Intelligent Environments in Manufacturing Ecosystems: Digital Platforms, Connected Intelligence, and Innovation Performance

1. Introduction

The world is undergoing profound change. The established global order is being rapidly reconfigured as transformations unfold across industries, infrastructure, and government. Transitions are driven by digitalisation, collaborative platforms, and AI; the global reorganisation of supply chains; renewed protectionism; climate change and green agendas [1,2]. Yet the trajectory toward an emerging order remains prolonged, conflictual, and uncertain [3].

Manufacturing sectors and ecosystems are particularly exposed to these transformations. At the enterprise level, changes occur in operations, skills, work organisation, transactions and exchanges, management practices, and the dynamics of innovation. At the sector level, changes are reflected in industrial trajectories and pathways, the expansion of novel industries such as electric vehicles, renewable energy, advanced materials, and robotics, alongside the decline or reconfiguration of fossil-fuel-based, labour-intensive, and technologically mature sectors [4,5].

Three groups of technologies occupy a central position in these transformations. First, digital platforms reshape organisational forms, work arrangements, business models, and industry leadership. Second, artificial intelligence, including machine learning, generative AI, and agentic AI, extends the scope of automation, prediction, coordination, and data-driven decision making. Third, green technologies in materials, energy systems, and carbon emissions reduction reorient industrial production toward sustainability and climate-neutrality objectives [6,7].

The challenge is to understand this emerging manufacturing landscape and develop responses to the complex transitions of manufacturing, including digitalisation, AI adoption, green systems, innovation capacity, and market volatility. While manufacturing innovation has focused on Industry 4.0, smart factories, automation, and data-driven operations, comparatively less attention has been devoted to ecosystem-level collaboration and experimentation, and the environments that enable collaborative innovation across firms, institutions, and industry sectors [8,9].

1.1. Problem Statement

Within this broader context, the problem addressed in this paper concerns the transformation of manufacturing ecosystems established at the sectoral level. The key question is how innovation can be organised beyond the boundaries of individual firms, at a level where knowledge, experimentation, and problem-solving are distributed across multiple organisations within a sector. It is well established that innovation in manufacturing is increasingly generated and diffused through interdependencies among firms, research organisations, technology consultants, support institutions, and technology providers. Innovation systems theory and X-helix models have shown that innovation depends on structured interactions among heterogeneous actors, including industry, universities, government, civil society, and, more recently, environmental and societal stakeholders [10,11,12].

The dynamics between internal and external capabilities and resources in innovation vary substantially by firm size, capabilities, and organisational reach. Smaller firms, in particular, depend more heavily on external resources across their innovation supply chains, since they often lack the financial, technological, and organisational capacity to manage transformation on their own. Their adaptation to digitalisation, AI, and green transition relies more on access to external digital, institutional, and physical environments, as well as on knowledge, services, infrastructure, and expert support provided by other organisations [13,14].

Therefore, our concern is with the wider environment that shapes innovation and transforms manufacturing ecosystems. Successive waves of digitalisation, platform development, collaborative business models, and data- and AI-driven decision-making have substantially changed how innovation is produced and how it integrates different forms of intelligence and capabilities, combining human, collective, and machine intelligence that complement each other. Such intelligent environments emerge within digital-institutional-physical (DIP) spaces, where digital platforms, e-services, and AI agents support expert advice, collaborations and living labs that enable data sharing, learning, experimentation, problem-solving, and transformative innovation.

1.2. Research Hypothesis

The concern about the environment of innovation informs the research hypothesis of this work, which can be stated as follows:

“Within manufacturing ecosystems, digital platforms, e-services, and AI agents enable the recombination of human, collective, and machine capabilities distributed across experts, companies, and technological infrastructures. These interactions generate intelligent environments that improve the innovation performance of both ecosystem members and the ecosystem as a whole.”

The hypothesis is empirically verifiable and can be supported or rejected by estimating how combined human, collective, and machine capabilities affect the innovation performance of an ecosystem, with case studies providing evidence of this relationship.

Recent literature aligns with this hypothesis by framing intelligent environments as socio-technical ecosystems in which technological, organisational, and human capabilities are jointly configured, rather than as purely technological infrastructures [15,16]. Annapareddy [17] frames intelligent digital ecosystems as adaptive platform-based environments in which infrastructure, services, and digital intelligence are integrated into wider service ecosystems. Rzevski et al. [18] define smart ecosystems as systems of autonomous decision-making agents capable of allocating resources, planning, coordinating, monitoring, and controlling operations in real time, with emergent intelligence arising through conflict detection, negotiation, and consensus formation. Friston et al. [19] move further in this direction by framing ecosystems of intelligence as cyber-physical settings of natural and synthetic sense-making, in which humans remain integral participants of shared intelligence. Taken together, these contributions suggest that intelligent environments should be understood as organised settings in which human, collective, and machine capabilities are linked through platforms, institutions, and interaction mechanisms that support coordination, adaptation, and innovation across distributed actors.

To assess this hypothesis, several concepts are especially important, such as smart ecosystems, connected intelligence, and ecosystemic and transformative innovation, which provide the foundations for intelligent environments. These concepts are briefly outlined below, while a more detailed account can be found in Komninos [20].

Smart ecosystems are communities of organisations that collaborate within and across territories, supported by smart technologies, digital platforms, and artificial intelligence. Collaboration among ecosystem members drives innovation in operations, transactions, products, and services, thereby improving efficiency and sustainability [21,22,23,24]. Platform ecosystems provide the technological infrastructures and business models through which diverse actors are connected and coordinated to co-create value and innovation [25,26,27,28,29,30].

Connected intelligence denotes the coupling among human, collective, and machine intelligence and the integration of their capabilities. Human intelligence refers to the capabilities of human actors, including creativity, intuition, judgment, and reasoning under conditions of limited information [31,32,33,34,35]. Collective intelligence creates the conditions, institutions, and shared resources that enable collaboration and innovation systems to emerge [36,37,38,39]. Machine intelligence provides capabilities for data analysis, real-time and algorithmic reasoning, decision support, and prediction through generative and agentic AI [40,41,42,43,44]. Under certain conditions, connected intelligence emerges within smart ecosystems, generating transformative innovations and enabling state transitions. It is, therefore, a cornerstone in the transition and evolution of smart ecosystems toward intelligent environments and a driving force of collaborative problem-solving through data and AI.

Innovation involves substantial changes in the operating routines of an organisation, community, or territory. These changes may be incremental or radical, and they are typically associated with measurable improvements in performance. Transformative innovations entail deeper, often disruptive changes in routines and operational performance that go beyond optimisation, producing non-linear shifts in outcomes relative to prior or expected trajectories [45,46,47,48,49].

Intelligent environments emerge when smart ecosystems, connected intelligence, and transformative innovations interact and operate together. However, their emergence depends on three interrelated conditions. A contextual condition: the presence of platforms and ecosystems that enable collaboration among different types of agents, including experts, organisations, and AI agents, which allow their capabilities and resources to be combined. A procedural condition: the presence of catalysts that facilitate binary couplings among human, collective, and machine intelligence, materialised as services, tools, applications, and communication interfaces that enable interaction, coordination, and learning across agents. An evidential condition: the presence of innovation and transformation in the ecosystem, providing observable and measurable evidence of intelligence at work [20,50].

