Driving Sustainable Value. The Dynamic Interplay Between Artificial Intelligence Disclosure, Financial Reporting Quality, and ESG Scores

Abstract

Adapting contemporary business models to the challenges of implementing new technologies influences the sustainable value of companies. This study examines the disclosure practices of Romanian-listed companies regarding accounting estimates, their correlation with financial performance, ESG scores, and the use of artificial intelligence (AI). Financial data was gathered from annual reports and those regarding the use of AI on companies’ websites. Financial performance was measured through profitability and liquidity indicators. The results of the statistical regressions showed that company size can influence AI disclosure; however, industry is not a strong predictor, and the number of employees does not significantly influence AI disclosure. A positive relationship was found between AI transparency and the current ratio, suggesting that companies disclosing more information about their AI use may have higher current liquidity. Additionally, a statistically significant negative relationship was observed between the AI disclosure score and net profit, indicating that greater AI transparency is associated with lower net income. The results of interaction analysis proved that there may be a relationship between ESG exposure and financial performance when considering AI disclosure. However, this result may be considered controversial in a more conservative analysis, emphasizing the need for a more nuanced and multidimensional approach.

1. Introduction

High-quality financial reporting is one of the main persistent issues in the current literature, especially how to assess the quality of disclosed information [1]. Nguyen et al. [2] argue for a strong and positive relationship between the level of trust in narrative information disclosure and company value. Thus, companies should prioritize clear, consistent, and credible communication in their annual reports to positively influence investor perceptions and, consequently, their company’s value. At the same time, Nguyen et al. [2] noted that complex narrative information disclosure can conceal certain key information, generating confusion or uncertainty among investors, thereby risking the diminution of the company’s sustainable value. A holistic interpretation of the results suggests that narrative disclosure should be utilized as a tool to promote the value of the company. A particular field of interest in empirical research is the analysis of reporting and disclosure behavior regarding accounting estimates, as well as the use of artificial intelligence (AI) technologies by listed companies. In several empirical studies, researchers examined the nexus between disclosed accounting estimates, financial performance growth, and the share of R&D in value-added industries [3,4,5]. They find that higher-quality accounting disclosures are associated with business growth and development in equity-based industries. Other researchers focused on the AI technologies’ influence on accounting estimates' accuracy and reliability [6,7] and examined whether using AI technologies improves the timeliness and relevance of financial information or whether it introduces new challenges and risks in terms of accuracy and accounting estimate bias [8,9].

Through this study, we aim to understand how companies report and disclose financial information related to accounting estimates, performance, and the use of AI technologies. Disclosure of AI technologies and algorithms in the annual reporting of publicly traded companies is a growing area of interest. In this approach, we aim to seek answers to several questions: to what extent and how do companies disclose information about artificial intelligence? Does the quantity and quality of information disclosed by companies using artificial intelligence influence financial reporting and performance? Is there a correlation between ESG scores and AI disclosure? For this purpose, quantitative methods will be used in managing financial data, as well as qualitative methods in the content analysis of reports and other publicly available information.

Salawu and Moloi [10] pointed out the need to adopt standards that provide specific provisions for AI asset estimation, suggesting the use of a simple algorithm to measure them. Yingfan et al. [11] proposed an automatic method for measuring information disclosure in annual reports, based on digital-twins technology (DTL), which could be enhanced with AI technologies. Baldwin et al. [12] and Loukeris [13] pointed out the potential of AI to improve the accuracy and efficiency of accounting estimates in the financial reporting of listed companies. Machine learning, a subset of AI, is becoming particularly emphasized in empirical accounting research as a tool to improve accounting estimates reporting and disclosure [6,14]. The context and motivation for this research involve understanding how companies report and disclose financial and AI-related information by examining annual reports and information available on their websites. The financial information provided by companies is vital for investors and other stakeholders, and with the rise of AI in various fields in recent years, it is important to understand how this technology is used by companies, considering its ethical and social implications. Detailed analysis of these aspects will ensure a thorough understanding of how companies manage their financial and technological information, which can serve as a basis for improving transparency and decision-making in the contemporary business environment. Moreover, this research seeks to uncover any discrepancies or variations in communication practices that could inform the need for regulatory interventions or best practice guidance. It also examines the link between AI adoption disclosure and financial performance but also touches on the issue of ESG risk management by companies, a topic still insufficiently explored in the literature. Very recently, several authors [15,16,17] have focused their efforts on analyzing the correlations between ESG factors, whether they are examined in terms of performance, quality of information communication, or scores on risk management, and the adoption of AI systems and technologies. For instance, the results of the study conducted by Zhou et al. [17] reveal that AI contributes to a significant increase in corporate ESG performance, indicating an average increase in ESG scores among pilot-listed companies. The present study also falls within this context.

The structure of this paper includes the theoretical foundations of this study, the description of the data and the methods used in the analysis, the results and discussions, and in the last section, the conclusions, limitations, and possible further avenues of research are integrated.

