Sustainable finance has rapidly increased the demand for ESG data that is comparable, timely, and decision-useful. Yet ESG ratings and related outputs remain contested due to persistent cross-provider disagreement, heterogeneous methodologies, and an unclear boundary between ESG performance indicators and credible measurement of real-world impact. At the same time, the Artificial Intelligence era is reshaping ESG information production: providers increasingly deploy automation, machine learning, and natural language processing to scale the collection and processing of large volumes of unstructured corporate disclosures and external data. Against this background, this thesis examines how AI-enabled ESG providers operationalize ESG evaluation and how methodological and AI-related design choices influence divergence in ESG outputs, the interpretability of ESG metrics as proxies for impact, and their downstream use by financial actors. The study combines a semi-systematic literature review with a provider-level comparative analysis based on a structured matrix of 18 ESG information intermediaries. Findings show that heterogeneity starts from the output object itself (for example, ratings, datasets, alerts, or workflow artefacts) and is reinforced by recurring design drivers: benchmarking logic (relative vs. absolute), data direction and update cadence (inside-out, outside-in, hybrid), materiality and aggregation architectures, and the treatment of controversies and verification cues. AI is most often disclosed as a scaling layer for extraction/classification and reporting rather than as a replacement for core methodological assumptions. However, disclosures on governance, validation, and traceability remain uneven. Overall, ESG outputs primarily function as structured proxy signals that support financial workflows but are conditional as evidence of impact. These results underscore the need for greater transparency from providers and more responsible integration practices by financial actors.
La finanza sostenibile ha rapidamente aumentato la domanda di dati ESG comparabili, tempestivi e utili ai fini decisionali. Tuttavia, i rating ESG e i relativi output restano oggetto di controversia a causa del persistente disaccordo tra i diversi provider, di metodologie eterogenee e di un confine poco chiaro tra indicatori di performance ESG e una misurazione credibile dell’impatto reale. Allo stesso tempo, l’era dell’Intelligenza Artificiale sta rimodellando la produzione di informazioni ESG: i provider impiegano sempre più automazione, machine learning e natural language processing per scalare la raccolta e l’elaborazione di grandi volumi di disclosure aziendali non strutturate e dati esterni. In questo contesto, la presente tesi analizza come i provider ESG abilitati dall’IA operazionalizzano la valutazione ESG e in che modo le scelte di progettazione metodologica e legate all’IA influenzino la divergenza degli output ESG, l’interpretabilità delle metriche ESG come proxy dell’impatto e il loro utilizzo a valle da parte degli attori finanziari. Lo studio combina una literature review semi-sistematica con un’analisi comparativa a livello di provider basata su una matrice strutturata di 18 intermediari informativi ESG. I risultati mostrano che l’eterogeneità inizia dall’oggetto di output stesso (ad esempio rating, dataset, alert o artefatti di workflow) ed è rafforzata da driver ricorrenti di progettazione: logica di benchmarking (relativa vs. assoluta), direzione dei dati e frequenza di aggiornamento (inside-out, outside-in, ibrida), architetture di materialità e aggregazione, nonché il trattamento delle controversie e dei segnali di verifica. L’IA è più spesso dichiarata come un livello di scalabilità per estrazione/classificazione e reporting, piuttosto che come sostituto delle assunzioni metodologiche di base; tuttavia, le disclosure su governance, validazione e tracciabilità restano disomogenee. Nel complesso, gli output ESG funzionano principalmente come segnali proxy strutturati che supportano i workflow finanziari, ma sono condizionati come evidenza di impatto. Questi risultati evidenziano la necessità di una maggiore trasparenza da parte dei provider e di pratiche di integrazione più responsabili da parte degli attori finanziari.
AI-enabled ESG providers: how AI shapes ESG data and financial use in the AI era
NOVOSELOVA, IANA
2025/2026
Abstract
Sustainable finance has rapidly increased the demand for ESG data that is comparable, timely, and decision-useful. Yet ESG ratings and related outputs remain contested due to persistent cross-provider disagreement, heterogeneous methodologies, and an unclear boundary between ESG performance indicators and credible measurement of real-world impact. At the same time, the Artificial Intelligence era is reshaping ESG information production: providers increasingly deploy automation, machine learning, and natural language processing to scale the collection and processing of large volumes of unstructured corporate disclosures and external data. Against this background, this thesis examines how AI-enabled ESG providers operationalize ESG evaluation and how methodological and AI-related design choices influence divergence in ESG outputs, the interpretability of ESG metrics as proxies for impact, and their downstream use by financial actors. The study combines a semi-systematic literature review with a provider-level comparative analysis based on a structured matrix of 18 ESG information intermediaries. Findings show that heterogeneity starts from the output object itself (for example, ratings, datasets, alerts, or workflow artefacts) and is reinforced by recurring design drivers: benchmarking logic (relative vs. absolute), data direction and update cadence (inside-out, outside-in, hybrid), materiality and aggregation architectures, and the treatment of controversies and verification cues. AI is most often disclosed as a scaling layer for extraction/classification and reporting rather than as a replacement for core methodological assumptions. However, disclosures on governance, validation, and traceability remain uneven. Overall, ESG outputs primarily function as structured proxy signals that support financial workflows but are conditional as evidence of impact. These results underscore the need for greater transparency from providers and more responsible integration practices by financial actors.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Novoselova_Thesis_01.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Text of the thesis
Dimensione
3.08 MB
Formato
Adobe PDF
|
3.08 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_Novoselova_Executive Summary_02.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Text of executive summary
Dimensione
528.68 kB
Formato
Adobe PDF
|
528.68 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/252290