1.3. Research Methodology

The research methodology we follow to assess the hypothesis includes three interconnected stages to understand the operation and dynamics of intelligent environments in manufacturing ecosystems. The first stage develops the conceptual foundations, mathematical formalisation, and modelling framework needed to define the effect of intelligent environments on manufacturing ecosystems. The second stage is a case study that applies this formalisation to the experimental design of an intelligent environment. The third stage uses simulations to assess how interactions among human, collective, and machine capabilities affect the transition from the baseline state of an ecosystem to a more advanced state of innovation performance.

This methodology treats intelligent environments not as purely technological settings but as multidimensional socio-technical spaces. These spaces bring together digital platforms, e-services, AI agents, organisations and actors, institutional agreements, ecosystemic collaboration, living-lab experimentation, and co-design. The methodology is therefore both analytical and constructive: it aims not only to explain how innovation may emerge within manufacturing ecosystems but also to guide the design of environments that actively support such emergence.

Stage 1: Formalisation and modelling. The model formalises an intelligent environment in which interactions among human, collective, and machine capabilities establish a transformation engine that drives the transition of an ecosystem from a baseline innovation state to an advanced state. The selected model is a one-step vector autoregressive model, VAR(1). It includes the baseline and advanced innovation states of the ecosystem, as well as a transformation matrix that captures binary couplings among human intelligence (HI), collective intelligence (CI), and machine intelligence (MI). Each coupling reflects both a combination of capabilities and a direction of interaction. For instance, the HI–MI (HM) coupling reflects the combination of HI and MI capabilities, as well as the actualisation from HI to MI. These couplings drive the transformation of the initial state of a manufacturing ecosystem into an advanced state of higher innovation performance.

Stage 2: Experiment in the design of an intelligent environment.The case study begins with a multi-criteria analysis (MCA) to select the manufacturing sector that forms the foundation of the ecosystem in focus. Candidate sectors are compared against criteria related to innovation potential, digital readiness, green-transition needs, collaboration capacity, and implementation feasibility. MCA provides a transparent and systematic basis for selecting the sector in which the intelligent environment will be designed and tested. Additional criteria are then used to select a subset of companies within the sector, thereby defining the ecosystem population. A DIP socio-technical space creates the foundation for the intelligent environment within the selected ecosystem. This involves the ecosystem context and participating organisations; digital platforms, e-services, and AI agents; expert advice; and collaborative settings for experimentation and innovation. At the core of this environment is a transformation engine that combines binary couplings among human, collective, and machine intelligence to generate innovations in operations, capabilities, transactions, and products.

Stage 3: Simulation and analysis of transition dynamics. The third stage uses successive simulations with synthetic data to examine how the connected intelligence transformation engine affects the transition from the baseline to the advanced state of the ecosystem. Particular attention is given to changes in the eigenvalues and eigenvectors of the weighted transformation matrix, since these provide insight into the evolving structure, strength, and stability of interactions among the forms of intelligence. Simulations link the internal dynamics of the transformation engine and the transition from the baseline to the advanced state of the ecosystem. Starting from an initial weighted matrix, one coupling parameter is varied at a time. For each perturbation, the corresponding eigenvalues and eigenvectors are computed to assess changes in the structure and stability of interactions among human, collective, and machine intelligence. The same perturbed matrix is then used to compute the simulated advanced state. The analysis links changes in the dynamic structure of connected intelligence directly to changes in the performance variables that define the transition from the baseline to the advanced state.

Figure 1. Three-stage methodology for modelling, implementing, and analysing intelligent environments.

Figure 1. Three-stage methodology for modelling, implementing, and analysing intelligent environments.

Drawing on the problem statement, research hypothesis, and methodology outlined above, the paper is organised as follows. Section 2 outlines the mathematical formalisation used to capture the interactions among smart ecosystems, connected intelligence, and transformative innovation, conceptualising intelligent environments as outcomes of these interactions. Section 3 presents a case study. The design of an intelligent environment for a manufacturing ecosystem, its building blocks, DIP space, platforms and e-services, and operational logic. Section 4 develops the simulation experiment to examine the internal dynamics of intelligent environments. Section 5 returns to the research hypothesis and discusses what the formal model, case study, and simulation reveal about the dynamics of intelligent environments in manufacturing ecosystems. Finally, Section 6 draws conclusions by identifying the fundamental components of intelligent environments, their ecosystemic context, the connected-intelligence transformation engine, and transformative innovation as the documented outcome.

2. Foundations and Formalisation of the Transition

The formalisation proposed here describes how intelligent environments, which combine digital platforms, e-services, and agentic AI with expert advice and inter-company collaboration, can strengthen the innovation performance and drive the transition of a manufacturing ecosystem. It provides a mathematical representation of how an intelligent environment grounded in connected intelligence moves an ecosystem from a baseline state to a more advanced state, as documented by changes in routines and performance metrics.

The model includes (a) the baseline state of a manufacturing ecosystem, (b) a transformation engine based on connected intelligence, (c) the advanced state of the ecosystem, (d) key performance indicators describing and capturing changes in the ecosystem, and (e) the parameters that define and explain the transition from the baseline to the advanced state. Linear algebra and dynamic systems provide a suitable formalism for expressing this transition, showing how, in the context of smart ecosystems, human intelligence (HI), collective intelligence (CI), and machine intelligence (MI) interact to generate transformative innovation (I). In this way, the model captures the operation of an intelligent environment as a system of connected intelligence that generates innovation. The parameters of the transformation engine describe interactions among autonomous forms of intelligence, while innovation remains an emergent phenomenon that may involve nonlinear, socially constructed, and path-dependent processes.

2.1. Baseline State

The baseline state of the ecosystem is represented by a vector of performance variables that capture innovation inputs, outcomes, and impacts, measured at the level of ecosystem members, and aggregated as ecosystem-level averages.

Yt-1 = [Xt-1, Ot-1, Pt-1]^T

(1)

where:

• Yt-1 is the baseline-state vector of performance variables;
• Xt-1 is the average innovation input variable;
• Ot-1 is the average innovation outcome variable;
• Pt-1 is the average innovation impact variable.

These variables may be selected from established innovation scoreboards. Input variables may refer to R&D expenditure, digitalisation, and personnel education or skills. Outcome variables may include patent applications, sales from new or improved products, or productivity gains. Impact variables may include growth, emissions reduction, or safety improvements, depending on the sector in focus. The baseline-state vector should include all three types of variables. In a simplified specification, it may include one variable for each dimension: one input indicator, one outcome indicator, and one impact indicator.

2.2. Advanced State

In the advanced state of the ecosystem, the initial state descriptors have evolved. Innovation is therefore represented not as a static condition but as the rate of change in the relevant performance variables that capture inputs, outcomes, and impacts.

