Olga Cherednichenko

Entrepreneurship and Sustainability Issues is a peer-reviewed journal which publishes original research papers and case studies

In the contemporary landscape of Software Engineering and Business Process Management (BPM), the integration of generative artificial intelligence has precipitated a paradigm shift from manual, deterministic specification to automated, probabilistic generation. While offering scalability, this transition introduces a fundamental volatility known as the "Probabilistic Gap"—the chasm between the fluid, high-variance output of Large Language Models (LLMs) and the strict, zero-tolerance syntactic requirements of execution engines like DMN (Decision Model and Notation). This paper addresses the "Struc-Bench Paradox," highlighting the limitations of transformer architectures in generating complex structured data without rigid orchestration. The study formally defines and implements the "Dispatcher," a pivotal control plane component designed to function as an intelligent resource arbiter and quality gatekeeper within a neuro-symbolic architecture. The theoretical framework shifts the economic focus from Baumol’s Cost Disease, which addresses production speed, to Boehm’s Law of Software Economics, which emphasizes the exponential cost of defects propagated to production. To operationalize this, the Dispatcher represents a discrete deterministic process modeled using Cost-Colored Petri Nets rather than Finite State Machines (FSMs). The Petri Net formalism allows for precise modeling of concurrency, state accumulation, and the strict enforcement of "Retry Budgets," thereby mathematically guaranteeing system termination and preventing infinite loops of costly regeneration. The architectural implementation utilizes a "Test-First" generation philosophy: the system first synthesizes validation criteria (JSON test cases) utilizing Schema Injection and RAG, and subsequently grounds the generation of DMN logic (XML) in these pre-validated scenarios. Experimental analysis was conducted using a controlled set of 200 generation cycles to evaluate two distinct error-recovery strategies: Strategy A (Independent regeneration of DMN tables only) and Strategy B (Joint/Dynamic regeneration of both DMN and Test Cases). Quantitative results demonstrate that Strategy B is economically superior, achieving a 6.06% reduction in total cost and an 8.44% reduction in token consumption compared to the independent patching approach. The findings indicate that simultaneous regeneration empowers the LLM to resolve semantic incoherence and hallucinations more effectively than iterative repairs, prioritizing logical consistency over partial code retention. The study concludes that the Dispatcher effectively bridges the neuro-symbolic divide by transforming validation from a post-production manual review into a pre-production automated cycle. By enforcing a "Stop-Loss" mechanism driven by economic constraints, the framework minimizes the Total Cost of Ownership and serves as a critical "Trust Proxy," mitigating automation bias and ensuring that AI-generated artifacts meet the rigorous reliability standards required for enterprise deployment.

This paper explores the use of large language models (LLMs) to support collaborative business intelligence in the tourism domain through two key tasks: extracting travel-related tags from user queries and generating Cypher queries for accessing knowledge graphs. We focus on evaluating the performance of compact and efficient LLMs, aiming to balance accuracy with computational feasibility. To assess tag extraction, we evaluated Phi-3 Mini, LLaMA 3.2, and Gemma 3 using the DeepEval framework with G-Eval scoring. Phi-3 Mini showed the best balance between accuracy and efficiency, while Gemma 3 achieved the highest scores at the cost of increased resource usage. For Cypher query generation, we tested more powerful models: Mistral Small 3.1, Phi-4, Gemma 3, and ChatGPT-4o. ChatGPT-4o achieved the highest correctness, while Mistral Small demonstrated the best trade-off among smaller models. Our results suggest that lightweight LLMs are suitable for basic natural language processing tasks, but structured query generation remains challenging and requires stronger models. Further research is needed to improve the reliability of generated queries and to develop robust validation mechanisms. This study introduces a comparative evaluation of lightweight and standard LLMs specifically applied to collaborative business intelligence in the tourism domain. It highlights the feasibility of using compact LLMs for natural language processing tasks while demonstrating the challenges of structured query generation, which requires more powerful models.

The paper addresses the problem of formalization and automated processing of textual business rules in decision support systems, taking into account the existing data structure and the contextual constraints of the decision-making domain. The study proposes mathematical models covering three key stages: identifying the structural components of a textual business rule, generating a formal DMN decision table, and validating the syntax and semantics of the resulting model in accordance with logical consistency rules. The structural analysis model extracts conditions, actions, and dependencies while incorporating the available data schema; the DMN generation model maps textual constructs to decision table elements with respect to the context of the decision support system; the validation model detects logical inconsistencies, incompleteness, and coherence errors in the formalized model. A prototype implementation is presented, enabling experimental evaluation of the proposed models. The results demonstrate correct formalization, completeness of business rule transformation into the DMN table, and effective detection of structural, syntactic, and semantic errors in the generated decision model.