2. Theoretical Background and Research Hypotheses

Artificial intelligence is defined as the ability of a system to effectively understand information from the external environment, gather knowledge from that information, and use the knowledge to adapt in a versatile way in performing certain operations and achieving goals [18]. Also, in recent years, financial reporting practices have been under continuous transformation due to the technological revolution generated by artificial intelligence [19]. Improving external financial reporting through the adoption of artificial intelligence would translate into increasing the accuracy of information as well as the efficiency of its communication with interested parties [20]. According to Ashraf [21], a lower incidence of material internal control deficiencies is significantly associated with the automation of operations. Therefore, companies that implement automation achieve higher quality financial reporting due to a stronger internal control environment. Previous studies conducted by Fülöp et al. [22], Commerford et al. [23], Gurmu and Miri [24], and Ding et al. [7] highlighted various issues related to accounting estimates and the use of AI within companies. For example, the mentioned works looked at how companies use data estimation models and advanced technologies to improve accounting processes and financial decisions. Also, these studies highlighted the influence of AI implementation on the efficiency and business performance of companies, underlining the advantages and challenges related to the use of these systems in the business environment. Shiyyab et al. [25] conducted investigations on credit institutions examining the influence of AI use on financial indicators. The results showed that the Jordanian banks are still at the beginning of the process of adopting and assimilating AI systems in their ongoing operations, and the disclosure degree of AI information is low. The findings of the study indicate that AI keywords disclosure influences the financial performance of banks. AI has a positive impact on financial performance, concerning return on assets (ROA) and return on equity (ROE), and a negative influence on total expenses, aspects that confirm the prevailing hypothesis that AI increases revenues and decreases certain categories of costs [25]. The analysis performed in this study inspired and made us determined to choose the financial indicators ROA and ROE in this research. Another study conducted by Bonsón and Bednarova [26] showed that AI voluntary disclosure is rare and indicates an absence of standardization, being mostly prevalent in the financial technology and telecommunications sectors. Recent research highlights an emerging trend in the use and reporting of AI, discussing the need to establish a global standard for such communication [26,27].

AI systems can rapidly process huge amounts of data, quickly perform repetitive tasks of grouping, cleaning, sorting, and ranking, and can quickly analyze complex data sets, identifying patterns and generating insights that can assist human decision-making [28,29,30]. AI can increase the efficiency and effectiveness of decision-making processes, leading to improved task performance compared to classical methods.

Definitely, in the context of the increasing integration of AI into the operational processes of companies, to produce high-quality financial reports, the focus is also on shaping the skills of accountants and managers, both to improve financial analysis skills and to increase the relevance of information provided to stakeholders [31,32,33]. Managers’ openness to new perspectives on the integrated business model, based on AI implementation, as well as the assimilation of specific skills, is also emphasized by Estep et al. [34], noticing that when the company employs AI, managers register higher audit restatements based on audit evidence obtained through the use of technologies, rather than from human specialists. On the other hand, more and more researchers are interested in the correlation between the effects of the implementation of integrated reporting and its quality analysis concerning the performance of companies, allowing the comparative analysis of companies that use financial reporting complemented by that for sustainability [35].

At the same time, it is worth examining the relationship between company size indicators and investments in AI systems, as well as the level of transparency in communicating this information. In this regard, Babina et al. [36] developed an evaluation model that enables the analysis of both the determinants and the results of AI investments made by companies across various economic sectors. Babina et al. demonstrated a positive relationship between AI investments and company size. Thus, AI investments are concentrated among the largest companies, and as these companies invest in AI technologies, they grow, increasing their sales, employment, and market share. Companies that invest in AI systems expand through innovation [36]. In line with current trends, the results of the predominantly qualitative study conducted by Przegalinska et al. [37] suggest that incorporating AI technologies can significantly enhance organizational performance in aspects related to automation, assistance, creative efforts, and innovation processes. The authors examine the possibility of strategic implementation of AI in the context of business organizations, noting that generative AI, compared to human intelligence, triggers more positive feelings, uses simple, friendly language, and has a narrower vocabulary of terms.

Similarly, the results obtained by Liu et al. [38] converge, indicating that the green innovation effect of industrial robot applications is more pronounced in companies with high operational levels. As the authors note, large corporations have greater financial resources to acquire robotic systems, compared to SMEs, thus facilitating the development of green innovation technologies. At the same time, large companies have a higher degree of compliance with environmental regulations and are more receptive and open to implementing government policies aimed at sustainability through green innovation [38,39,40,41]. Other authors, such as Odugbesan et al. [42] and Wen et al. [43], believe that the integration of robots in companies increases productivity and innovation capacity. However, at the same time, the possibilities of reuse and significantly superior performance can reduce energy costs for production and promote green development. Odugbesan et al. [42] show that advanced AI technologies positively moderate corporate innovation and the development of green talents.

