Multidimensional Sustainability Assessment

• Multidimensional sustainability assessment frameworks are formalized methodologies that evaluate systems across environmental, economic, social, technical, and governance indicators.
• The approach involves sequential steps of indicator measurement, normalization, weighting, aggregation, and scenario-based comparison to ensure reproducible evaluations.
• Empirical validations demonstrate significant reductions in energy, CO2, and water usage for AI-assisted and agentic AI scenarios compared to manual processes.

A multidimensional sustainability assessment framework is a formalized approach to evaluating and comparing systems, workflows, or policies across several interrelated environmental, economic, and social indicators—frequently extended to include technical and governance dimensions. Such frameworks provide rigorous, quantitative, and reproducible methodologies for measuring the sustainability impacts of complex human and technological systems. In advanced applications, these frameworks integrate systematic indicator design, normalization, interdependency modeling, aggregation, scenario analysis, and replication protocols, enabling transparent governance and informed trade-off analysis across sustainability objectives (Gosmar et al., 10 Nov 2025).

1. Structural Principles and Objectives

Multidimensional sustainability assessment frameworks are characterized by a formalized architecture consisting of sequential measurement, normalization, weighting, aggregation, and scenario-based comparison steps. The core structural sequence is:

1. Indicator Measurement: Direct quantification of core sustainability indicators (e.g., energy consumption $E$, CO$_2$ emissions $C$, water usage $W$) and performance metrics (e.g., throughput, operator count, processing time) for each system or workflow scenario.
2. Normalization and Scaling: Mapping raw indicator values to a standardized $[0,1]$ interval via min–max normalization. This homogenizes disparate units and scales (such as kWh, kgCO$_2$, L of water), allowing cross-comparison and aggregation.

$$x'_i = \frac{X_i - \min_j(X_j)}{\max_j(X_j) - \min_j(X_j)}$$

where $X_i$ is the raw value for scenario $i$, and $j$ indexes all scenarios under evaluation.

1. Weighting: Assignment of relative importance (weights) $w_k$ to each sustainability dimension. The sum of weights is constrained as $\sum_k w_k = 1$.
2. Aggregation: Computation of a single composite sustainability score $S_i$ for scenario $i$ as a weighted sum of normalized indicators:

$S_i = \sum_{k} w_k\,x^{(k)}_i,\quad S_i \in [0,1]$

Lower values of $S_i$ indicate more sustainable (lower-impact) configurations.

1. Scenario Analysis and Benchmarking: Systematic comparison across distinct operational modes (e.g., manual, AI-assisted human-in-the-loop, fully agentic AI) to quantify and visualize sustainability gains, trade-offs, and opportunity costs. Rigorous scenario selection supports fair comparison by normalizing for task volumes and workloads (Gosmar et al., 10 Nov 2025).
2. Replicability and Adaptation: Explicit guidance on parameter selection, relevant data sources, normalization methodology, weight selection (ESG priorities), and adaptation pathways for other domains ensures reproducibility and transferability to heterogeneous organizational or sectoral contexts.

Framework objectives include:

• Enabling holistic quantification and benchmarking of environmental and resource footprints across system alternatives.
• Supporting integration of sustainability metrics into ESG governance and reporting structures.
• Facilitating decision-maker interpretation of trade-offs between automation gains, efficiency, compliance, and ecological impact.

2. Indicator Definitions, Formulations, and Normalization

Central to the multidimensional approach is the formal definition and per-scenario calculation of distinct sustainability indicators, along with their normalization for aggregation:

Energy consumption ($E$):

• $E_i$ per scenario is partitioned into human-operated workstation energy ($E_{\mathrm{human},i}$) and AI/cloud inference energy ($E_{\mathrm{cloud},i}$):

$E_i = E_{\mathrm{human},i} + E_{\mathrm{cloud},i}$

$$E_{\mathrm{human},i} = n_{\mathrm{ops},i} \times t_{\mathrm{shift}} \times P_{\mathrm{ws}}$$

$E_{\mathrm{cloud},i} = D \times e_{\mathrm{doc},i}$

$e_{\mathrm{doc},i} = \frac{E_i}{D}$

CO$_2$ emissions ($C$):

• Computed using a grid emission factor EF (gCO$_2$/kWh):

$$C_i = E_i \times \frac{\mathrm{EF}}{1000},\quad c_{\mathrm{doc},i} = e_{\mathrm{doc},i} \times \frac{\mathrm{EF}}{1000}$$

Water usage ($W$):

• Using Water Usage Effectiveness (WUE):

$$W_i = E_i \times \mathrm{WUE},\quad w_{\mathrm{doc},i} = e_{\mathrm{doc},i} \times \mathrm{WUE}$$

Performance metric (throughput $T$):

• Standardized throughput per operator per day:

$$T_i = \frac{D}{n_{\mathrm{ops},i} \times t_{\mathrm{shift}}}$$

All indicators are mapped per scenario and normalized as described in Section 1 (min–max normalization).

