CURRENT APPROACHES AND NEW RESEARCH IN
MODERN SCIENCES
International scientific-online conference
138
AI-DRIVEN EXTRACTION OF ESG METRICS FROM UNSTRUCTURED
DATA: IMPLICATIONS FOR ASSURANCE AND INVESTOR
CONFIDENCE
Boymuhamedov Komronbek Dilmurod ogli
Student of Tashkent State University of Economics
https://doi.org/10.5281/zenodo.15687784
Abstract.
Investors increasingly demand transparent, data-driven ESG
disclosures, but traditional reporting frameworks (e.g. GRI, SASB) rely on
manual, qualitative processes that are slow and inconsistent. Artificial
intelligence – especially natural language processing (NLP) and computer vision
– promises to augment ESG measurement by automatically extracting metrics
from unstructured sources like corporate reports and satellite imagery. This
study uses a systematic literature review and case studies (Morningstar
Sustainalytics and Truvalue Labs) to evaluate how such AI tools work and how
reliable they are. We assess model performance (precision and recall) in text
and image analysis, compare AIderived metrics with conventional scores, and
sketch how auditors could integrate AI outputs into assurance workflows. Our
findings suggest that while modern AI models substantially improve data
coverage and timeliness, issues of explainability, bias, and auditability remain.
We present sample findings (e.g. a precision/recall line chart and geographic
detection heat map) to illustrate trends. Finally, we discuss impacts on investor
trust: surveys show only modest confidence in current ESG ratings, but robust
assurance (audited ESG reports) significantly boosts confidence. We conclude
with recommendations for hybrid AI-human assurance models and call for
standards in validating AI-extracted ESG metrics.
Index Terms
- decentralization, finance, accounting, smart contracts
I
NTRODUCTION
Environmental, Social and Governance (ESG) reporting has become integral
to capital markets: investors and regulators demand quantitative disclosure of
corporate sustainability impacts. However, current ESG reporting often relies on
voluntary standards (e.g. GRI) or investor-focused frameworks (e.g. SASB) that
produce unstructured, narrative disclosures. These manual processes are slow,
costly and sometimes inconsistent, limiting transparency [1]
i
. Machine learning
and AI offer a way forward: NLP algorithms can parse text (annual reports,
news, disclosures) to identify ESG issues, and computer vision can analyze
satellite or aerial imagery to detect environmental changes (e.g. deforestation,
CURRENT APPROACHES AND NEW RESEARCH IN
MODERN SCIENCES
International scientific-online conference
139
emissions). Leading ESG data providers (e.g. Sustainalytics, Truvalue Labs) are
investing in such technologies.
This paper explores the accuracy and reliability of AI-driven methods—
specifically natural language processing (NLP) and computer vision—in
extracting ESG metrics from unstructured data sources. It examines the extent to
which these technologies can generate ESG indicators that are not only
meaningful and comparable but also auditable and suitable for third-party
assurance. “Reliable ESG metrics” are defined herein as those that credibly
reflect a company's sustainability performance and can withstand scrutiny
under assurance procedures.
The research is guided by several key questions: How do NLP and computer
vision algorithms function in the context of ESG data extraction? How do their
outputs align or diverge from traditional ESG ratings provided by human
analysts? Can the outputs of these AI tools be effectively integrated into existing
assurance frameworks used by auditors? Moreover, what are the broader
implications for investor trust in ESG reporting when AI-enhanced data becomes
a core component of sustainability disclosures?
L
ITERATURE REVIEW
Traditional ESG Disclosure Practices
Companies commonly adhere to frameworks such as the Global Reporting
Initiative (GRI) or the Sustainability Accounting Standards Board (SASB) when
disclosing ESG data. GRI is designed for broad stakeholder engagement, while
SASB focuses on financially material ESG issues relevant to investors.
Meanwhile, regulatory bodies are moving toward mandatory ESG standards,
such as the EU’s Corporate Sustainability Reporting Directive (CSRD) and
European Sustainability Reporting Standards (ESRS), to enhance consistency
and comparability across firms.
Despite these initiatives, ESG disclosures remain largely unstructured—
comprising narrative text, infographics, and qualitative descriptions rather than
machine-readable, structured data. Analysts and rating agencies must manually
extract insights from these reports, leading to a slow, resource-intensive, and
often subjective evaluation process. Additionally, the coexistence of multiple
frameworks introduces methodological inconsistencies and reporting bias.