ESG scores are well-established concepts in capital markets, reflecting companies’ sustainability actions and enabling investors to evaluate non-financial performance in correlation with financial performance [44,45,46,47]. Most studies examining the relationship between ESG scores and artificial intelligence focus on measuring ESG performance using AI tools, rather than exploring the link to the degree of disclosure of AI use. Fluharty-Jaidee and Neidermeyer [48] identify the use of AI to predict poor sustainability reporting and ESG performance measurement practices generated by various more or less ethical behaviors regarding incentives for inappropriate sustainability reporting. Research results obtained by Zhang [49] show that AI acts as a deterrent to greenwashing while enhancing sustainable development. Thus, AI contributes to eliminating the negative impact of greenwashing by also increasing the quality of ESG rating scores in disclosure practices. As Zhang notes, the effectiveness of AI in reducing greenwashing behaviors is well-appreciated in state-owned enterprises, companies operating in less polluting industries, regions with strict environmental regulations, and those with less developed green finance sectors. Tian and Shi’s [50] qualitative study uses textual content analysis to count the frequency of words referring to AI in companies’ annual reports and their ESG scores. The research results argue that the application of AI technology can effectively reduce the occurrence of greenwashing behavior, and the influence of devices indicates that green innovation partially mediates the relationship between AI and corporate greenwashing. Detailed analyses show that the inhibitory effect of AI on corporate greenwashing is more pronounced in private enterprises, large enterprises, and enterprises in high-pollution industries. Despite the increasing number of studies addressing the link between the use of AI technologies and tools and corporate reporting in the context of sustainable business development, there is a clear lack of studies examining the various aspects of this connection.

In the context described above, we proposed to test the following assumptions:

Hypothesis H1.

The AI technologies information disclosure is significantly influenced by the size of the companies, the industry type, and the average number of employees, developed in the following sub-hypotheses: H1(a): There is a significant positive correlation between the average disclosure score of AI technologies and company size; H1(b): There is a significant positive correlation between the average AI technology disclosure score and industry typology; and H1(c): There is a significant positive correlation between the average disclosure score of AI technologies and the average number of employees.

Hypothesis H2.

There is a significant positive correlation between companies’ financial performance and AI disclosure, with the following sub-hypotheses: H2(a): There is a significant positive correlation between current liquidity (CR) indicator and AI disclosure; H2(b): There is a significant positive correlation between return on assets (ROA) indicator and AI disclosure; and H2(c): There is a significant positive correlation between return on equity (ROE) indicator and AI disclosure.

Hypothesis H3.

There is an interaction between ESG performance scores and AI disclosure in terms of the impact on companies’ financial performance. Companies that have better ESG scores and disclose AI information on their websites may have superior financial performance compared to those that have poorer ESG scores or do not disclose AI information.

3. Materials and Methods

This research focuses on assessing the quality of financial information reporting and disclosure, and those regarding the use of AI technologies, and uses a combination of quantitative and qualitative methods that provide a balanced and comprehensive approach to the subject. Furthermore, we present the methodological tools used.

Sample selection. To establish the sample, we focused on the analysis and the level of innovation in the private entities listed on the Bucharest Stock Exchange (BVB), with an emphasis on the main segment and the Premium category. The choice of this field meets the need to understand more deeply how elite companies evolve and develop in the specific economic context of the capital market in Romania. One of the main reasons for this choice is the increasing relevance of financial markets and the impact these companies have on the economy. Companies listed and included in the main segment and the Premium category have a significant influence on market development, often being representative of key sectors of the economy. To constitute the sample, the banking sector and companies that, for various reasons, were included in the Special Watch List were eliminated, leaving fifty-five companies grouped in ten industries. The analyzed period refers to the years 2018–2022.

Data description and collection. Public financial information was accessed by consulting the annual reports of the selected companies. Three files were built to centralize the collected information. After the complete collection of all financial data, the financial indicators, CR (current ratio), ROA (return on assets), and ROE (return on equity) were calculated. Starting from Farcane et al.’s [5] work, the data set on accounting estimates was constituted by applying the qualitative research method. We examined the public information from the explanatory notes to the annual financial statements regarding the selected categories of accounting estimates (17 in number), using a separate file to quantify the information presentation quality through the scoring technique. To calculate the average disclosure degree of information, a scoring range from 0 to 3 was used, where 0 represents the total lack of information related to the subject; 1 indicates an unsatisfactory level of presentation, i.e., the information is poorly presented or very little information was disclosed; 2 shows that the information is presented at an average or satisfactory level; and 3 that the information is very well presented and communicated in detail. Information on ESG scores was gathered from https://bvbresearch.ro/ReportDashboard/ESGScores (accessed on 15 March 2024), accessed in March 2024, following the Sustainalytics methodology [51]. The information on ESG scores downloaded from https://bvbresearch.ro/ReportDashboard/ESGScores (accessed on 15 March 2024) and used in this study refers to ESG risk, ESG exposure, ESG management, and ESG risk ranking score. The content analysis of annual reports revealed that companies do not disclose AI information for the analyzed period. Considering the fast evolution of technology and its impact on various fields, research and analysis of specific technologies are becoming increasingly important to understand and adapt to ongoing changes. In this context, the analysis and evaluation of technologies such as Blockchain, Machine Learning (ML), Natural Language Processing (NLP), Robotic Process Automation (RPA), Internet of Things (IoT), and Resource Planning Systems (SAPs) are topics of great interest within the academic, industrial, and research communities [52,53]. The choice of these technologies for research is justified by their relevance in various fields, their potential to bring innovations and significant changes in how certain activities or processes are carried out, and their association with competitive advantages, operational efficiency, and adaptability to market demands [54,55]. Since no information was found on the use of AI technologies and systems in annual reports, other public means of communicating information were examined, and thus, the buttons and information accessible on the companies’ websites were analyzed. Hence, information disclosed on AI utilization was collected manually from the companies’ websites. Information disclosed on Blockchain technologies, ML, NLP, RPA, IoT, and SAP solutions was inventoried. Information about the use of AI was mainly identified under the following information buttons: projects, programs and strategies, news, and the administrator’s report. These means of communication provide essential information on employing AI in companies, covering various enforcement issues and the impact of techniques in different contexts and initiatives. Therefore, to perform analyses, information regarding the use of AI in operational activities found on the companies’ websites was noted with 1 and with 0 cases where no information on the use of AI was disclosed.

