Decision-making in industrial environments increasingly depends on the ability to align operational data with strategic goals. Data-driven approaches often focus on the available data while overlooking the specific informational needs of decision makers. To bridge this gap, this paper proposes a multi-driven approach that integrates both data-driven and requirements-driven perspectives, supported by the use of Large Language Models (LLMs). Through prompt-based interactions, LLMs generate conceptual and analytical models from metadata and user requirements, accelerating the modeling process while ensuring relevance and coherence through the integration of both data and analytical requirements. The approach is applied to a real-world industrial case characterized by operational complexity, high variability, and technological heterogeneity. The results show how the combined use of LLMs and a structured modeling approach can support the development of analytical systems that are technically feasible and strategically aligned.
In industrial environments, decision-making processes increasingly rely on the ability to extract
actionable insights from operational data. Modern manufacturing settings employ advanced analytics to
monitor operations in real time, detect anomalies, and optimize workflows - with Industry 5.0 studies
reporting improvements in decision outcomes of up to 46% following the implementation of data-driven
methods [
To address this gap, this paper proposes an approach that combines a multi-driven perspective,
integrating the strengths of both data-driven and requirements-driven approaches. The goal is to align
what data reveals with what users need, ensuring that the resulting analytical models are both feasible
and meaningful. This dual orientation enables organizations to capture not only the structure of the
available datasets but also the strategic, decision, and information goals of stakeholders. Although
such alignment can be performed manually, it is often labor-intensive and error-prone, particularly
in complex industrial contexts characterized by diverse, fragmented data sources. To overcome this
challenge, we propose leveraging Large Language Models (LLMs) to assist in constructing analytical
models. Hence, a major contribution of this work is the proposal and evaluation of a multi-driven
and LLM-assisted analytical modeling approach. This contribution have been grounded in a research
process guided by the Design Science Research Methodology for Information Systems [
Recent work has demonstrated that LLM agents can automate multiple stages of industrial
analytics, including process understanding, concept extraction, and iterative self-refinement, efectively
accelerating the end-to-end data-to-dashboard pipeline [
This paper applies a structured approach that combines process modeling, conceptual and analytical data modeling, and requirements analysis, supported by prompt-based interactions with an LLM. It is applied to a real-world industrial case known for high variability, technological diversity, and constant adaptive challenges. The results show how this multi-driven approach supports targeted and efective decision-making by aligning business objectives with data.
This paper is structured as follows. Section 2 analyzes related works. Section 3 describes the use case where the proposed approach will be applied. Section 4 details the steps of the proposed approach. Section 5 shows an instantiation of the use case. Section 6 presents the conclusions and future work.
The increasing complexity of modern industrial environments poses significant challenges for efective
decision-making. Frequent changes in operational settings, safety constraints, and regulatory
requirements demand iterative cycles of requirements engineering, process modeling, and data alignment.
In such settings, aligning the semantics of available data with user information needs is essential but
dificult to automate - particularly in the conceptual design of Data Warehouses (DWs) or domain
models [
Recent studies have explored the use of LLMs to assist engineers in automating or accelerating parts
of this process. Nouri et al. [
Domain modeling has been another area of interest. Yang et al. [
Other works have focused on supporting conceptual modeling tasks. Cámara et al. [
Overall, while LLMs have shown potential in helping with requirements engineering and model creation, studies consistently indicate that human involvement is necessary to ensure that the meanings are correct and fit with specific industry standards. This supports the development of hybrid, humanin-the-loop approaches — as explored in this paper — where LLMs accelerate the initial modeling efort, but outputs are validated and refined collaboratively.
Unlike previous works, which often focus on isolated modeling tasks, our approach integrates both data-driven and requirements-driven perspectives in a unified pipeline that explicitly targets the alignment of user requirements with available industrial data, supported by LLMs.
Colep Packaging1, a company of the RAR Group2, is the Iberian leader in general line packaging and an
important European supplier of aerosols. With production in Portugal, Spain, Poland, and Mexico, and
more than 700 employees, it operates with a vertical integration model, developing metal and plastic
solutions for several markets [
The core production flow is represented in Fig. 1, using a BPMN (Business Process Model and Notation) diagram. This visual model details each stage of the process and explicitly highlights key decision points and sub-processes. For example, depending on the type of product or volume, diferent printing technologies (ofset or digital) may be selected, each with its own implications in terms of setup time, cost, and flexibility. The diagram helps clarify the operational logic while also making visible the possible paths and conditional branches that increase the complexity of execution and planning. It also makes it easier to identify which parts of the process are more sensitive to variability and where analytical support is more critical, since BPMN explicitly models alternative paths and decision points (e.g., the choice between ofset and digital printing), highlighting where execution may diverge.