Scenario example results (for $D=5,000$ documents/day, EF = 288 gCO$_2$/kWh, WUE = 0.18–0.30 L/kWh):

Scenario	$n_{\mathrm{ops}}$	$E$ (kWh/d)	$C$ (kgCO$_2$/d)	$W$ (L/d)
Manual	70–400	36.3–194.7	10.5–56.1	35.1–58.4
AI-Assisted	7–28	6.1–16.2	1.8–4.7	1.1–4.9
Agentic AI	7–28	9.8–20.5	2.8–5.9	1.8–6.2

(Gosmar et al., 10 Nov 2025)

3. Aggregation, Scenario Comparison, and Empirical Validation

Aggregation through ESG-Weighted Scoring:

The composite ESG-oriented sustainability score $S_i$ for each scenario is:

$$S_i = w_E\,x'^{(E)}_i + w_C\,x'^{(C)}_i + w_W\,x'^{(W)}_i,\quad S_i \in [0,1]$$

where the default is equal weighting ($w_E = w_C = w_W = 1/3$), but can be set according to governance or stakeholder priorities. This formalizes ESG compliance and prioritization.

Empirical Validation:

Empirical results show:

• Energy reductions vs. manual process: HITL, 70–90%; agentic AI, 73–90%.
• CO$_2$ reductions: HITL, 83–92%; agentic AI, 73–90%.
• Water reductions: HITL, 94–97%; agentic AI, 91–97%.

Quantitative reductions confirm the framework’s empirical effectiveness in discriminating sustainability performance (Gosmar et al., 10 Nov 2025).

Aggregated result table:

Reduction (%) vs. Manual	HITL	Agentic AI
Energy	70–90	73–90
CO$_2$	83–92	73–90
Water	94–97	91–97

4. Extension, Response to Trade-offs, and Adaptation Protocols

Parameterization and Replication:

• All main parameters (document volume $D$, per-document energy $e_{\mathrm{doc},i}$, emission factor EF, WUE, PUE, weights $w_k$) are explicitly specified.
• Emphasizes real-world data acquisition (time-and-motion studies, cloud energy reports, data center benchmarks).
• The template can be instantiated for other domains by adapting input parameters and indicator definitions.

Potential Extensions:

• Incorporation of Scope 3 life-cycle assessment (hardware manufacturing, EOL).
• Additional ESG dimensions (e.g., labor metrics, supply-chain equity).
• Dynamic or participatory weighting schemes.
• Scenario routing (assigning tasks according to typological sustainability/complexity).
• Real-time dashboards for governance monitoring.

Adaptation Guidance:

Vertical (across business units) and horizontal (across supply chains) roll-out is supported, with templates for aggregation and benchmarking provided. Adaptive cycles (annual reweighting, data refresh, scenario tuning) ensure the method remains relevant under dynamic operational and regulatory conditions (Gosmar et al., 10 Nov 2025).

5. Best Practices, Limitations, and Future Directions

Best Practices:

• Comprehensive parameter documentation and normalization.
• Use multi-source, reproducible data.
• Engage stakeholders in weight-setting.
• Employ scenario benchmarking and validation experiments.
• Document and review boundary conditions for adaptation to new sectors.

Limitations:

• Scope is presently limited to ecological and resource metrics; social and governance metrics are not yet embedded, though the framework can be extended.
• The default aggregation (linear weighted sum) presumes independence among indicators; if synergy or redundancy (non-additivity) is important, integration with a Choquet integral or similar non-additive methods may be warranted (Angilella et al., 2018).
• Sensitivity to indicator selection and subjective weighting, mitigated through transparent documentation and stakeholder inclusion.

Future Research:

• Inclusion of multi-objective or non-additive aggregation to capture indicator interactions.
• Extensions to time-dependent, trajectory-based sustainability assessment.
• Automated scenario splitting based on task complexity.
• Integration with ESG dashboards and life-cycle analytics.
• Empirical benchmarking across domains, including finance, healthcare, legal document processing, and others.

6. Practical Implementation and Domain Transferability

The framework is designed for practical implementation with a four-step workflow:

1. Measurement: Gather and quantify all relevant indicators for each scenario.
2. Normalization: Apply min–max scaling across all scenarios to bring different units to a common basis.
3. Aggregation: Calculate the weighted sum ESG-score.
4. Comparison/Benchmarking: Rank, visualize, and interpret scenario sustainability, identifying the lowest-impact (best) alternatives.

This architecture is immediately transferable to other document-intensive and data-centric workflows, providing a rigorously validated, transparent, and adaptable methodology for multidimensional sustainability assessment in AI-enabled operational contexts (Gosmar et al., 10 Nov 2025).