Description of variables and statistical data processing. Following accelerated technological progress, the embrace of AI in the accounting domain has become an essential aspect for companies in various sectors. This research investigates the quality of financial reporting and disclosure and the use of AI, examining numerical and non-numerical financial information on transparency and the impact of this technological development. In the construction of the empirical analysis, the variables presented in

Table 1. Description of variables.

Variables	Abbreviation	Measuring	Type of Variable
Financial performance	Fin_perform (FP)	Classical/traditional financial indicators	Dependent variable
Current ratio	CR	Current assets/current liabilities	Dependent variable
Return on assets	ROA	Net profit/total assets	Dependent variable
Return on equity	ROE	Net profit/equity	Dependent variable
Net result	NR	Net income	Dependent variable
The average disclosure degree of accounting estimates	Di_Est	$Di\_Est={\displaystyle \frac{{\sum }_{i=1}^n{x}_i}{n}}$ x = information quantified according to the level of detail of the presentation in the financial reports n = number of quantified accounting estimates	Independent variable
The average disclosure score on changes in accounting estimates	Modif_Est	$Modif\_Est={\displaystyle \frac{{\sum }_{i=1}^n{z}_i}{u}}$ z = information quantified by 1 for the change in accounting estimate u = number of changes in accounting estimates	Independent variable
The average disclosure score on information regarding the use of AI	Di_IA	$Di\_IA={\displaystyle \frac{{\sum }_{i=1}^n{y}_i}{t}}$ y = information quantified by 1 for the AI technologies disclosed on the website t = number of AI technologies used	Dependent variable for testing H1/independent variable for testing H2
The average number of employees	Sal	The average value of the number of employees for the analyzed period	Independent variable
Company size (total assets)	Size (TA)	The total value of assets for the investigated period	Independent variable
The industry in which the company operates	Ind	Industry type	Independent variable
Owner’s equity	Capital_pr.	Owner’s equity = Total assets—Liabilities	Independent variable
ESG performance scores	ESG risk score	Measures the magnitude of risks not managed by the company. Lower scores indicate that the company’s risks are not managed are lower [51].	Independent variable
	ESG exposure	The Sustainalytics Exposure Score measures a company’s exposure to various significant ESG issues, based on industry-specific factors. The higher the score, the higher the company’s exposure to relevant ESG issues [51].	Independent variable
	ESG management	The Sustainalytics ESG management score measures the robustness of a company’s ESG programs, practices, and policies. A higher score indicates better performance in managing ESG risks [51].	Independent variable
	ESG risk ranking Score	According to the Sustainalytics methodology [51], it shows how the company ranks among industry representatives.	Independent variable

were selected.

Several computer programs and various AI tools were used in the statistical analysis of the data. To test the hypotheses, Gretl 2023b and Python vs. 3 were used, with the Matplotlib 3 tool. In addition, other tools available in Python, such as Scikit-learn, StatsModels, or TensorFlow, were used to perform linear regression, logistic regression, and the Random Forest model. It is recommended in accounting research to use logistic regression as a robust econometric method in situations where the dependent variable is binary, because it provides valid results [56,57]. The significant advantage of logistic regression is that it does not require the dependent variable to be normally distributed nor that the residuals be normal, unlike linear regression. This flexibility makes it suitable for accounting data that often does not meet normality assumptions. Its ability to model the probability of an event and identify the influence of various factors makes it a valuable tool for obtaining valid results and meaningful interpretations [58]. As artificial intelligence (AI) becomes a key component of corporate strategy, the disclosure of AI-related information by listed companies is gaining increasing importance. Investors, analysts, and stakeholders are increasingly interested in understanding how AI investments and initiatives translate into business performance. In these circumstances, the Random Forest method is emerging as a powerful tool for testing and identifying the complex correlations between these variables. Disclosure of information on AI, in our case, a complex, qualitative variable, measured by the presence of sections dedicated to AI and investments in advanced technologies on the websites of the sampled companies, and financial performance (measured by ROA, ROE, NetR, and CR indicators) are not necessarily related by a simple linear function. Random Forest, as an ensemble of decision trees, excels in identifying nonlinear patterns and relationships, including interactions between different types of disclosures, presenting a major benefit over linear models [59]. As Hastie et al. [60] argued, the Random Forest method can integrate multiple types of data, both structured and unstructured. Analyzing corporate data, often “noisy” with many potentially irrelevant variables, can be misleading, but Random Forest is less susceptible to overfitting and can more effectively filter out weak predictive variables, focusing on those that have a real impact. This flexibility makes the method suitable for datasets with high dimensionality and potential multicollinearity [61].