Yt = [Xt, Ot, Pt]^T

(1)

The innovation, in discrete time, will be

$${\mathrm{I}}_{\mathrm{t}}={\displaystyle \frac{{\mathrm{Y}}_{\mathrm{t}}-{\mathrm{Y}}_{\mathrm{t}-1}}{{\mathrm{Y}}_{\mathrm{t}-1}}}$$

(3)

while in continuous time, it will be

$$\mathrm{I}\left(\mathrm{t}\right)={\displaystyle \frac{\mathrm{d}\mathrm{Y}\left(\mathrm{t}\right)}{\mathrm{d}\mathrm{t}}}$$

(4)

and

$$\mathrm{A}\left(\mathrm{t}\right)={\displaystyle \frac{{\mathrm{d}}^2\mathrm{Y}\left(\mathrm{t}\right)}{\mathrm{d}{\mathrm{t}}^2}}$$

(5)

This second derivative captures acceleration and distinguishes a transformation accelerating, slowing, or at a constant rate:

A(t) > 0: the transformation is accelerating

A(t) < 0: the transformation is slowing down

A(t) = 0: innovation proceeds at a constant rate

where:

$Y_{t-1}$ is the performance vector at the baseline state

$Y_t$ is the performance vector at the advanced state

$I_t$ is the vector of relative changes in innovation performance between the two states, and

$A\left(t\right)$ is the innovation acceleration in continuous time.

In the present specification, innovation is measured in discrete time as a change in one time lag between the baseline and the advanced state. Innovation acceleration, however, would require additional observation periods and is therefore not included in the simplified one-lag model.

This formulation allows innovation to be interpreted as the observed transformation of an ecosystem’’s performance profile rather than as a single scalar outcome. Each component of $I_t$captures the relative change in one performance dimension, such as input capacity, innovation outcome, or sectoral impact.

2.3. Transformation Engine Representing Connected Intelligence Couplings

The transformation engine is the core mechanism of the intelligent environment. It consists of weighted binary couplings among human, collective, and machine intelligence that drive the ecosystem’’s transition from its baseline state to an advanced state of innovation performance. It involves the actualisation of the intelligent environment within the DIP space of the ecosystem, enabling rich interactions among human, collective, and machine intelligence. These forms of intelligence do not merge into a single entity; rather, they remain autonomous and interact through socio-technical interfaces such as digital platforms, e-services, expert advisory systems, living labs, collaborative platforms, data spaces, and agentic AI tools. The advanced state is therefore not simply the outcome of technological adoption, but the product of connected intelligence operating across human, collective, and machine capabilities.

The transformation engine represents the dynamic interactions among human, collective, and machine intelligence, together with their respective capabilities. These capabilities are captured by KPIs and organised into vectors, while their interactions are represented through a coupling matrix that governs the rates of change in innovation inputs, outcomes, and impacts. In this way, the transformation engine models how different forms of intelligence interact over time to shape sectoral innovation dynamics.

The model does not represent human (HI), collective (CI), and machine (MI) intelligence as separate capability vectors. Instead, it focuses on binary couplings among them, treating innovation as an outcome of their interactions rather than of each form of intelligence in isolation. These couplings are operationalised through KPIs that measure the effects of interactions between human, collective, and machine intelligence.

The transformation engine is represented by a weighted matrix of binary couplings of connected intelligence that link HI, CI, and MI and capture interactions among these autonomous forms of intelligence. The weights in the matrix express the intensity and effects of interactions at the level of each observation unit, here the company. They may be estimated through OLS derived from survey responses and observable metrics, such as participation in collaborative platforms, use of AI applications, expert advisory support, data-sharing practices, living-lab engagement, and related indicators (see the metrics proposed in Section 3).

Each matrix coefficient combines an estimated weight and a KPI-based interaction metric. For instance, in the coupling between human and collective intelligence, the coefficient xHC measures the KPI-based effect of the interaction, and wHC quantifies its contribution to innovation performance. The KPI metric is defined empirically, while the estimated weight is obtained through regression.

$$\left[\begin{array}{ccc}0& \mathrm{w}\mathrm{H}\mathrm{C}\mathrm{x}\mathrm{H}\mathrm{C}& \mathrm{w}\mathrm{H}\mathrm{M}\mathrm{x}\mathrm{H}\mathrm{M}\\ \mathrm{w}\mathrm{C}\mathrm{H}\mathrm{x}\mathrm{C}\mathrm{H}& 0& \mathrm{w}\mathrm{C}\mathrm{M}\mathrm{x}\mathrm{C}\mathrm{M}\\ \mathrm{w}\mathrm{M}\mathrm{H}\mathrm{x}\mathrm{M}\mathrm{H}& \mathrm{w}\mathrm{M}\mathrm{C}\mathrm{x}\mathrm{M}\mathrm{C}& 0\end{array}\right]$$

(6)

where

• W is the weighted matrix of binary couplings of connected intelligence
• wHC, wHM, wCH, wCM, wMH, wMC are estimated weights
• xHC, xHM, xCH, xCM, xMH, xMC are measured KPI interaction effects metrics.

2.4. The Transition Model

To represent the transition from the baseline to the advanced state of the ecosystem, through the operation of the transformation engine, we use a vector-based socio-technical modelling framework. The transformation from the initial state to an advanced state can be represented as a first-order vector transition. The baseline state is represented by a performance vector Yₜ₋₁, while the post-transformation state is represented by the vector Yₜ. If the main performance change is expected to occur over one lag between these two states, a VAR(1)-type transition model can be used to estimate how baseline conditions and interactions among variables influence the transformed state.

A first-order Vector Autoregressive model, VAR(1), for three variables per vector is as follows:

Yₜ = b₀ + W₁Yₜ₋₁ + eₜ

(7)

In matrix form

Yt = b0 + WYt−1 + et

(8)

or

$$\left[\begin{array}{c}{Y}_{1t}\\ Y_{2t}\\ Y_{3t}\end{array}\right]=\left[\begin{array}{c}{b}_{01}\\ b_{02}\\ b_{03}\end{array}\right]+\left[\begin{array}{ccc}0& w_{12}& w_{13}\\ w_{21}& 0& w_{23}\\ w_{31}& w_{32}& 0\end{array}\right]\left[\begin{array}{c}{Y}_{1,t-1}\\ Y_{2,t-1}\\ Y_{3,t-1}\end{array}\right]+\left[\begin{array}{c}{e}_{1t}\\ e_{2t}\\ e_{3t}\end{array}\right]$$

(9)

The matrix equation expands into the following three equations {10}

Y₁ₜ = b₀₁ + w₁₁ · Y₁,ₜ₋₁ + w₁₂ · Y₂,ₜ₋₁ + w₁₃ · Y₃,ₜ₋₁ + e₁ₜ

Y₂ₜ = b₀₂ + w₂₁ · Y₁,ₜ₋₁ + w₂₂ · Y₂,ₜ₋₁ + w₂₃ · Y₃,ₜ₋₁ + e₂ₜ

Y₃ₜ = b₀₃ + w₃₁ · Y₁,ₜ₋₁ + w₃₂ · Y₂,ₜ₋₁ + w₃₃ · Y₃,ₜ₋₁ + e₃ₜ

where:

• Y₁ₜ, Y₂ₜ, Y₃ₜ, the three endogenous variables observed at time t;
• Y₁,ₜ₋₁, Y₂,ₜ₋₁, Y₃,ₜ₋₁, their one-period lagged values;
• W, the coefficient matrix, each lagged variable affects each current variable;
• b₀₁, b₀₂, b₀₃ the intercept terms for each equation;
• wᵢⱼ (i,j = 1,2,3) the autoregressive coefficients corresponding to HC, HM, CH, CM, MH, MC binary couplings
• e₁ₜ, e₂ₜ, e₃ₜ, the noise error terms at time t.

