An Agentic LLM-Human Collaboration Framework for Enterprise Modeling in Resource-Constrained Organizational Contexts

1. 1. Introduction

Enterprise modeling (EM) plays a central role in organizational design and digital transformation by providing formal representations of business processes, structures, and strategies that support decision-making, systems development, and knowledge management [1]. Dominant modeling languages—BPMN for process flows, ArchiMate for enterprise architecture, and UML for software design—have reached high levels of methodological maturity. However, this maturity has developed predominantly within the epistemic context of large, digitally capable organizations in the Global North, where trained modeling engineers, stable technological infrastructure, and high organizational formalization are assumed as starting conditions.

Organizations in peripheral or resource-constrained environments—characterized by limited internet connectivity, low digital literacy, high labor informality, and linguistic diversity—remain largely absent from the EM literature as subjects of study or intended beneficiaries of design. This constitutes a significant equity gap: these organizations represent a large share of economic activity in the Global South and yet are systematically excluded from the benefits of structured process knowledge and enterprise transformation [2].

At the same time, the emergence of LLMs and multi-agent AI architectures opens new possibilities for AI-assisted modeling [3]. LLMs demonstrate remarkable capabilities in natural language understanding, process narrative elicitation, formal notation generation, and contextual explanation—capabilities that could substantially reduce the expertise barrier to enterprise modeling. The central theme of PoEM 2026, “Rethinking Enterprise Modeling in the Age of AI,” explicitly frames this duality: modeling for AI and modeling with AI as complementary directions that require critical examination.

This article addresses the intersection of these two phenomena. The research question is: How can agentic LLM systems be structured to support enterprise modeling in resource-constrained organizational contexts, where human modeling expertise is scarce and organizational conditions are informal? To answer this question, ALMEC (Agentic LLM-based Model Engineering Collaboration) is proposed, a conceptual framework that defines agent roles and human–AI collaboration structures throughout the enterprise model lifecycle.

The framework is grounded and validated through a case study of multimodal logistics organizations operating in the Alto Amazonas region (Loreto, Peru)—an empirically rich context characterized by dependence on river transportation, Indigenous multilingualism, and structural digital limitations. Although specific, this context is argued to be representative of a broad category of resource-constrained organizations globally, making ALMEC’s design principles transferable.

Article contributions: (1) A novel conceptual framework (ALMEC) for agentic LLM–human collaboration in EM, structured around four agent roles and four lifecycle stages; (2) a taxonomy of human–AI accountability and trust mechanisms by lifecycle stage; (3) an empirical instantiation in the Amazonian logistics context; and (4) design guidelines for EM in resource-constrained organizational environments.

Enterprise modeling encompasses a set of methods, languages, and tools for representing organizational knowledge in formal and computable structures [4]. Key paradigms include process-oriented modeling (BPMN), architecture-oriented modeling (ArchiMate, TOGAF), and goal-oriented modeling (i* framework, BMM). These methods share an implicit assumption of organizational formalization—documented processes, defined roles, and stable IT systems—which makes them poorly suited to contexts of informality and infrastructural fragility [5].

Research on EM adoption in small and medium-sized enterprises (SMEs) has documented barriers including technical complexity, lack of modeling expertise, and poor tool usability [6]. In the Global South, additional structural barriers—digital divide, linguistic diversity, and low institutional capacity—amplify these difficulties. The literature reports that EM adoption in emerging economies is significantly lower than in OECD countries, even after controlling for organizational size [7].

Large language models have demonstrated the ability to generate process descriptions, BPMN XML, and entity-relationship schemas from natural language inputs [8]. Recent work has explored LLM-assisted generation of BPMN models [9], ArchiMate pattern suggestion [10], and ontology population from unstructured text [11]. However, these contributions assume technically skilled users who can verify, correct, and refine model outputs—an assumption that fails in low-expertise contexts.

The specific challenge of explainability in AI-generated enterprise models has received limited attention. Transparency requirements differ fundamentally between technical stakeholders (who need explanations of model correctness) and operational stakeholders (who need explanations of process rationale) [12]. No existing framework addresses both simultaneously within an agentic architecture.

