitions established in Peru's Law 31814 and Brazil's Bill No. 2338 federal framework); and institutional accessibility (assessing whether agent outputs are actionable by organizations with limited technical capacity, including MSMEs and public sector entities).

5.6 Scope conditions and boundary restrictions

LATAM-CAF is a conceptual reference architecture, not a deployment-ready implementation. Its applicability is subject to several conditions. First, the framework addresses the Latin American context as a structurally heterogeneous region and acknowledges that the four gap dimensions present with different intensity across countries: the infrastructure dimension is most acute in Central America and Andean countries; the linguistic dimension is most significant in communities with a high proportion of indigenous language speakers; the regulatory dimension is most advanced in Peru and Brazil; and the access/adoption dimension is most pronounced in the region's smaller, lower-income economies. Second, the framework is designed for general-purpose agentic applications and would require domain-specific extension for safety-critical sectors such as healthcare or financial services. Third, the framework addresses the technical and structural barriers to deployment and does not directly address the behavioral and social dimensions of AI adoption—including trust, digital literacy, and organizational change management.

6 Theoretical conceptual corroboration of the LATAM-CAF against secondary evidence

To evaluate the conceptual coherence of the LATAM-CAF, each of its three layers was cross-referenced with real-world situations documented in the literature and supported by secondary evidence; the results are summarized in Table 3. It is important to clarify the epistemic status of this exercise: the evidence presented below was instrumental in constructing the framework's dimensions in Section 4, and its use here therefore constitutes corroboration of internal consistency rather than independent empirical validation. The LATAM-CAF remains, at this stage, a conceptual reference architecture rather than an implemented or empirically tested system. Establishing its operational validity (through prototype implementation, case studies, or expert elicitation methods such as the Delphi technique) constitutes a necessary direction for future research.

Table 3. LATAM-CAF validation against regional evidence sources.

The five evidence sources are independent of each other, reinforcing the convergent validity of the framework's four dimensions.

The infrastructure claims underpinning the L1 Context Profiler and L2 Degradation Engine are supported by multiple independent data sources that converge on consistent conclusions. JP Morgan's 2025 regional analysis, ILIA 2025, and sectoral research on fixed network deployment document the same structural condition: significant but uneven connectivity progress since 2012, alongside a persistent and significant gap relative to the conditions assumed by dominant agent architectures. The supercomputing concentration documented by ILIA 2025 (90% in Brazil) directly supports the rationale for hybrid-edge execution modes as an unavoidable design requirement for regional deployment.

The linguistic evidence base for the L2 Language Router is particularly strong. The LatamGPT initiative represents a direct institutional response to the documented training data gap, confirming that this gap is recognized and consequential at the highest levels of regional AI policy. The FLARE-ES benchmark results provide quantitative confirmation that standard LLM performance degrades systematically on Spanish financial tasks, validating the technical rationale for language routing at the agent architecture level (Hebert et al., 2025).

The regulatory evidence supports LATAM-CAF's modular design. Peru's Law 31814 and its 2025 Regulations offer a concrete implementation case for risk-based AI governance in an emerging economy context, demonstrating both the feasibility and the specificity of regulatory requirements that agent systems must meet. The differentiated implementation timelines for micro and small enterprises in Peru's regulation directly validate the institutional capacity dimension of the framework.

The access and adoption evidence provided by CEPAL (2025) and Loaiza (2025) validates the framework's positioning: the region's pattern of broad but shallow adoption, concentrated in consumer-oriented tools, confirms that the primary challenge for agent-based system deployment is not demand, but structural readiness, precisely the gap that LATAM-CAF is designed to address.