The ability to explore the complex and potentially nonlinear relationships between AI disclosure and the financial performance of listed companies, and to handle diverse data types, identify interactions, and measure the importance of predictors, makes this method a useful tool in deeply understanding how AI transparency influences corporate value. Thus, managers, investors, and other stakeholders can gain valuable insights from analyzing the results obtained by applying this method. In statistical hypothesis testing, the Pearson correlation coefficient, which measures the extent of linear association between two continuous variables, the t-test, and the correlation matrix were used to assess the relationships between variables in the data set [62,63,64]. Also, statistical regression was used, which allows the evaluation of relationships between variables, the testing of hypotheses, as well as the estimation and prediction of results [65,66]. It helps to identify predictors and control variables, providing a robust method for investigating the topic. To test the assumptions, a linear regression model was used to investigate the factors influencing AI disclosure (Di_IA), including company size, industry, and average number of employees (Di_IA_i = β0 + β1Sizei + β2Indi + β3Employli + + εi), and four models to assess the effect of AI disclosure (Di_IA) on business performance. Hypothesis H3 suggests that sustainability and corporate governance initiatives, combined with technological innovations such as AI, can lead to competitive advantage and, ultimately, better financial performance. It is tested whether ESG and AI act as complementary forces in generating value for the company. For a limited number of companies, information is known at the level of 2022 between: ESG risk score, ESG exposure, ESG management, and ESG risk ranking score. To test this hypothesis, the regression model could be extended to include both ESG scores and the AI disclosure variable, as well as their potential interactions:

$$\begin{gathered} {{\mathrm{F}\mathrm{i}\mathrm{n}}_{\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{m}}}_{\mathrm{i}}={\mathsf{\beta }}_0{+\mathsf{\beta }}_1{{\mathrm{D}\mathrm{i}}_{\mathrm{I}\mathrm{A}}}_1+{\mathsf{\beta }}_2{{\mathrm{E}\mathrm{S}\mathrm{G}}_{\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}}_{\mathrm{i}}+{\mathsf{\beta }}_3{{\mathrm{E}\mathrm{S}\mathrm{G}}_{\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}}}_{\mathrm{i}}+{\mathsf{\beta }}_4{{\mathrm{E}\mathrm{S}\mathrm{G}}_{\mathrm{M}\mathrm{a}\mathrm{n}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}}}_{\mathrm{i}} \\ +{\mathsf{\beta }}_5{{\mathrm{E}\mathrm{S}\mathrm{G}}_{{\mathrm{R}\mathrm{i}\mathrm{s}\mathrm{k}}_{{\mathrm{R}\mathrm{a}\mathrm{n}\mathrm{k}\mathrm{i}\mathrm{n}\mathrm{g}}_{\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}}}}_{\mathrm{i}}+{\mathsf{\beta }}_6\left({{\mathrm{D}\mathrm{i}}_{\mathrm{I}\mathrm{A}}}_{\mathrm{i}}\times {{\mathrm{E}\mathrm{S}\mathrm{G}}_{\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}}_{\mathrm{i}}\right)+{\mathsf{\epsilon }}_{\mathrm{i}} \end{gathered}$$

(1)

ESG_Score_i, ESG_Exposure_i, ESG_Management_i, and ESG_Risk_Ranking_Score_i are variables that measure the ESG aspects of company i. Di_IAi×ESG_Score_i is an interaction term that shows how ESG scores moderate the impact of IA disclosure on financial performance, and εi is the error term for company i. As applied methods, considering that the sample of companies contains information at the level of 2022, we will apply multiple regression analysis with hypothesis testing on residuals and their correction. The analysis followed several stages: (i) multicollinearity check: testing for multicollinearity between independent variables to verify that they are not so correlated as to distort the regression results; (ii) multiple Regression Estimation using financial performance as the dependent variable and IA disclosure score, company size, and industry as independent variables; (iii) statistical tests: t-test to assess the significance of the coefficients and F-test to assess the overall significance of the model to determine whether the independent variables have a significant impact on the dependent variables; (iv) model diagnosis: models assessed to identify any specification problems, such as heteroscedasticity or autocorrelation, and appropriate corrections (robust standard errors) applied; and (v) residual analysis: checking residual assumptions to identify possible deviations from classical regression assumptions.