In the proposed restricted VAR(1) model with a three-variable vector, the specification includes three intercepts and six autoregressive coefficients, corresponding to the six directed binary couplings among human, collective, and machine intelligence. If the three error variances are also included, the model involves 12 parameters in total. The number of available observations should be sufficient to estimate the proposed model with its intercepts, autoregressive coefficients, and error variances.

The values of the baseline-state and advanced-state descriptors, together with the interaction indicators (xHC, xHM, xCH, xCM, xMH, xMC) are measured across the firms participating in the manufacturing ecosystem. These observations are used to estimate the corresponding weighted interaction coefficients (wHC, wHM, wCH, wCM, wMH, wMC) of the transformation engine. The estimated coefficients quantify the strength of the directed interactions among the dimensions of connected intelligence and reveal how these interactions influence the advanced-state innovation inputs, outcomes, and impacts.

2.5. System Design and Selection of Variables

The model described here requires substantial design effort, as it does not formalise an existing environment. Instead, it seeks to represent a dynamic configuration comprising the manufacturing sector, the services that actualise the ecosystem, the binary couplings, the creation of the transformation engine, and the resulting outcomes and impacts of transition.

Treating the intelligent environment as a socio-technical system composed of a manufacturing ecosystem, its baseline state, the transformation engine, and the advanced state requires specifying the structure and properties of these subsystems. It is necessary to define how the ecosystem is created, its key transformation processes, the forms of interaction among actors, technologies, and institutions, including the digital platforms, data-driven activities, collaboration services, knowledge-sharing spaces, and inter-firm exchanges.

All these parameters affect the selection of variables used to represent the ecosystem’s baseline state. The selection must account for the trade-off between the dimensionality of the baseline-state vector and the number of observations required for robust parameter estimation. The available observations constitute the panel data used to estimate the model parameters, with each observation providing values for all selected variables. The central trade-off, therefore, is between the richness of the baseline-state description and the number of observation points required for reliable estimation.

With a sufficient number of observation units (for example, 100 organisations), a VAR(1) model containing approximately 15 to 20 parameters is feasible to estimate. This is because the number of observations remains substantially larger than the number of parameters, providing an acceptable basis for estimation and leaving sufficient degrees of freedom for statistical analysis. In regression-based estimation, each additional parameter reduces the residual degrees of freedom; therefore, maintaining a favourable ratio between observations and estimated parameters is important for obtaining reliable and interpretable coefficient estimates [51,52,53].

The design of the transformation engine specifies how human, collective, and machine capabilities are connected and activated. Their corresponding vectors represent activation of capabilities already present within the ecosystem, as well as those created through targeted interventions, such as expert advisory systems, living labs, collaborative platforms, AI assistants, and agents for advice, benchmarking, learning, and collaboration.

The advanced state is defined by the rates and, where relevant, the acceleration of change in the main variables used to characterise the baseline state. It therefore expresses not a different descriptive framework, but the transformation of the initial state through changes in routines, performance, and innovation dynamics.

All variables and processes to be defined are sector-specific and depend on the maturity of the industry under consideration, as well as on the propensity for collaboration, digitalisation, innovation, and transition. However, additional services, tools, and support mechanisms must also be introduced to facilitate and accelerate this transition. The manufacturing ecosystem thus provides the initial conditions for innovation through the interaction of social and technological entities, including knowledge actors, organisations, and digital agents. Creating an intelligent environment requires the development of additional infrastructures, platforms and e-services, new institutional arrangements, spaces for experimentation, and AI-enabled mechanisms. The objective is to create, through design, a coherent socio-technical system capable of accelerating the transition of an ecosystem from the baseline to the advanced state of innovation performance.

3. SmartGreenEcos: An Intelligent Environment for Manufacturing Ecosystems

The empirical basis for testing the formal model described is provided by the SmartGreenEcos experiment, which focuses on the smart and green transition of the dairy industry across four Mediterranean countries. SmartGreenEcos is a NEXT MED Interreg project that develops digital platforms, e-services, targeted agentic AI, mobilises specialised consultants, and uses living-lab approaches to support collaboration and innovation within an industry ecosystem of 100 companies from Greece, Cyprus, Tunisia, and Jordan. It provides an experiment in which the proposed methodology for designing, modelling, and optimising intelligent environments for innovation is illustrated and assessed.

3.1. The SmartGreenEcos Innovation Model

SmartGreenEcos is building a mission-oriented manufacturing ecosystem that supports the digital and green transition of the dairy sector through digital platforms, e-services, artificial intelligence, expert advice, and inter-company collaboration. The operations, skills, transactions, and products of participating companies are treated as potential fields of innovation and change. SmartGreenEcos aligns with three major pillars of current EU policy. It contributes directly to the new European industrial policy, which emphasises resilient and innovative industrial ecosystems as drivers of competitiveness, collaboration, and cross-border value creation. It supports the digital transition by accelerating the uptake of advanced digital services, data platforms, and AI solutions that help companies modernise processes, share knowledge, and integrate into the European data economy. It advances the green transition and carbon-neutrality agenda by promoting energy efficiency, circular-economy practices, and sectoral roadmaps to reduce greenhouse gas emissions.

SmartGreenEcos creates a cross-border ecosystem of dairy companies, research organisations, technology providers, and experts. The ecosystem brings together 100 manufacturing companies, and additional consulting firms and research laboratories from four Mediterranean countries. SmartGreenEcos deploys an innovation pathway comprising four main stages: (1) creation of the ecosystem; (2) development of sector-specific digital platforms, e-services, and AI agents; (3) provision of expert advice and innovation support; and (4) engagement of companies in collaboration and use of the platforms and services, initially with expert support and subsequently on their own. This pathway operationalises the model outlined in Section 2, in which the ecosystem’s baseline state evolves into an advanced state through the operation of a transformation engine. This engine comprises targeted platforms, e-services, AI agents, and binary couplings among human, collective, and machine capabilities (Figure 2).