7 Discussion

7.1 Theoretical contributions

This article makes three theoretical contributions to the AI research literature. First, it introduces the concept of architectural context-specificity as an explicit analytical category: the observation that AI system designs embed contextual assumptions that constitute structural constraints on their deployment outside their original environment. This concept extends existing work on algorithmic bias and fairness, which has focused primarily on training data and model outputs, to encompass the system architecture level. Second, it provides the first systematic four-dimensional characterization of the LATAM operational gap for agentic AI, grounded in recent regional evidence and linked to specific architectural implications. Third, it proposes LATAM-CAF as a principled framework for contextual adaptation, whose design principles and three-layer architecture can serve as a reference model for parallel work in other emerging economy contexts.

7.2 Practical implications

For AI practitioners designing or deploying agent systems in Latin America, LATAM-CAF offers practical guidance at multiple implementation levels. At the architectural level, the framework specifies which components require regional adaptation (context profiling, language routing, evaluation suites) and which can be adopted from existing global frameworks (core reasoning loops, tool integration patterns). At the procurement and policy level, the framework provides a structured vocabulary for evaluating vendor claims about regional readiness and for specifying minimum regional suitability requirements in public procurement.

The framework also has implications for the regional AI talent and infrastructure agenda. If the four identified gap dimensions are accepted as constitutive of the implementation challenge in Latin America, then regional AI investment should prioritize: edge-capable model deployment tools; Spanish and Portuguese model development (with a roadmap for indigenous language coverage); regulatory sandbox infrastructure to accelerate governance development; and Latin America-specific benchmark development. The LatamGPT initiative addresses the linguistic dimension; the LATAM-CAF framework provides a broader architectural context for this and similar initiatives.

7.3 Limitations and future research

The primary limitation lies in the conceptual nature of LATAM-CAF: the framework has not been empirically validated through implementation and deployment. Future work should develop prototypes of each framework layer and conduct empirical evaluations in deployment contexts representative of Latin America, including variations in connectivity class, linguistic context, and regulatory environment. Controlled experiments comparing the performance of standard agent architectures with and without LATAM-CAF adaptation would provide the quantitative evidence base needed for formal validation.

A further limitation pertains to the intra-regional heterogeneity of the Latin American context. The framework treats the region as structurally heterogeneous but addresses it through a unified adaptive architecture.

Finally, the framework addresses structural and technical barriers but does not directly model the human and organizational dimensions of agentic AI adoption in Latin America. Future work should integrate sociotechnical perspectives, including studies on AI trust, digital literacy, and organizational change in regional contexts, to offer a more complete picture of the implementation challenge.

8 Conclusion

This article argues that the global advance of agentic AI systems poses a specific and underrecognized challenge for Latin America and the Caribbean. The dominant agentic architectures emerging from leading AI research institutions embed structural assumptions, regarding infrastructure quality, linguistic environments, regulatory contexts, and organizational capacity, that are systematically misaligned with the operational reality of most Latin American deployment contexts. Deploying these architectures without deliberate contextual adaptation risks not only performance degradation, but also the amplification of existing structural inequalities by concentrating the benefits of AI agency in the most resource-rich segments of an already unequal regional landscape.

The LATAM Contextual Adaptation Framework (LATAM-MAC) is proposed as a well-founded response to this challenge. Its three-layer architecture (context profiling, adaptive execution, and contextual evaluation) provides a structured basis for designing, adapting, and evaluating agent systems that can function effectively and equitably under the heterogeneous conditions of the Latin American context. The framework draws on recent regional evidence from ECLAC, CENIA, and national regulatory developments, and is designed to evolve as that evidence base grows.

The implications of this work extend beyond Latin America. As agentic AI systems spread from their origins in the Global North to the diverse operational contexts of the Global South, the question of contextual adaptation will become a central element of equitable and effective AI deployment worldwide. LATAM-CAF represents a contribution to a methodology for addressing this question, a methodology urgently needed and for which this article hopes to provide a useful foundation.

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.

Footnotes

Edited by: Narcisa Medranda - Morales, Universidad Politecnica Salesiana - Sede Quito, Ecuador

Reviewed by: Evelio Geronimo, National Pedagogic University, Mexico

Laura Reyes Alardo, University Federico Henriquez y Carvajal, Dominican Republic