The presence of multicollinearity was assessed using the nonparametric Spearman correlation matrix. If the presence of multicollinearity is validated, the variables that induce multicollinearity will be eliminated. Thus, the analysis will include sequential steps, and as the basic model, a simple model was estimated that includes the initial variables, Fin_performi = β0 + β1Di_IAi + β2Sizei + β3Indi + εi. Subsequently, the ESG scores were added one by one to the regression model, testing each one for its impact on financial performance. After each addition, changes in R-squared and F-statistic were checked to determine whether the new variable significantly improved the model. For each newly estimated model, diagnostic tests were performed to check for new multicollinearity issues (by calculating VIF for newly introduced variables) and to ensure that the residuals met the linear regression assumptions.

4. Results and Discussion

4.1. Results

Data analysis shows differential adoption of emerging technologies by industry, with a greater focus on automation and data analytics in some sectors, while others remain less influenced by these technology trends. This could indicate both growth opportunities and challenges in digital transformation for different sectors of the economy. Investigating the use of AI technologies reveals a significant concentration of information related to SAP solutions, indicating widespread adoption of these programs in the corporate environment. Consequently, a growing number of companies resort to implementing software solutions that integrate AI techniques and algorithms. The trend reflects a growing interest in AI technologies’ employment in the business environment, highlighting their importance and relevance in the current context of digital transformation and technological evolution. This evolution can be interpreted as a response to the increasingly complex demands and opportunities of the contemporary business environment, where AI is becoming increasingly integrated and indispensable for optimizing operations and improving organizational performance [67]. Analysis of companies by the average AI techniques disclosure score shows a higher concentration of companies in the manufacturing and communications industries, suggesting more advanced adoption of AI or greater transparency in these areas. Bittnet Systems stands out as the company that utilizes the most AI and reports the most relevant information. The higher degree of AI disclosure may suggest the company’s proactive approach to adopting and implementing AI-based solutions in its operational or strategic activities and may indicate an increased commitment to transparency and communication regarding technological innovations. This observation could provide significant insight into how certain companies are more advanced in adopting and using AI compared to others in the industry.

The analysis of the financial results of listed companies shows a variety of performances in terms of return on assets (ROA) and return on equity (ROE), with some companies having outstanding results that influence the average, although most companies present more moderate values. Current liquidity varies considerably among companies, indicating that some have much more current assets relative to short-term liabilities, indicating a different ability to manage short-term obligations. Net profit also varies, indicating that there are substantial differences between companies’ profitability. AI disclosure is generally low, revealing that not all companies are openly communicating about the adoption of AI technologies or that only a limited number have implemented these technologies in their operational processes. Regarding ESG aspects, the scores show a varied concern for sustainability and governance among companies, with some companies having higher scores than others (

Table 2. Descriptive statistics for 2022.

Panel A
Financial performance indicators
	Mean	Median	Std. Dev.
ROA	0.100021325	0.04513833	0.36161
ROE	0.051465779	0.060043015	0.191651
CR	225.336114	5.704113	962.5243
Net result	4.272653793	2.269164494	9.156484
Aggregate score for disclosure of information on the use of artificial intelligence (AI)	0.121	0	0.204
ESG information
ESG risk score	44.47692308	48	14.67834
ESG exposure	49.98461538	48.9	14.71569
ESG management	22.97692308	23	4.993188
Company size
Large company	1
Medium company	4
Small company	50
Percentage of companies that recorded changes in accounting estimates
1 (changes are present)	43.64%
0 (nonexistent changes)	56.36%
Number of companies in each sector
The extractive industry	3
Manufacturing industry	35
Energy production and supply	2
Constructions	1
Wholesale and retail trade	3
Transportation and storage	4
Hotels and restaurants	4
Information and communications	1
Professional, scientific activities	1
Health and social care	1
Frequency of disclosure of information on the use of AI (%)
BKCHN	1.82
ML	7.27
NLP	3.64
RPA	21.82
IoT	10.91
SAP	27.27

).

The size distribution of the companies in the sample is skewed towards small companies, with only one large company present, indicating that most companies have not yet reached a certain operating scale. Almost half of the companies reported changes in accounting estimates, which could reflect adjustments to regulations, changes in market conditions, or revisions to financial forecasts. The number of companies in each sector indicates a dominance of the manufacturing industry in the studied sample, followed by sectors such as transport and storage, hotels, and restaurants, indicating a sectoral diversity in the sample (Figure 1). The frequency with which companies disclose information about the use of specific AI technologies varies, with a higher percentage of companies communicating about SAP and RPA solutions, which could indicate a focus on automation and efficient data management.

The matrix in

Table 3. Correlation matrix for 2022.