The company combines ofset and digital printing technologies, each with distinct operational tradeofs. Ofset printing is more suitable for large production volumes and ensures high color consistency, while digital printing, although slower and more limited in print quality, allows for quick adjustments, minimal setup time, and greater flexibility in small or customized batches. Managing this hybrid setup requires the coordination of materials, design files, and production schedules, increasing the need for adaptable and responsive planning systems.
Beyond the technological setup, the complexity of the use case becomes more evident when considering the dynamic nature of the industrial scenario. The company faces constant changes that afect production and decision-making processes, such as the integration of new equipment, frequent reconfiguration of the shop floor layout, supply chain disruptions, coordination between multiple work centers, adaptation to new labor models (e.g., shift systems or four-day weeks), and compliance with sustainability and energy eficiency targets. These factors create a high-variability environment that challenges traditional planning approaches and requires flexible, data-driven systems capable of integrating information from multiple sources.
To support decision-making under these conditions, Colep Packaging relies on digital tools such as Manufacturing Execution Systems (MES) and industrial simulation models, which allow for real-time monitoring, predictive analysis, and optimization of production flows. However, even with such systems in place, the integration between business objectives and operational processes remains a challenge. The variability and heterogeneity of data sources, combined with evolving user needs, highlight the need for approaches that can simultaneously consider what the data reveal (data-driven) and what the decision-makers require (requirements-driven). This context of high complexity justifies the relevance of applying a multi-driven decision-making approach, as proposed in this paper.
To address the challenges of aligning business goals with complex data environments, this paper applies a structured, LLM-supported approach that integrates process and data modeling and requirements analysis. This section presents the approach in two stages: the first describes the overall approach that combines data and user requirements to build a multi-driven analytical model; the second details the structured prompts used to guide each interaction with the LLM.
Although diferent implementations can be followed, this paper adopts the analytical data model as a
DW system [
As mentioned, two analytical data models can be identified, one derived from the data and the other from the requirements, both considering the identified conceptual data model of the domain. In the several tasks, an LLM supports this approach with increased eficiency (the models can be obtained in a short period of time) and handling the complexity that is usual in complex industrial settings, with a high diversity of data sources and data sets. These models are later integrated by the LLM. Along all the steps, the data engineer validates the proposals and makes the necessary suggestions/corrections towards the model to be implemented. Following this two-fold path ensures that the data available for analysis guides the identification of the analytical data model, but this model is then refined considering the user’s needs. In case of a lack of data for specific user requirements, this approach ensures that the data engineer and the organization are aware of it, taking proper actions to collect such data in the future.
With the multi-driven analytical data model, the LLM supports the identification of useful
visualizations. To be possible, this work considers that the user requirements are specified in iStar [
Besides the data and user requirements as input, another key component of the approach is the prompt strategy to interact with the LLM. To obtain useful results and avoid hallucination [15], the prompts must detail the input of each step and the expected output. For the input, this work considers: • Data: include the activities of the business processes and the diferent events associated with those activities. Activities usually relate diferent entities of the domain, such as Customer and Order for place order, with a set of time-series events. To avoid mixing entities, activities, and events, the prompt must clarify the several concepts present in the input data. • User requirements: include the SGs to be addressed and the way they are supported by the DGs and IGs. These concepts should be clarified in the prompt, which must also highlight that the IGs commonly point to business or process indicators associated with the main entities of the domain or with the supporting activities, respectively. • Visualization tasks: useful visualizations must align the analytical task, the type of data, and the cardinality of the data, among other characteristics. Pointing to or adopting a specific taxonomy or classification for visualization tasks enhances the resulting visualizations.