Agentic AI systems—autonomous agents that pursue goals through sequences of tool-mediated actions—have emerged as a productive paradigm for complex, multi-step tasks [13]. Multi-agent frameworks such as LangGraph, AutoGen, and CrewAI enable role specialization, inter-agent communication, and iterative refinement through structured workflows [14]. Applied to enterprise modeling, agentic architectures offer the possibility of decomposing the complex EM task into specialized subtasks—elicitation, formalization, validation, and explanation—each handled by a purpose-specific agent.

The human-in-the-loop (HITL) dimension of agentic EM is critical and insufficiently explored. Existing HITL frameworks in AI systems design [15] do not account for the knowledge asymmetry between expert AI agents and non-expert human stakeholders that characterizes resource-constrained contexts. This gap motivates the accountability structures incorporated into ALMEC.

1. 3. THE ALMEC FRAMEWORK

ALMEC (Agentic LLM-based Model Engineering Collaboration) is a conceptual framework that structures collaboration between specialized LLM agents and human process engineers throughout the enterprise model lifecycle. Table 1 presents the overall architecture as an agent–lifecycle matrix. The framework is organized along two dimensions: agent roles (vertical) and lifecycle stages (horizontal).

Table 1. ALMEC Agent–Lifecycle Matrix

Each ALMEC agent is implemented as an LLM instance with a role-specific system prompt, access to a defined set of tools, and a structured communication protocol with adjacent agents and human engineers.

▸ Elicitor Agent: Conducts semi-structured process interviews through natural language dialogue. In multilingual contexts, it operates in Spanish and Indigenous languages (Shawi, Kukama) using translation subroutines. It produces structured process narratives in SIR (Structured Intermediate Representation) format, consumable by the Formalizer.

▸ Formalizer Agent: Transforms SIR narratives into formal EM constructs. Depending on the modeling objective, the output may be BPMN 2.0 XML, ArchiMate 3.1 JSON-LD, or an OWL ontology. The Formalizer operates under organizational constraints (connectivity limitations, tool availability) that condition the selection of the notation.

▸ Validator Agent: Applies structural validation (syntactic correctness of formal models), semantic validation (consistency with business rules), and pragmatic validation (alignment between the model and stakeholder intent). In resource-constrained contexts, validation prioritizes high-impact process elements over exhaustive coverage.

▸ Explainer Agent: Generates natural-language explanations of model elements, validation results, and change rationales, targeting three stakeholder profiles: operational workers, middle managers, and process engineers. Explanations are calibrated according to literacy level and organizational role.

ALMEC defines five accountability checkpoints (AC) distributed throughout the lifecycle, at which human engineers must review, approve, or override agent outputs. These checkpoints are designed to preserve human accountability for the final model while maximizing automation of technically complex subtasks:

▸ AC-1 (Discovery Gate): Human approval of the elicited narrative before formalization begins.

▸ AC-2 (Formalization Review): Human inspection of the structure of the generated formal model.

▸ AC-3 (Validation Sign-off): Human endorsement of the validation report before model publication.

▸ AC-4 (Explainability Audit): Human verification that Explainer outputs are accurate and culturally appropriate.

▸ AC-5 (Maintenance Trigger): Human authorization of model updating following detected drift.

ALMEC Design Principle — Asymmetric Explainability

In resource-constrained contexts, the Explainer Agent bears a disproportionately high accountability burden compared with standard EM environments. When stakeholders lack modeling literacy, the interpretive gap between formal model and organizational reality can only be closed through continuous and culturally sensitive explanation—not through training or tool simplification. ALMEC treats explainability as a first-class modeling artifact, not an ex post facto add-on.

Alto Amazonas (Loreto, Peru) is a province accessible primarily through river transportation along the Huallaga and Marañón rivers, with Yurimaguas serving as a commercial and logistics hub. The region’s logistics organizations operate multimodal transportation chains integrating river vessels, light ground freight, and air cargo under conditions of seasonal variability, intermittent satellite connectivity, and a predominantly bilingual workforce (Spanish–Shawi or Spanish–Kukama) with limited adoption of digital tools.