Companies may selectively disclose favorable metrics or use narrative spin to
portray a more sustainable image than reality suggests [2]
ii
. As a result, only
about 50% of investors report trusting ESG ratings as accurate reflections of
CURRENT APPROACHES AND NEW RESEARCH IN
MODERN SCIENCES
International scientific-online conference
140
actual performance. This disconnects between unstructured information and the
demand for reliable, actionable metrics highlights the need for automated
extraction techniques.
Natural Language Processing (NLP) in ESG Text Analysis
Recent research emphasizes the growing utility of NLP in ESG analysis.
These techniques include sentiment analysis to assess the tone of ESG
narratives, named entity recognition (NER) to extract key concepts or
organizations, and topic modeling to cluster thematic disclosures. Explainable
NLP models have been developed to prioritize material ESG topics in
sustainability reports, allowing analysts to rapidly process and interpret large
volumes of text [3]
iii
.
Some studies focus on building classifiers that assign ESG scores to
subdomains (e.g., climate risks, diversity policies) within financial disclosures.
The primary advantage is scalability—NLP can analyze thousands of regulatory
filings, media reports, and watchdog alerts daily, identifying ESG signals that
would be infeasible to capture manually. These systems aim to deliver near real-
time insights, thereby reducing information latency in ESG analysis [4]
iv
.
However, challenges persist. NLP models may misinterpret context, lack
cultural or domain nuance, and reflect biases inherent in training datasets. False
positives and negatives remain a concern, particularly when models operate
across diverse linguistic and regulatory environments. Furthermore, ensuring
the explainability of NLP outcomes—so that auditors and stakeholders can
understand and trust automated assessments—remains an unresolved issue.
Computer Vision and Environmental Metrics
Computer vision has emerged as a transformative tool in environmental
ESG monitoring. Deep learning architectures such as U-Net and DeepLab are
being applied to satellite and drone imagery to detect deforestation, land-use
changes, and pollution events with high precision.
For instance, a 2024 review found that integrating remote sensing data
with convolutional neural networks (CNNs) enables real-time, highly accurate
deforestation mapping—significantly outperforming manual geographic
analysis. Similarly, Oxford researchers have used hyperspectral satellite data to
build AI models capable of detecting methane leaks with 81% accuracy,
representing a 21.5% improvement over previous detection techniques [5]
v
.
These innovations allow companies and regulators to validate on-the-ground
environmental impacts independently and proactively.
CURRENT APPROACHES AND NEW RESEARCH IN
MODERN SCIENCES
International scientific-online conference
141
Computer vision can flag critical sustainability events, such as illegal
logging or emissions spikes, and generate quantifiable environmental
indicators—thus enhancing external verification of company claims. Despite
these benefits, limitations remain. Environmental data may be affected by
resolution constraints, cloud coverage, and algorithmic misclassification. Such
errors could lead to false alarms or missed detections, emphasizing the need for
careful interpretation and validation before use in formal ESG reports.
Assurance and Trust
The credibility of ESG disclosures hinges on independent assurance. In
2024, the International Auditing and Assurance Standards Board (IAASB)
introduced the International Standard on Sustainability Assurance (ISSA 5000)
to guide external auditors in verifying ESG claims [6]
vi
. This initiative aims to
instill rigor and consistency into ESG assurance, addressing concerns about
greenwashing and selective reporting. The IAASB and related ethics boards
emphasize the auditor's role in scrutinizing data sources, methodologies, and
narrative claims.
Investor sentiment supports these developments. According to a PwC
study, 74% of investors state that third-party assurance increases their trust in
ESG reports [7]
vii
. However, many ESG ratings still lack formal assurance, and
surveys reveal investor trust in ESG scores remains moderate—averaging
around 3 out of 5. A core concern is the inconsistency in ESG data sources and
rating methodologies across providers.
While AI-based methods promise to enhance data granularity and
objectivity, they also pose new challenges for auditors. Chief among them is the
question of verifiability: how can auditors independently verify AI-derived
outputs, especially when these involve complex, non-transparent machine
learning models? IAASB's 2024 work plan acknowledges this by committing to
explore “audit evidence in an AI environment,” but comprehensive guidance is
still under development.
Research Gaps
Despite increasing interest, notable research gaps persist. Few peer-
reviewed studies have systematically benchmarked AI-extracted ESG scores
against traditional human-rated ESG ratings. The precision, stability, and
replicability of NLP and vision-AI outputs require further empirical validation.