As a model of ecosystemic innovation, SmartGreenEcos takes a different path from startup-driven innovation, which typically centres on small, high-risk ventures developing niche solutions. Instead, it seeks to mobilise the sectoral base of established dairy companies to innovate collectively. The focus is both on innovation at the individual company level and on the wider transformation of the dairy sector. Rather than promoting isolated experiments, SmartGreenEcos creates shared infrastructures and collective capabilities through digital platforms, e-services, AI agents, expert support, and inter-company collaboration. In this way, innovation becomes a sector-wide transformation process rather than a startup-driven phenomenon. By diffusing smart and green technologies across the ecosystem, the model also addresses size inequalities in the industry, ensuring that smaller companies with limited resources can access advanced technological capabilities, knowledge, and innovation resources.

SmartGreenEcos uses collaborative mechanisms, digital platforms, e-services, and AI applications to support innovation across company operations, skills, transactions, and products. Participating companies gain access to tools to assess ESG performance and digital maturity, optimise production processes, reduce energy and raw material costs, strengthen AI capabilities, and improve procurement through better supplier selection and input-cost management. The ecosystem also supports B2B and B2C sales channels, supply-chain innovation, collaborative new product development, and knowledge exchange. By joining SmartGreenEcos, companies can benefit from shared technological capabilities, AI-enabled learning solutions, expert support, and networking opportunities with companies, research organisations, and technology providers across the Mediterranean.

3.2. Selecting a Manufacturing Sector as Foundation of the Ecosystem

SmartGreenEcos develops an innovation model that is applicable across all manufacturing sectors. However, as a real-life experiment, it must be implemented within a specific sectoral focus to ensure the relevance of applied practices across companies. The first methodological challenge concerns the appropriate level of sectoral granularity: whether the design of the system will be conducted at the level of an industry division, such as NACE Divisions 10–32; an industry group, such as NACE 10.1–32.9; or an industry class, such as NACE 10.11–32.99. Each option presents advantages and limitations, including the availability of a sufficient number of firms, common sectoral challenges and production processes, and the likelihood of attracting sufficient company interest, given that participation in setting the ecosystem is voluntary. In practice, the choice is narrowed down to the industry-division and industry-group levels, since sector size and data availability at the industry-class level are often restricted to avoid the identification of individual firms.

Based on statistical data from Greece, Cyprus, Tunisia, and Jordan, the food industry was selected as the broader field for the experiment. Within this division, the next step was to identify one industry group among the nine groups that make up food manufacturing. The first criterion considered was the size of each industry group (NACE 10.1, 10.2, 10.4, 10.6, 10.8, and 10.9) across the four countries. Groups with a relatively small number of firms in any participating country were excluded because this would constrain the formation of the 100-member ecosystem.

A multicriteria analysis (MCA) was then applied to the three candidate industry groups: processing and preserving of fruit and vegetables (NACE 10.3), manufacture of dairy products (NACE 10.5), and manufacture of bakery and farinaceous products (NACE 10.7). The MCA criteria included: (1) sector size, (2) dynamism, measured as change over the last five years, (3) turnover, (4) exports, (5) supply-chain breadth, (6) green transition potential, (7) digital transition potential, and (8) ecosystem-setting propensity. The MCA, normalised on a 0–1 scale at the criterion level, produced a clear and robust ranking of the three candidate industry groups across Cyprus, Greece, and Jordan, with scores of 16.66 for fruit and vegetables, 18.00 for dairy products, and 7.60 for bakery and farinaceous products.

The dairy industry was therefore selected for the SmartGreenEcos experiment. It scored consistently highly across economic-relevance criteria, including turnover and exports, supply-chain depth and integration, green and digital transition potential, and ecosystem maturity, as reflected in the presence of clusters, cooperatives, and shared platforms. The next step, implemented in each participating country, was to conduct a dissemination campaign targeting companies in the selected industry group. The campaign presented the SmartGreenEcos model, invited firms to join the ecosystem, explained the benefits of participation, and collected declarations of interest. This process led to the selection of 100 dairy companies to participate in the ecosystemic innovation experiment.

3.3. The Transformation Engine: Platforms and E-Services Actualising Binary Couplings

The transformation engine is the core component of the experiment, alongside the ecosystem itself, and the variables used to describe the ecosystem’s baseline and advanced states. Formally, the transformation engine is represented by a six-element interaction matrix that encodes weighted binary couplings among different forms of intelligence: HC, HM, CH, CM, MH, and MC. These binary couplings are directed. For example, HM differs from MH because each coupling indicates a different initiating intelligence.

In reality, the engine driving the transition from the baseline to the advanced state is far more complex. Each binary coupling can be materialised through multiple services, mechanisms, or interventions, operating at different levels of effectiveness. If the six binary couplings are considered across ten normalised effectiveness levels, from 0.1 to 1.0, this produces 60 potential coupling-effectiveness elements, with the number of possible configurations exceeding 50 million. This illustrates the very large configuration space underlying even a simplified transformation engine. Moreover, the engine’s configuration also depends on the variables used to describe the baseline state, since the impact of each binary coupling is conditioned by the corresponding baseline-state descriptors. The transformation engine is therefore defined not only by its internal coupling structure, but also by its relationship to the specific characteristics of the ecosystem being transformed.

The SmartGreenEcos transformation engine comprises four digital platforms, eight e-services, and multiple AI agents, focused on operations, skills, transactions, and product development (Figure 3). These platforms and e-services are activated by experts and companies within the ecosystem. Experts contribute human capabilities; companies contribute collective capabilities and resources; while digital platforms, e-services, and AI agents introduce machine capabilities into the ecosystem. Within each platform, the full set of binary couplings among human, collective, and machine intelligence occurs: HC, HM, CH, CM, MH, and MC. Human, collective, and machine actors work both as initiators of interactions and recipients of their effects. The digital platform, e-services, and AI agents serve as operational mechanisms that link, coordinate, and mobilise distributed capabilities for ecosystem transformation. Given the ecosystem’’s distributed nature across four countries, all services are implemented digitally as e-services.

The “Benchmarking and AI-based Optimisation” platform focuses on operations and includes two e-services that address Environmental, Social, and Governance (ESG) performance and Digital Maturity (DM) performance. Both e-services are structured around a three-tier architecture. The first tier, the data entry section, collects structured ESG and DM information from participating companies, including relevant environmental, social, governance, digital, and operational indicators. The second tier, data analysis and maturity-level identification, transforms these inputs into ESG and DM scores, thematic sub-scores, benchmarking results, and maturity-level classifications. This analysis enables each company to understand its position in relation to ecosystem averages, sectoral benchmarks, and relevant sustainability or digitalisation standards. The third tier, the improvement action plan, converts the ESG or DM diagnosis into a customised roadmap for improvement. Agentic AI applications support this process by identifying ESG and DM gaps, prioritising improvement actions, suggesting certification or compliance pathways, and estimating expected environmental, social, and business impacts on different operations. As a final outcome, each e-service provides the company with a customised ESG or DM operations improvement action plan.