Three representative organizations participated in the ALMEC instantiation: (1) a medium-sized freight company with 45 employees that manages cargo consolidation at the Port of Yurimaguas; (2) a cooperative of Indigenous river transporters (Kukama community, 28 active members) operating on the lower Huallaga; and (3) a regional logistics intermediary coordinating the palm oil supply chain from producers in San Martín to distributors in Iquitos.

The ALMEC framework was instantiated using a LangGraph multi-agent architecture with GPT-4o as the underlying LLM. Each agent was prompted with role-specific system instructions and given access to a shared organizational knowledge graph populated from the initial document analysis (transport manifests, route records, informal process descriptions).

Table 2. ALMEC Instantiation in the Alto Amazonas Case Study

Human engineers approved checkpoints AC-1 through AC-3 with modifications in 38% of cases, primarily cultural corrections to process terminology and corrections to descriptions of geographic routes. AC-4 (explainability audit) generated the highest modification rate (61%), with engineers reviewing Explainer outputs to remove technical jargon and adapt metaphors to Amazonian operational realities. AC-5 was autonomously triggered by the Validator Agent on three occasions, with human authorization obtained in all cases within 24 hours via WhatsApp.

ALMEC extends enterprise modeling theory in three directions. First, it introduces the concept of context-adaptive formalization, according to which the formal notation selected by the Formalizer Agent is determined by the infrastructural constraints of the deployment context, rather than solely by the modeling objective. This relativizes the tool–method relationship assumed in standard EM approaches.

The case study instantiation involves a single geographic context (Alto Amazonas) and three organizations, which limits the generalizability of the specific findings. The Elicitor Agent’s multilingual capability was limited to Spanish–Shawi translation; elicitation in Kukama was conducted through human intermediaries, introducing an interpretation overhead. Evaluation of ALMEC results relied on self-reported stakeholder agreement measures rather than formal model-quality metrics, representing a construct-validity limitation. Future research should include model-quality assessment by panels of EM experts.

The ALMEC case demonstrates that responsible AI deployment in enterprise modeling requires attention to contextual equity in addition to technical accuracy. In peripheral contexts, a highly accurate model that is inaccessible to its primary users—because it is formalized in a notation they cannot interpret or explained in a language they do not speak—does not provide organizational value. The Explainer Agent and ALMEC’s accountability checkpoints represent design mechanisms for equity-sensitive EM, with implications for PoEM’s engagement with the Global South and underrepresented organizational forms.

This article presented ALMEC, a conceptual framework for agentic LLM–human collaboration in enterprise modeling under resource-constrained conditions. Through a case study of multimodal logistics organizations in the Peruvian Amazon, it demonstrated that ALMEC’s four-agent architecture (Elicitor, Formalizer, Validator, Explainer) and five accountability checkpoints enable effective management of the model lifecycle in contexts where conventional EM adoption has historically failed. The framework contributes to PoEM 2026’s central theme—rethinking EM in the age of AI—by placing equity, explainability, and contextual adaptability as first-class design concerns.

Future work will include: (1) formal evaluation of ALMEC model quality against established EM quality frameworks (SEQUAL, QPER); (2) extension to additional resource-constrained contexts (rural healthcare, informal markets, Indigenous economic cooperatives); (3) development of an ALMEC reference ontology formalizing the taxonomy of agent roles; and (4) empirical study of human trust calibration at ALMEC accountability checkpoints in longitudinal deployments.