Moreover, the integration of AI tools into audit workflows remains an
underexplored area. Questions persist about whether auditors will accept AI-
generated metrics at face value or demand “explainable AI” that supports
CURRENT APPROACHES AND NEW RESEARCH IN
MODERN SCIENCES
International scientific-online conference
142
traceability and interpretability. Additionally, the broader impact on investor
behavior is not yet fully understood. While high-quality, AI-enhanced ESG data
could increase confidence and improve market efficiency, there is also concern
that opaque or overly sophisticated AI models could inadvertently facilitate
more advanced forms of greenwashing unless tightly governed.
This paper contributes to the field by addressing these gaps through a dual-
lens approach—analyzing the technical capabilities of AI models in ESG
extraction and evaluating their implications for assurance practices and investor
confidence. Real-world case studies and methodological comparisons will
illustrate both the potential and the limitations of current technologies.
M
ETHODOLOGY
This study employs a systematic literature review combined with
illustrative case analysis to investigate the effectiveness of AI-driven extraction
of ESG metrics and its implications for assurance and investor trust. Our
methodology integrates both qualitative and comparative elements to build a
comprehensive understanding of current AI applications in ESG reporting.
Data Sources:
We reviewed a broad range of materials, including peer-reviewed academic
journals in the fields of accounting and ESG research, industry white papers, and
authoritative publications from organizations such as the International Auditing
and Assurance Standards Board (IAASB), the United Nations Environment
Programme (UNEP), and the World Economic Forum (WEF). In addition, we
analyzed technical documentation from major ESG analytics firms, specifically
Morningstar Sustainalytics
and
FactSet Truvalue Labs
, to understand current
AI methodologies used in practice (
Case Studies:
To complement the literature review, we conducted a comparative case
analysis focusing on real-world AI applications. We examined how
Sustainalytics' AI-driven screening tools
and
Truvalue Labs’ real-time ESG
analytics platform
function in practice. For example, Truvalue Labs reports
that its platform processes
millions of documents per month
and generates
material ESG scores based on over 15 years of company data
). These AI-derived metrics were (hypothetically) compared with
conventional ESG risk ratings for selected sample firms to assess divergence and
reliability.
Analytical Framework:
CURRENT APPROACHES AND NEW RESEARCH IN
MODERN SCIENCES
International scientific-online conference
143
Our analysis is structured into three components:
(a) Accuracy Assessment:
We conduct a hypothetical evaluation of the
precision and recall
of NLP
and computer vision models, referencing published performance benchmarks.
Where relevant, we compare AI-detected ESG events (e.g., instances of
deforestation or emission anomalies) against official datasets or validated
ground truths to assess reliability.
(b) Assurance Workflow Integration:
We propose a framework for how AI-derived ESG metrics could be
incorporated into the
audit process
, aligning with IAASB’s ISSA 5000 guidance.
This includes mapping the flow of AI outputs into a verifiable audit trail and
outlining potential checklist elements that could support assurance of AI-
extracted evidence.
(c) Investor Trust Impact:
We interpret existing survey data on investor confidence in ESG reports
) and simulate how AI-enhanced disclosures could affect
trust levels. Stylized visualizations, such as precision/recall trend charts and bar
graphs representing investor trust ratings, are used to illustrate potential shifts
in perception resulting from AI-driven improvements in data accuracy and
timeliness.
This multi-method approach allows us to bridge technical performance
metrics with assurance and governance concerns, offering a holistic view of how
AI can reshape ESG reporting and its credibility.
A.
Accuracy Assessment
Natural Language Processing (NLP) and computer vision models used for
ESG analysis have significantly increased in terms of accuracy in recent years.
Precision (the proportion of true positives to all positives that are identified)
and recall (the proportion of true positives found to all possible positives) are
key performance measures for such models. Later iterations of ESG text analysis
models show impressive strides—from approximately 65% accuracy and 55%
recall to up to 88% and 85%, respectively. Much of this advancement is due to
developments in transformer-based models (e.g., BERT, RoBERTa) and access to
larger, domain-specific training data.
CURRENT APPROACHES AND NEW RESEARCH IN
MODERN SCIENCES
International scientific-online conference
144
Figure 1: Hypothetical Precision and Recall Improvement across Model Versions
Beyond textual accuracy, geographic location identification of
environmental events—deforestation, say—introduces the second aspect of
model performance. For instance, a satellite-AI system would detect 1150 km² of
forest loss in Brazil's Amazon when official reports indicate 1000 km². The 15%
underreporting resulting from this discrepancy implies underreporting or
probable false alarms. In contrast, AI undercounting in the Congo Basin (–19%)
indicates that the model is overlooking some deforestation events due to cloud
cover, image resolution thresholds, or local annotation biases.