The “Capacity Building” platform focuses on strengthening capabilities, human capabilities, company capabilities, and, above all, AI capabilities within the firm. It includes a series of learning modules that facilitate the use of AI agents for the company’s digital and green transitions. The agents are based on the nexos.ai platform, which supports user-friendly agent design and data analysis. Their design relies on context engineering, using prompts informed by a selected library of publications on dairy industry operations. This ensures that the agents are grounded in sector-specific knowledge and aligned with the operational, digital, and green transition needs of dairy companies. Agents are specialised on demand forecasting, estimating upcoming demand at category and Stock Keeping Unit level; production planning, translating forecasted and confirmed demand into feasible production schedules; inventory and shelf-life management, monitoring stock exposure across finished goods, raw materials, packaging, and semi-finished products, with particular attention to ageing, remaining shelf life, and operational usability; procurement, supporting purchasing decisions for raw milk, ingredients, cultures, packaging materials, and other production inputs; predictive maintenance, identifying equipment at risk of failure before breakdowns occur; energy and resource monitoring, analysing energy use, water consumption, emissions indicators, and resource efficiency by product line, utility area, or facility; costing and margin analysis, calculating profitability at product and customer or channel level; and customer complaint and service management, organising and analysing complaints related to product quality, packaging, delivery conditions, temperature abuse, and expiry issues.

The “Transactions and Supply Chain” platform comprises two e-services: Digital Procurement and B2B Transactions. The Digital Procurement e-service provides a unified environment for companies to organise purchasing agreements, either individually or through pooled/common procurement, within consortium workspaces. It addresses fragmented purchasing, weak negotiating power, inconsistent approval procedures, limited supplier visibility, and inefficient Request for Quotation / Request for Proposal (RFQ/RFP) cycles. Its main functions cover the procurement workflow from requisition, approval, RFQ/RFP preparation, evaluation, and award to optional invoice matching or export, without including payment execution. The e-service supports buyers and, at the ecosystem level, enables procurement optimisation, collective purchasing, supplier assessment, and more transparent sourcing decisions. The B2B Transactions e-service operates as a complementary service that enables companies to sell products cross-border to commercial buyers, such as agents, importers, distributors, and resellers, through a controlled B2B marketplace. It addresses fragmented export sales, weak buyer qualification, limited control over catalogue access, poor order-to-cash visibility, and manual dispute handling. Its main functions include product catalogue publishing, buyer onboarding and verification, cart and checkout processes, seller order management, issue and return handling, and payment or settlement tracking. The e-service supports sellers, commercial buyers, and payment service integrations.

The “New Product Development” platform comprises two e-services focused on collaborative product development. The Co-Lab Space e-service is a Living Lab that provides a secure innovation environment where dairy companies can publish innovation challenges and collaborate with researchers, technology providers, and experts to solve them. It is a product innovation space on shelf-life extension, bio-based packaging, salt reduction, water-use reduction, whey valorisation, and functional dairy product development. Its main functions include challenge publication, digital NDA signing, secure Virtual Collab Rooms, document sharing, discussion channels, proposal evaluation, and decision board validation. The e-service also supports structured prototype development by enabling innovation teams to document formulations, technical specifications, process parameters, materials, trial results, and successive prototype iterations. The Eco-Simulator e-service provides a simplified Life Cycle Assessment that enables companies to evaluate the environmental impact of new products, materials, processes, or logistics scenarios before making physical investments. The service uses input data such as energy consumption, water use, packaging materials, transport distances, product bills of materials, and technical sheets to calculate carbon and water footprints, as well as comparative environmental impacts. It is a simulation space for “what-if” scenario assessment, environmental dashboards, and sustainability reports that can support financing, certification, or export requirements. Typical production-oriented applications include comparing alternative production recipes, assessing changes in process parameters, evaluating equipment upgrades, testing energy- and water-saving measures in processing lines, analysing waste and by-product valorisation options, and estimating the environmental benefits of scaling up a new product variant.

3.4. Indicators to Capture and Assess the Transition

Indicators are used to define an analytical set of variables to capture changes in firms’ innovation performance and in the ecosystem as a whole. They combine input, outcome, and impact dimensions, allowing assessment of how resources and capabilities mobilised by the transformation engine, through digital platforms, e-services, and AI agents, affect a company’s performance and the ecosystem’s digital and green transition.

The metrics used to describe the baseline and advanced states are drawn from established innovation scoreboards, which provide a wide range of indicators for assessing the performance of a given setting, whether a sector, region, or country. These indicators typically distinguish between inputs, outcomes or results, and broader economic or environmental impacts. Input indicators, such as personnel with digital skills (% of total employment), personnel with tertiary education (% of total employment), researchers and engineers (% of total employment), R&D expenditure (% of turnover), capture firms’ internal capacity to develop and adopt innovations. Outcome indicators, such as patent applications (per 100 employees), e-commerce sales (% of total sales) and sales from new or improved products (% of total sales), reflect digital market growth and direct innovation results, respectively. Impact indicators, such as labour productivity (value added per employee), recycled material input rate (% of total material inputs), or GHG emissions reduction (tCO2 equiv. avoided), reflect progress towards economic performance and greener production practices. Together, these variables provide a concise representation of the baseline and advanced states. They enable assessment of whether the SmartGreenEcos transformation engine’s operation contributes to measurable improvements in innovation, productivity, growth and environmental performance.

Twenty-four metrics describe effects from the transformation engine, defined by binary couplings among human, collective, and machine intelligence (Table 1). They capture the extent to which binary couplings affect company operations, capabilities, transactions, and product development. The coupling-effect KPIs are measured as counts of improvements attributable to each directional interaction between the source and target forms of intelligence. Then, counts are normalised to the interval [0, 1] and used as proxies for the effect of binary couplings within the transformation engine on the advanced state of the company and the ecosystem. Normalisation is performed separately for each column of interaction KPIs, across operations, capabilities, transactions, and product development, to smooth differences in scale across these domains of change and ensure that the resulting values remain comparable across dimensions before being introduced into the coupling matrix.

These interaction-effect metrics, together with the baseline and advanced-state measurements of the dairy ecosystem, provide the dataset needed to estimate the weighted matrix of the transformation engine, which is specific to this manufacturing ecosystem.

The KPIs are based on measurements collected from firms participating in ecosystems where platforms, e-services, and AI agents enable connected intelligence and support improvements in innovation performance across operations, capabilities, transactions, and products.

The weights are estimated empirically using OLS or similar methods. The estimated weights, together with the performance KPIs, define the parameters of the transformation engine. These parameters can then be used to improve the transformation engine by identifying which directed binary couplings have the strongest impact on innovation performance. This makes it possible to prioritise and strengthen the platforms, e-services, expert support mechanisms, and AI-enabled tools associated with the most influential couplings. Gradient descent can also be used to estimate the direction in which specific weights should be adjusted to improve a defined innovation-performance objective.

In the SmartGreenEcos experiment, the intelligent environment includes the baseline and advanced states of the ecosystem, platforms, e-services, and AI agents that drive the transition, as well as the weighted matrix that quantifies the impact of each directed binary coupling among human, collective, and machine intelligence. The baseline state describes the initial innovation performance of the dairy ecosystem, while the advanced state captures the expected improvements in operations, skills, transactions, and products. Between these two states, the transformation engine operates through digital platforms, expert support, inter-company collaboration, and AI-enabled tools, which materialise couplings among different forms of intelligence. The weighted matrix provides the analytical representation of this process, showing how each directed interaction contributes to the transition from the baseline to the advanced state.