	INDUSTRY ID	Aggregate_Score AI	Company Size	Debts	Provisions	ESG MN	ESG EXPS	ESG RISK_SCORE
INDUSTRY_ID	1	−0.30044	0.196039	−0.11939	−0.25583	−0.51735	−0.30131	0.102332
Aggregate_Score_AI	−0.30044	1	−0.1823	−0.15942	0.091096	0.327376	0.065475	−0.15088
Company size	0.196039	−0.1823	1	−0.7308	−0.812	−0.5684	−0.6902	−0.39247
Debts	−0.11939	−0.15942	−0.7308	1	0.884615	0.510989	0.82967	0.813187
Provisions	−0.25583	0.091096	−0.812	0.884615	1	0.697802	0.917582	0.774725
ESG_MANAGEMENT	−0.51735	0.327376	−0.5684	0.510989	0.697802	1	0.851648	0.401099
ESG_EXPOSURE	−0.30131	0.065475	−0.6902	0.82967	0.917582	0.851648	1	0.763736
ESG_RISK_SCORE	0.102332	−0.15088	−0.39247	0.813187	0.774725	0.401099	0.763736	1

presents the results of the correlation coefficients. Empirical results support the fact that between debts and provisions has a very strong correlation, suggesting that as a company’s real debt increases, so do its provisions. This may reflect prudent accounting management in which companies allocate more provisions as their debt levels increase; debts and ESG management also have a strong correlation, indicating that firms with higher debts tend to have better ESG management scores. This may suggest that firms with higher debts are also those that place a greater emphasis on sustainable practices and social responsibility; provisions and ESG management show a similarly strong correlation, suggesting that companies with stronger ESG management may be more inclined to make adequate provisions to manage their risks; provisions and ESG exposure also show a significant correlation, indicating that firms with higher levels of provisions tend to be more exposed to ESG risks, possibly reflecting greater awareness and preparedness to manage these risks; ESG management and ESG exposure have a very strong correlation with each other, which is expected, as firms that are more exposed to ESG risks will need more effective management of these risks.

Given the limited number of observations in the sample (55) and the presence of multicollinearity in the data induced by the strong correlations above, it was decided to eliminate the variables that can induce multicollinearity (real debts and real provisions) and sequentially test the variables related to ESG scores. The models were estimated using both the HAC (Heteroskedasticity and Autocorrelation Consistent) method and the Huber–White method (also known as “robust standard errors”), which are used to adjust the standard errors of the coefficients in regressions to provide more reliable statistical inferences in the presence of certain types of data problems. The HAC method is applied when the data exhibits both heteroscedasticity (the variability of the residuals changes depending on the level of one or more independent variables) and autocorrelation (residuals correlated over time or other ordering). The Huber–White method adjusts the standard errors of the regression coefficients to allow for heteroscedasticity, without assuming a specific form of the variability of the residuals.

The clustered heatmap (Figure 2) reveals how the numerical indicators in the 2022 dataset interconnect. Two dominant red blocks stand out. Block A (top left) groups together several profitability ratios, gross margin, operating margin, ROE, and ROA. Their deep red squares show pairwise correlations above ≈ 0.85, indicating that these measures rise and fall together. Block B (center-right) connects market indicators, indicating the market values of companies in these lenses. A blue diagonal band running through these block marks strong negative relationships between profitability metrics and cost or leverage variables. Companies with higher margins tend to have lower leverage, which mathematically lowers the covariance. ESG-oriented scores also fall into a smaller cluster on the lower left. Their intra-correlations are moderate (≈ 0.4–0.6), but they have weak links with financial ratios, implying that ESG performance movements are somewhat independent of short-term financial indicators. On the other hand, liquidity indicators form a red mini-cube, moderately positive among themselves but almost uncorrelated with profitability. This separation warns that strong solvency does not automatically translate into superior performance.

The combined visualization of the scatter and bar chart (Figure 3) allows the identification of clusters, for example, of companies adopting more AI technologies, and the interpretation of the left panel leads us to conclude that more intense adoption of AI may accompany stronger ESG risk control. On the other hand, the right panel reveals that only a few companies use more than three AI technologies at the level of the analyzed year. The slightly downward trend shows that a greater involvement of AI in operational processes may coincide with a slightly improved ESG risk, but the difference is wide; therefore, in-depth statistical testing would be considered for more reliable results.

4.2. Discussion

To test H1(a), the Pearson correlation coefficient was applied to assess the linear relationship between the average AI disclosure score (Di_IA) and total asset values (TA). The correlation coefficient is 0.2829, with a p-value of 0.0364, which indicates a weak to moderate positive correlation between the two variables, which is statistically significant, suggesting a linear connection. In terms of testing H1(b), the comparison of Di_IA with industry type involves a categorical variable (industry) and a numerical variable (Di_IA). We will check the distribution of mean scores to decide on the appropriate test method for analyzing differences between types of industries. The distribution of mean AI disclosure scores assessed by the Shapiro–Wilk test indicates that the scores do not follow a normal distribution. Given this lack of normality, for H1(b), it is more appropriate to use a nonparametric test to compare the medians of average scores across different industry categories, such as the Kruskal–Wallis test. The result of the Kruskal–Wallis test for H1(b) indicates a statistical value of 44.7889 and a p-value of 0.3967, showing that there are no statistically significant differences between the medians of the average AI disclosure scores by industry, according to the standard level of significance (p < 0.05). For testing H1(c), which involves comparing Di_IA with the average number of employees, similar to H1(a), the Pearson correlation coefficient was employed to evaluate the connection between the variables. The correlation coefficient takes the value of 0.1648, with a p-value of 0.2294, and indicates a weak positive correlation between the variables, but this relationship is not statistically significant, suggesting that there is no strong association between them. The correlation matrix revealed a strong relationship between total assets and the number of employees and weaker correlations of the average AI disclosure score with the other variables (average accounting estimates disclosure score, average accounting estimates changes score, and equity).