For data, only the metadata needs to be made available (ensuring confidentiality, since no sensitive values are shared). This is important given privacy and confidentiality concerns. This metadata includes the available tables and the corresponding attributes. The proposed prompts are next detailed. 4.2.1. From the Event Data to the Data-driven Conceptual Model Here, the metadata of the available data is used as input to obtain the data-driven conceptual model. The prompt is as follows:
Consider a dataset with the following tables and the associated attributes. With this metadata, infer the several relationships between the entities of the domain, and propose an entity-relationship diagram that represents the conceptual model of the domain. This conceptual model must include the entities and their attributes, as well as the relationships among these entities. The conceptual model can only include entities, attributes, and relationships inferred from the available metadata. Also, consider that the entities of the domain have events that are occurrences of the activities that afect entities over time. These events or activities cannot be considered entities of the domain. The tables and their attributes are: «to be defined». 4.2.2. From the Data-driven Conceptual Model to the Data-driven Analytical Model The conceptual data model obtained in the previous step is used as input to infer an analytical data model. The prompt is as follows:
Consider the conceptual data model obtained in the previous prompt. Assume we want to create a data warehouse system for the analysis of that data. Adopt a constellation schema if multiple fact tables are needed, integrating the several fact tables through shared dimensions. For each star schema of the constellation, describe the fact table, the dimension tables, and the attributes of these tables. For each fact table, list the measurable indicators (metrics). 4.2.3. From the User Requirements and the Data-driven Conceptual Model to the
Requirements-driven Analytical Model Here the conceptual model guided by the data and the user requirements is used to derive an analytical model with fact and dimension tables. The prompt is as follows:
Consider user requirements specified using the iStar extension for Data Warehouses. In it, Strategic Goals (SG), Decision Goals (DG), and Information Goals (IGs) are used. SGs are related to the main objectives of the business process that are being enhanced, representing a desired change from a current situation to a future one. DGs represent decisions that use information to provide benefits for the organization, operationalizing the SGs into actions by answering the question, "How can a strategic goal be achieved?" IGs guide the information needed to achieve a DG by responding to the question, "How can decision goals be achieved in terms of information required?" IGs outline the data that must be gathered, usually through analysis. As a result, they can be described in terms of goals or in terms of the analysis process. The requirements are: «to be defined» Based on the conceptual data model obtained from the data and the user requirements mentioned above, propose an analytical data model suitable for a data warehouse system that addresses the entities of the domain. Identify one or more fact tables that meet the IGs, and define the dimension tables needed to provide the relevant contextual information. Use a constellation schema if needed. For each fact table, list the measurable indicators (metrics), and for each dimension, list the attributes required to support filtering or grouping. Clearly indicate how each IG is supported by the analytical data model.
Analytical Model 4.2.4. From the Data- and Requirements-driven Analytical Models to the Multi-driven Now the LLM must merge the two previous analytical models, maintaining the alignment between available data and requirements. The prompt is as follows:
Given the two analytical data models — one derived from the available data and the other from user requirements in the context of the entities of the domain - integrate both into a unified, multi-driven analytical model. Use a constellation schema if multiple fact tables are needed. When dimensions or facts overlap, unify them; when they diverge, include both with appropriate relationships.
Highlight where the models complement each other or where gaps in data exist. 4.2.5. From the Multi-driven Analytical Model to the Analytical Visualizations Based on the final model, the LLM should propose useful visualizations for users based on the IGs. The prompt is as follows:
Based on the multi-driven analytical model and the IGs, suggest analytical dashboards or visualizations that can support decisionmaking. For each visualization, specify: i) the SG and DG it supports; ii) the data it uses (fact and dimension tables); iii) the type of chart (e.g., time series, bar chart, heatmap); and iv) the recommended aggregation and filtering dimensions. Whenever possible, organize the visualizations by SG, showing how each IG contributes to it.
This section presents the step-by-step application of the multi-driven decision-making approach. The LLM used in the instantiation was the ChatGPT language model3. All prompts were detailed in the previous section. For reproducibility, the used prompts and obtained results are available at [16].
The first step identified a conceptual data model based on the structure of the available industrial dataset (Fig. 3). Only metadata with the tables names and their attributes was made available. The resulting model identifies key entities within the production domain, such as Customer, Order, Material, Operation, and WorkCenter, and outlines the relationships between them. For instance, it shows that customers place orders, which in turn triggers the production process. Each order uses a specific set of materials required for the production of the ordered items. Orders include a sequence of operations that are performed at a particular work center (physical or logical area of the industrial setting, such as
lithography or cutting, where a specific type of operation takes place). The model also incorporates energy consumption data, linking each work center to its associated gas and electricity usage through the entities ConsumptionGas and ConsumptionElectricity.
It is important to note that this conceptual model does not represent process sequences or execution lfow; rather, it describes the static relationships among the core business entities relevant to the domain. This conceptual structure provides a consistent and realistic representation of the production environment and serves as a solid foundation for analytical modeling in the following steps. This model (Fig. 3) was already validated by the data engineer, who disregarded a redundant relationship between orders and work centers identified by the LLM, as operations already relate these two entities.