4. The Dynamics of Intelligent Environments

Simulation can be used to examine the engine’s internal dynamics and its impact on the transition from the ecosystem’s baseline state to its advanced state. Starting from an initial weighted matrix, simulations can perturb the coupling coefficients and assess how the modified matrix affects the transition equation (Yt = b0 + W Yt-1 + et), which estimates the ecosystem’s advanced-state performance. In this way, simulations link changes in the coefficients and structure of connected intelligence to changes in the performance variables that define the advanced state. They therefore provide insight into the operation of the intelligent environment as a whole, from the baseline state, the directed binary couplings effects, the transformation engine weights, and the resulting innovation performance.

4.1. Simulations with Synthetic Data: Creation of the Dataset

The simulation begins from an initial matrix configuration and proceeds by gradually varying each matrix entry while holding the remaining entries constant. The six entries of the weighted coupling matrix (HC, HM, CH, CM, MH, MC) can be treated as bounded parameters of the transformation engine. At each step of the simulation, the eigenvalues and eigenvectors of the matrix are computed, together with the resulting advanced-state vector (Yt). The set of simulated perturbations thus generated provides the basis for a sensitivity analysis of how different coupling configurations affect both the internal dynamics of the transformation engine and the transition from baseline to advanced state.

The analytical dataset is generated from a (3 x 3) transformation matrix with six free coupling values (see (6) or (9)). Values of each of these six coefficients are varied in increments of 0.1, yielding 60 perturbation cases in total. For each perturbation case, the simulation produces the resulting state vector Yt = (y1t, y2t, y3t), the three eigenvalues (λ1, λ2, λ3), and the three eigenvectors (v1, v2, v3). The eigenvectors describe the structural modes of organisation of each perturbed matrix, while the eigenvalues are used both in deriving these vectors and in understanding dominance properties.

Since the six matrix parameters are varied systematically in fixed increments, the analytical focus is placed on the system’s behaviour across the resulting set of perturbation cases rather than on any single initial configuration. The simulation, therefore, constitutes an organised sensitivity analysis of the transformation engine and its impact on the advanced-state vector. In methodological terms, the analysis is conducted in both structure-space and parameter-space, since it focuses on the eigenvalues, eigenvectors, and resulting advanced-state vectors arising from matrix perturbations. It may provide three types of insight: first, matrix dynamics, through changes in eigenvalues and eigenvectors; second, transition effects, through the simulated change from Yt-1 to Yt; and third, sensitivity patterns, through repeated variation of each coupling value across the simulation runs.

To create the perturbed dataset, the nexos.ai platform was used to design an agent with relevant skills to generate a synthetic dataset from the VAR(1)-type model described in section 2.

Box 1. Agent-based synthetic dataset generation.

4.2. Simulation Analysis: Matrix Coefficients, Eigenvalues, and Eigenvectors

The synthetic dataset was generated using an AI-assisted workflow. The Excel file contains 60 observations and 17 columns: case ID, parameter perturbed, Yt1, Yt2, Yt3, lambda_1, lambda_2, lambda_3, and the nine eigenvector components. A random manual check of the eigenvalues and eigenvectors against the corresponding perturbed matrices confirmed the accuracy of the generated dataset.

In this simulation, the advanced-state descriptors are not defined as increments over the baseline state. Rather, they are estimated as the combined effect of the transformation engine, the intercept term, and random disturbances. Values of Yt below Yt-1 do not indicate errors; they highlight conditions in which the combined effect of the transformation engine remains insufficient to exceed the baseline state. Conversely, values of Yt above Yt-1 indicate that the transformation engine has crossed the threshold required to generate growth.

The simulated dataset (Table 2) reveals that changes in the advanced-state descriptors (Y1t, Y2t, Y3) follow the row structure of the weighted matrix. The components of Yt do not all change simultaneously under each perturbation. Instead, perturbing a coefficient in a given row of the matrix affects only the corresponding component of Yt. For example, perturbations of w12 and w13 affect Y1t, while Y2t and Y3t remain unchanged. This occurs because Y1,t-1 = Y2,t-1 = Y3,t-1. So the effect of each row of W on Yt depends on the row sum of the perturbed matrix.

The six binary coupling coefficients affect the advanced-state descriptors in structurally equivalent ways. Because Y1,t-1 = Y2,t-1 = Y3,t-1, perturbing either of the two non-zero coefficients in the same row of $W$produces the same effect on the corresponding component of Yt. Consequently, w12 and w13 have equivalent effects on Y1t, w21 and w23 have equivalent effects on Y2t, and w31 and w32 have equivalent effects on Y3t. In this dataset, the transformation of the advanced-state vector (Yt) is therefore determined primarily by the magnitude of the perturbed coefficient and its row position in the weighted matrix.

The eigenvalues are repeated across the six perturbation blocks. For each perturbation magnitude, the same eigenvalue triplet appears regardless of whether the perturbed coefficient is w12, w13, w21, w23, w31, or w32. This indicates that, under this experimental design, the perturbed matrices have equivalent spectral properties. The eigenvalues mainly reflect the magnitude of the perturbation rather than the specific structural location of the perturbed coefficient.

A main conclusion of the simulation is that changes in the advanced-state descriptors $Y_t$and in the eigenvalues (λ1, λ2, λ3) depend primarily on the magnitude of the transformation-matrix coefficients rather than on the specific structural configuration of the matrix. The six binary couplings behave equivalently with respect to $Y_t$ and the eigenvalues.

Analysis of the eigenvectors captures changes in the transformation engine’s internal structure, providing additional information about its configuration. Unlike the eigenvalues, the eigenvector components vary according to the position of the perturbed coefficient. While the eigenvalues capture repeated dynamic magnitudes across the perturbation blocks, the eigenvectors capture how the relative contribution of the three components changes under different coupling configurations.

A correlation analysis between the perturbed matrix coefficients and the resulting eigenvector components shows how each binary coupling reconfigures the relative contributions of the three dimensions of the transformation engine (Table 3). The strongest and most systematic correlations appear in the first eigenvector. For each perturbation block, the component corresponding to the perturbed row tends to increase strongly, while the other two components tend to decrease. This confirms that the first eigenvector captures the dominant directional effect of the perturbation on the system structure.

The second and third eigenvectors show more differentiated patterns. Their correlations are less uniform across perturbation blocks, indicating that they capture secondary structural adjustments rather than the main direction of change. These patterns are useful because they reveal how perturbations redistribute the relative importance of the three dimensions within the transformation engine, even when the eigenvalues remain unchanged across blocks. In addition, lambda-1 (λ1) is consistently the largest eigenvalue across all perturbation cases. Together, these observations indicate that the first eigenvector (v1) can be considered as the dominant eigenvector of the transformation matrix.