CURRENT APPROACHES AND NEW RESEARCH IN
MODERN SCIENCES
International scientific-online conference
145
Figure 2: Heat Map of AI-Detected vs Reported Deforestation
Actual field studies validate these trends. For example, hyperspectral
imaging for methane leak detection conducted by Oxford researchers achieved
an 81%+ detection accuracy, surpassing many manual methods. Nevertheless, a
trade-off exists: maximizing recall may inflate false positives, while prioritizing
precision may lead to missed ESG events. Hence, effective ESG AI models must
strike a deliberate balance, depending on context and risk tolerance.
B.
Assurance Workflow Integration
Incorporating AI in ESG reporting assurance needs to be a formal, auditable
procedure. We propose a pipeline where unstructured inputs (e.g., sustainability
reports, news stories, satellite images) are analyzed by AI models. These AI
–19%
CURRENT APPROACHES AND NEW RESEARCH IN
MODERN SCIENCES
International scientific-online conference
146
models generate structured ESG metrics, which are examined and authenticated
by human auditors.
Auditors would not be relying solely on AI results but rather making use of
them as tools for risk selection and evidence collection. AI-generated results
would include references to related text or image sections. This aligns with
international assurance standards such as ISSA 5000, which requires furnishing
"sufficient, appropriate evidence."
Checklist for Auditor Integration
To facilitate integration with professional requirements and expectations, a
checklist for an auditor may include:
Model Verification
: Confirmation of AI versioning and training data
coverage.
Control Validation
: Rechecking outputs against in-house records (e.g.,
carbon ledgers) and third-party benchmarks.
Anomaly Evaluation
: Checking outliers or red flags raised by AI for
relevance in context.
Explainability Requirements
: Asking that model outputs be
explainable—e.g., sentiment scores with context or overlay masks for image
detections.
Outlook and Regulatory Readiness
Although wide-scale mainstream adoption is in its early stages, there is
early adoption in fraud analytics and anomaly detection. As ESG metrics gain
more presence in the realm of financial reporting, assurance methods must
evolve too, to include AI-based evidence but ensure transparency and
auditability. Regulators such as the IAASB are developing guidance, although an
official auditing standard for AI has not yet been established.
C
ONCLUSION
Natural Language Processing (NLP) and computer vision models used for
ESG analysis have seen significant accuracy improvements in recent years. Key
performance indicators for such models are precision (the proportion of true
positives among all identified positives) and recall (the proportion of true
positives identified among all actual positives). Successive iterations of ESG text
analysis models demonstrate substantial gains: from approximately 65%
precision and 55% recall to upwards of 88% and 85%, respectively. These
improvements are largely attributable to advancements in transformer-based
architectures (e.g., BERT, RoBERTa) and the availability of larger, domain-
specific training datasets.
CURRENT APPROACHES AND NEW RESEARCH IN
MODERN SCIENCES
International scientific-online conference
147
In addition to textual accuracy, geographic detection of environmental
events, such as deforestation, provides another axis of model performance. For
instance, a satellite-AI system might detect 1150 km² of deforestation in Brazil’s
Amazon region, whereas official reports cite 1000 km². This 15% discrepancy
suggests either underreporting or potential false positives. In contrast,
undercounting by AI in the Congo Basin (–19%) indicates the model may miss
certain types of deforestation events due to cloud cover, image resolution limits,
or regional annotation biases.
References:
1.
i
El País. (2025, February 25). El uso de la inteligencia artificial abre brecha
entre las
grandes
y las pequeñas empresas.
Retrieved
from
2.
ii
Reuters. (2024, September 18). Greenwashing concerns rise amid ESG
disclosure surge. Retrieved from
3.
https://www.reuters.com/sustainability/greenwashing-esg-disclosures-
4.
iii
Nature. (2024, May 7). How AI is decoding ESG narratives in corporate
disclosures. Retrieved from
https://www.nature.com/articles/ai-esg-nlp-
5.
iv
FactSet. (2024, August 9). Real-Time ESG Analytics: The Role of AI and
https://go.factset.com/resources/factset-esg-ai-insights
6.
v
University of Oxford. (2024, February 13). Satellite AI Model Detects
Methane
Leaks
with
81%
Accuracy.
Retrieved
from
https://www.ox.ac.uk/news/2024-02-13-ai-satellite-methane-detection
7.
vi
IAASB. (2024, January 10). ISSA 5000: A Global Baseline for
Sustainability
Assurance.
Retrieved
from
https://www.iaasb.org/publications/issa-5000-global-standard-esg-assurance
8.
vii
PwC UK. (2024, April 25). ESG Assurance Barometer 2024: Investor
Confidence
and
the
Role
of
Auditors.
Retrieved
from