 Sandkuhl, K., Fill, H.-G., Hoppenbrouwers, S., Krogstie, J., Matthes, F., Opdahl, A. L., Schwabe, G., Uludag, Ö., & Winter, R. (2018). From expert discipline to common practice: A vision and research agenda for extending the reach of enterprise modeling. Business & Information Systems Engineering, 60(1), 69–80. https://doi.org/10.1007/s12599-017-0516-y

 Pereira, C. M., & Sousa, P. (2005). Enterprise architecture: Business and IT alignment. In Proceedings of the 2005 ACM Symposium on Applied Computing (pp. 1344–1345). ACM. https://doi.org/10.1145/1066677.1066980

 Fillottrani, P. R., & Keet, C. M. (2021). Toward AI-assisted enterprise modeling. Enterprise Modelling and Information Systems Architectures (EMISAJ), 16(2), 1–24.

 Frank, U. (2014). Multi-perspective enterprise modeling: Foundational concepts, prospects and future research challenges. Software and Systems Modeling, 13(3), 941–962. https://doi.org/10.1007/s10270-012-0273-9

 Ssebuggwawo, D., Hoppenbrouwers, S., & Proper, E. (2010). Assessing collaborative modeling quality based on modeling artifacts. In J. Ralyté, X. Franch, S. Brinkkemper, & S. Wrycza (Eds.), Advanced Information Systems Engineering Workshops (Lecture Notes in Business Information Processing, Vol. 68, pp. 307–319). Springer. https://doi.org/10.1007/978-3-642-17722-4_24

 Stirna, J., Persson, A., & Sandkuhl, K. (2007). Participative enterprise modeling: Experiences and recommendations. In O. Pastor & J. Falcão e Cunha (Eds.), Advanced Information Systems Engineering (Lecture Notes in Computer Science, Vol. 4495, pp. 546–560). Springer. https://doi.org/10.1007/978-3-540-72988-4_38

 Vom Brocke, J., Simons, A., & Cleven, A. (2011). Towards a business process-oriented approach to enterprise content management: The ECM-blueprinting framework. Information Systems and e-Business Management, 9(4), 475–496. https://doi.org/10.1007/s10257-009-0124-6

 Bellan, P., Dragoni, M., & Ghidini, C. (2022). Extracting business process entities and relations from text using pre-trained language models and in-context learning. In Enterprise Design, Operations, and Computing (pp. 182–199). Springer. https://doi.org/10.1007/978-3-031-17604-3_11

 Bernardi, M. L., Casciani, A., Cimitile, M., & Marrella, A. (2024). Conversing with business process-aware large language models: The BPLLM framework. Journal of Intelligent Information Systems, 62(3), 1607–1629. https://doi.org/10.1007/s10844-024-00898-1

 Hinkelmann, K., Gerber, A., Karagiannis, D., Thoenssen, B., van der Merwe, A., & Woitsch, R. (2016). A new paradigm for the continuous alignment of business and IT: Combining enterprise architecture modelling and enterprise ontology. Computers in Industry, 79, 77–86. https://doi.org/10.1016/j.compind.2015.07.009

 Hertling, S., & Paulheim, H. (2023). Transformer-based semantic relation typing for knowledge graph integration. In A. Gangemi, R. Navigli, M.-E. Vidal, P. Hitzler, R. Troncy, L. Hollink, A. Tordai, & M. Alam (Eds.), The Semantic Web: ESWC 2023 (Lecture Notes in Computer Science, Vol. 13870, pp. 105–121). Springer. https://doi.org/10.1007/978-3-031-33455-9_7

 Hevner, A. R., March, S. T., Park, J., & Ram, S. (2004). Design science in information systems research. MIS Quarterly, 28(1), 75–105. https://doi.org/10.2307/25148625

 Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., Zhao, W. X., Wei, Z., & Wen, J.-R. (2024). A survey on large language model based autonomous agents. Frontiers of Computer Science, 18, 186345. https://doi.org/10.1007/s11704-024-40231-1

 Wu, Q., Bansal, G., Zhang, J., Wu, Y., Zhang, S., Zhu, E., Li, B., Jiang, L., Zhang, X., & Wang, C. (2023). AutoGen: Enabling next-gen LLM applications via multi-agent conversation framework. arXiv. https://doi.org/10.48550/arXiv.2308.08155

 Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. International Journal of Human–Computer Interaction, 36(6), 495–504. https://doi.org/10.1080/10447318.2020.1741118