This pattern adds an important structural layer to the analysis. It shows that the location of the perturbed coefficient affects the internal configuration of the transformation engine, even when the eigenvalues remain unchanged. The strong correlations between the perturbed matrix coefficients and the components of v1 suggest that changes in the six binary couplings primarily affect the dominant mode of transformation of the system. In practical terms, the dominant eigenvector captures the principal configuration through which human intelligence (HI), collective intelligence (CI), and machine intelligence (MI) combine to produce the advanced state (Yt). By contrast, the second and third eigenvectors, associated with smaller eigenvalues, capture secondary modes of transformation and account for more limited structural adjustments within the system. Their weaker correlations with the perturbed coefficients indicate that the binary couplings have a much smaller influence on these secondary modes.

This finding is consistent with the model, suggesting that innovation performance is governed primarily by a single dominant configuration of H–C–M binary interactions. The remaining eigenvectors represent secondary organisational effects and structural refinements rather than alternative dominant conditions of transformation.

5. Discussion: Intelligent Environments and Ecosystemic Innovation

The three stages of the methodology, presented in Section 2 and Section 3, and 4, describe the conceptualisation of an intelligent environment for a manufacturing ecosystem, its implementation in the SmartGreenEcos experiment, and the analysis of the factors that determine its effectiveness in advancing innovation performance. Together, these stages support the hypothesis that interactions among human, collective, and machine intelligence can generate environments that improve the innovation performance of both ecosystem members and the ecosystem as a whole. The combination of expert knowledge, digital platforms, e-services, AI agents, and collaborative settings, such as living labs, co-labs, and data-sharing mechanisms, creates conditions for an intelligent environment to emerge. This environment enables companies to access shared capabilities, coordinate innovation activities, and transform operations, skills, transactions, and products across the ecosystem.

This environment is materialised as a DIP space [54]. The digital dimension encompasses collaborative platforms, e-services, and AI agents; the institutional dimension comprises agreements for collaboration, data sharing for optimisation, living labs, co-labs, joint procurement, and product development; the physical dimension comprises companies, their personnel, production infrastructure, and experts and consultants.

The model used to describe the transition is a vector-based one-time-lag autoregressive model. Its advantages include the clear structuring of the baseline state, the transformation engine, and the transition to the advanced state; the use of established innovation performance metrics; and the representation of the transformation engine through a weighted matrix of binary couplings among different forms of intelligence. The matrix coefficients are composite, combining a coupling indicator with an estimated weight that expresses the strength of the corresponding directed interaction. Other types of models, such as agent-based models, can also be used to formalise the operation of intelligent environments and assess the transition from baseline to advanced states. In such models, the agents may represent experts, companies, and digital or AI-enabled services, while their interactions are mediated through e-services, collaborative platforms, and communication mechanisms. The outcomes of these interactions would then alter the state of each company and, through aggregation, the state of the ecosystem as a whole. The advantage of agent-based models is that they can capture heterogeneity among firms, decentralised decision-making, nonlinear interactions, learning effects, and emergent patterns of innovation that may not be fully represented in a vector autoregressive formulation.

The case study shows the design of the transformation engine, the platforms and e-services deployed, and the metrics used to capture the transition. The design was collaborative, engaging experts and companies, and adapted the services to the real-life challenges of the manufacturing industry in focus. The companies were selected in the four countries following an open call and had no prior collaborations. Gradually, through the platforms and collaboration services, the sector evolves into an ecosystem. The collected data enables computation of the specific transformation engine matrix for this industry.

The way the transformation engine is described, through a weighted matrix of binary couplings among expert advice, collaborative settings and agreements, digital platforms, e-services, and AI agents, reinforces the understanding of intelligent environments as socio-technical constructs. The capabilities mobilised in this process are provided by experts, collective arrangements, and digital systems operating together. At the same time, the four subsystems outlined in the SmartGreenEcos experiment—data optimisation with AI, learning to use AI agents, transaction intensification, and product improvement—are not specific to the dairy sector alone. They are sufficiently generic to be adapted to other manufacturing sectors, although their concrete implementation must be tailored to the specific operations, value chains, and transition needs of each sector.

Simulations provide a deeper understanding of how the intelligent environment operates. The KPIs used to measure the effects of binary couplings (Table 1) capture the dynamics of the ecosystem under study, showing how specific interactions among human, collective, and machine intelligence affect intra-company operations. Simulations, by contrast, examine the internal operation of the transformation engine in greater detail. They reveal how changes in matrix coefficients and vector structure influence the transition from the baseline to the advanced state. The weights in the matrix adapt the model to specific ecosystem contexts by expressing the relative strength of each directed coupling.

The intelligent environment model we describe is specific to the manufacturing sector. This limitation defines its transferability to other ecosystems, such as energy, mobility, services, or housing. While intelligent environments are pertinent for transforming any type of ecosystem, their specific forms should be adapted to the context of the reference ecosystem (Komninos, 2026). Also, the effect of prior transformation momentum, expressed as $Y_{t-1}-Y_{t-2}$, is not included in the present model in order to keep the transformation engine simple and transparent. This contextual effect can be examined in future works as an extension of the VAR(1)-type formulation.

6. Conclusions

In this work, we framed the intelligent environment as a socio-technical construct in which forms of connected intelligence operate, each characterised by binary coupling capabilities. Intelligent environments bring into alignment:

• the manufacturing ecosystem itself, including the industry sector, participating organisations, challenges, and transition objectives;
• digital platforms, e-services, and AI agents that support innovation in operations, skills, transactions, and product development; and
• human actors, experts, and personnel who implement and use these services and tools to transform the routines and innovation performance of companies.

The interaction of these actors and resources defines the intelligent environment not merely as a technological infrastructure, but as an organised setting in which human, collective, and machine intelligence interact to support sectoral transformation. The ecosystem and its platforms provide the context and organisational structure of this environment; the transformation engine constitutes its operational core, generating advanced capabilities and knowledge functions through connected intelligence; innovation emerges both as the outcome of binary interactions and as observable evidence of intelligence in operations, learning, transactions, and products.

Innovation within this environment emerges from the transformation of baseline-state descriptors and key performance indicators (KPIs) that articulate selected objectives of change and improvement. The platforms and e-services of the intelligent environment are aligned with these objectives, while the KPIs associated with the transformation engine are selected for their ability to influence and transform the same variables. In this way, the baseline conditions of a system, the operation of the transformation engine, and the indicators used to assess transition are integrated within a common framework that guides the evolution of the ecosystem from a baseline state towards a more advanced state.

This understanding fundamentally alters conventional views of intelligent environments as primarily technological constructs. Rather than treating intelligence as embedded solely in digital infrastructures, the proposed perspective conceptualises intelligent environments as socio-technical DIP environments in which innovation emerges from interactions and couplings among human, collective, and machine intelligence. Platforms, data, and AI agents are therefore not ends in themselves, but components of a broader framework designed to generate advanced capabilities, propel innovation, and facilitate systemic change within ecosystems.