After the conceptual data model, the LLM suggested three main fact tables: one focused on order analysis (Fact_Orders), another on operation execution (Fact_Production), and a third on energy consumption (Fact_EnergyConsumption). These fact tables are connected to dimension tables, such as Date (Dim_Date), Time (Dim_Time), Customer (Dim_Customer), Work Center (Dim_WorkCenter), and Operation (Dim_Operation). This model is discussed in subsection 5.4.
Considering the concepts of the domain and the user requirements specified in iStar (Figure 4), the LLM suggested a structured mapping between the IGs and the existing analytical data model. It should be noted that the iStar model was designed by the authors, based on requirements elicited in collaboration with the industrial partner. Each IG was linked to the relevant fact and dimension tables, conrfiming that the previously defined constellation schema could support all analytical needs derived from the SGs and DGs.
The result confirmed the need for a constellation schema, but centered on two fact tables: Fact_Production, capturing detailed records of operational execution (e.g., duration, idle time, and energy per operation); and Fact_EnergyConsumption, summarizing energy usage per time period and work center. These are linked to dimensions such as WorkCenter, Operation, Material, Date, and Time, depending on the case, enabling flexible filtering and aggregation.
It is important to highlight that: i) a traceability matrix was proposed by the LLM, identifying the metrics and dimensions required to support each IG. This ensured that the analytical model covers the full set of user requirements defined in iStar and clarified how each element of the schema contributes to addressing specific decision-making needs; ii) no fact table for orders was suggested by the LLM, which makes sense as the user requirements are focused on operational concerns, such as reducing shelf time and energy consumption, and not on more business-oriented objectives, such as increasing the number of orders over time.
Considering both analytical data models, Figure 5 depicts the integrated data model, highlighting the diferences between them when applicable.
As Figure 5 shows, most of the fact and dimension tables were suggested by both data- and requirements-driven perspectives (tables with headers represented in a white color). The requirementsdriven perspective suggested mainly additional metrics to include in the fact tables, addressing the diferent IGs and computing the indicators for answering the user’s requirements (these indicators are represented in a light-grey color). The data-driven perspective, with the tables suggested only by this perspective represented with the header in a dark-grey color, is related to information present in data that is not needed to answer this specific set of user requirements. However, these can be useful for further analysis, as mentioned previously in subsection 5.3, as the use case only included a limited set of user requirements with operational concerns. With organizational decision-making in mind, the integrated analytical model provides a comprehensive analysis of diferent business and process indicators, making available the proposal of a DW that is not devised for current informational needs but prepared to answer to future analyses.
The final step of the approach consists in translating the multi-driven analytical model into visualizations that directly support decision-making. From the dashboards suggested by the LLM (with a semiautomatic alignment process, where the LLM proposes and a data engineer validates), we selected the one associated with the DG “Optimize Work Centers” (Figure 6) of the SG "Reduce Energy Consumption," allowing the analysis of IG3, IG4, and IG5.
First, a DW system was physically instantiated in the industrial setting, following the multi-driven analytical model. Afterwards, a dashboard that includes the visualization of energy consumption by work center and time period was implemented, addressing IG4 and IG5. For IG3, a key limitation was found. Although the production type exists in the dataset, it contains only a single value - ofset. This lack of variability prevents any meaningful comparison between diferent production types, which is precisely what the goal requires. As a result, while IG3 is formally represented in the dashboard, the findings fall short. As illustrated in Figure 6, the bar showing the “ofset” production type visually highlights this limitation. Meaningful analyses would require collecting energy consumption data across multiple distinct production types. Pointing these findings is important for future data collection tasks, suggesting the definition of a data policy regarding current and future information needs.
Despite this limitation, the dashboard reveals valuable insights regarding IG4 and IG5. For example, as shown in the average consumption chart, work centers CS08 and CS13 stand out with the highest average electricity consumption. While in some cases high consumption aligns with longer operation durations, the dashboard also reveals exceptions - such as certain work centers with long duration but comparatively lower energy consumption - highlighting the importance of analyzing both dimensions jointly. Additionally, the heatmap related to IG5 reveals clear temporal patterns, with consumption peaking at 9:00am on Wednesdays and 8:00am on Thursdays. Such patterns may reflect production scheduling practices and highlight potential areas for optimization, such as load balancing or energy cost reduction strategies during peak hours. Therefore, the dashboard not only operationalizes the analytical objectives but also provides concrete evidence to guide energy-aware decision-making.
Moreover, one of the distinctive advantages of the multi-driven approach is that it not only reveals what can be done, but also what cannot. A clear example is DG3, "Prioritize Low-Energy Production Routes". While highly relevant from both operational and environmental perspectives, this goal could not be addressed analytically due to the absence of data on routing between work centers. IG6 would require detailed logs of inter-work center transitions, including operation sequences, timestamps, and energy consumption per route - information that is not currently captured in the available datasets. Nonetheless, this requirement was anticipated in the design of the analytical model, as suggested by the LLM, and the model was structured to accommodate such data in future iterations. As a result, while the data analysis is not possible yet, the model is already structured to incorporate it as soon as the necessary data becomes available. Still, the implemented dashboard shows the practical value of the multi-driven approach by transforming analytical objectives into visual tools that inform decision-making and expose data gaps that matter.
This paper proposed a multi-driven approach to industrial decision-making that combines data-driven and requirements-driven perspectives, supported by LLMs. By leveraging LLMs to analyze metadata and user requirements through prompt-based interactions, we demonstrated how it is possible to accelerate the generation of conceptual and analytical models while maintaining alignment with strategic business objectives. The approach was applied to a real-world industrial case at Colep Packaging, an environment marked by operational complexity and variability. The results showed that, although validation and refinement by a data engineer are still necessary, LLMs can anticipate a significant portion of the
Power BI Desktop 01-01-2022 31-12-2025 Work Center
CS05 CS06 CS07 CS08 CS09 CS11 CS13 CS15 CS16 CS17 CS18 EM01 LE02 LE04 LE06 LE08 LI05 LI11 LI15 LI17
IG3: Identify the Type of Production with the highest Consumption Average Electricity_Consumption Average Gas_Consumption Offset Average Electricity_Consumption Average Gas_Consumption Operation_Duration
IG4: Calculate the Average Energy Consumption by Work Center CS08 CS13 CS11 CS15 CS17 CS05 LE08 LI11 LI15 LI05 LI17 CS06 CS07 CS09 CS16 CS18 EM01 LE02 LE04 LE06 modeling work, reducing the initial manual efort and improving the alignment between available data and actual decision-maker needs. The approach proved especially valuable in revealing not only what can be analyzed with existing data but also the analytical limitations caused by information gaps.
In the context of the ongoing digital transition, the integration of LLMs into industrial modeling workflows exemplifies how advanced AI technologies can enhance data utilization and accelerate informed decision-making processes. Moreover, by enabling more eficient and targeted analyses, this approach supports green transition objectives by optimizing resource use, identifying ineficiencies, and guiding sustainable operational strategies. Thus, the proposed approach contributes to both digital innovation and environmental responsibility within industrial ecosystems.
Beyond these results, the case study revealed key learning points: LLMs accelerate model creation but require expert validation; prompt design is critical to avoid ambiguities; metadata alone can be enough preserving data confidentiality; and, the multi-driven approach both aligns data with requirements and exposes analytical gaps.
As future work, we could explore the use of alternative LLMs to assess their eficiency and efectiveness across the diferent modeling stages. Furthermore, the approach needs to be applied to diferent industrial use cases, addressing validity concerns since the current evaluation relies on a single case and a specific LLM version. Broadening the scope will strengthen the generalizability and reliability of the approach.
This work has been supported by FCT – Fundação para a Ciência e Tecnologia within the R&D Unit Project Scope UID/00319/Centro ALGORITMI (ALGORITMI/UM) and the European Union under the Next Generation EU, through a grant of the Portuguese Republic’s Recovery and Resilience Plan (RRP) Partnership Agreement, within the scope of the project PRODUTECH R3 – "Agenda Mobilizadora da Fileira das Tecnologias de Produção para a Reindustrialização". This paper uses icons made available by www.flaticon.com.
During the preparation of this work, the authors used ChatGPT and Grammarly for sentence polishing and rephrasing, and ChatGPT for supporting the proposed approach. All generated content was reviewed and edited by the authors, who take full responsibility for the final text [15] Z. Sun, Q. Wang, H. Wang, X. Zhang, J. Xu, Detection and mitigation of hallucination in large reasoning models: A mechanistic perspective (2025). doi:10.48550/arXiv.2505.12886. [16] C. Vieira, V. Sousa, P. Guimarães, D. Rodrigues, A. Vieira, M. Y. Santos, Multi-driven and llmassisted analytical modeling, 2025. URL: https://doi.org/10.5281/zenodo.15784597. doi:10.5281/ zenodo.15784597.