There is a moderate positive correlation between DI_IA and TA (0.28), as well as between DI_IA and Capital_pr. (0.25), which indicates that company size is positively related to the DI_IA indicator, suggesting that larger entities could register higher values of the DI_IA indicator. A lack of direct linear correlation is observed between DI_IA and DI_Est, with a correlation coefficient of almost zero (0.004), as well as between DI_IA and Modif_Esti (−0.007), indicating that these variables do not have a direct linear relationship with the DI_IA indicator. The observed correlations necessitate further analysis to elucidate the exact nature of the relationships and to determine whether there are causative factors or merely associations. For example, the strong relationship between TA and Capital_pr. is intuitive, but their impact on DI_IA may require more detailed exploration to identify the underlying mechanisms. The absence of a strong correlation between DI_IA and some variables, such as DI_Est or Modif_Est, shows that their influence on DI_IA could be mediated by other factors or that the relationships may be nonlinear and not captured by Pearson correlation coefficients.

To test the first regression model, linear regression was first employed. To perform the regression analysis, we need to approach the industry variable as a dummy variable. The mean squared error (MSE) is 0.0931, showing the mean squared difference between the predicted and actual values. A smaller value of MSE indicates a smaller error. The R² coefficient of determination is −0.0348, which measures the degree to which the predictions match the actual data. A negative R² in this context indicates that the model does not fit the data well and fails to accurately represent the variability in the response. This negative R² suggests that linear regression not only performs worse than predicting the mean but also fails to capture any meaningful patterns in the data. Therefore, the results indicate that the linear regression model may not be the most appropriate for this data set, possibly due to misalignment between the selected variables and DI_IA or due to the complexity of the relationships between variables that are not well captured by a simple linear model.

For these reasons, the logistic regression model and the Random Forest model were tested. To apply logistic regression, we need to have a binary dependent variable. Applying logistic regression to the dataset with the dependent variable Di_IA yielded results captured in the confusion matrix. The accuracy of the model is 54.55%, which indicates the percentage of correct predictions out of the total number of predictions. This is relatively modest, illustrating a limited degree of success for the model in discerning between the two categories based on the independent variables used. This result suggests that the model needs improvement by adjusting the independent variables, selecting new, more predictive variables, or exploring alternative modeling methods that better capture the relationships between the variables. The confusion matrix in Figure 4 shows 50 correct predictions for the “Below Median” class, 45 correct predictions for the “Above Median” class, 30 incorrect predictions where the actual “Above Median” values were predicted as “Below Median”, and 25 incorrect predictions where the actual values “Below Median” were predicted as “Above Median”. These results indicate the performance of the grouping model in distinguishing between the two classes: “Below Median” and “Above Median”, based on the Di_IA variable.

After training the Random Forest method on the dataset and making predictions, the confusion matrix shows five correct predictions for the negative class (true negatives), one incorrect prediction for the positive class as the negative class (false negative), five incorrect predictions for the negative as the positive class (false positives) and zero correct predictions for the positive class (true positives). Therefore, the accuracy of the model is about 45.45%. These results indicate a suboptimal performance of the model for binary classification on this dataset, given the relatively low accuracy and complete lack of correct predictions for the positive class. The results reveal that the model has difficulty effectively distinguishing between the two classes based on the available features, possibly due to the lack of relevant information in the features or the need to fine-tune model parameters to improve performance.

Analyzing hypothesis H1, according to which larger companies are more likely to have the resources necessary to implement AI more efficiently and thus obtain greater financial benefits from this transparency, the negative coefficients of medium- and small-sized companies to the company’s financial performance evaluated by ROA or ROE or the net result of the year highlight precisely this aspect. Thus, medium- or small-sized companies are associated with inferior financial performance in certain terms, possibly due to the initial costs of adopting AI that outweigh the short-term benefits. The more negative coefficients for small companies, compared to medium-sized ones, could indicate that smaller companies feel the costs and challenges of AI adoption more acutely, possibly having fewer resources to absorb these initial costs and to effectively implement AI technologies. The improvement in current liquidity for both size categories suggests that, despite overall financial performance challenges, companies may experience increased operational efficiency or better access to capital following the adoption of AI and disclosure. This may reflect a positive market or investor response to AI-related initiatives. The improvement in current liquidity may indicate that, in the medium to long term, companies of all sizes may start to reap financial benefits from adopting AI, particularly in terms of operational efficiency and investor attractiveness. This conclusion suggests that, over time, companies can overcome initial challenges and reap greater financial benefits from transparency in AI usage.

The interpretation of the results for hypothesis H2, which examines the correlation between financial indicators and AI information presented on the website, is described next. In the case of H2(a), the Pearson correlation coefficient records the value 0.1943 and p-value 0.0014, which results in a significant positive correlation between current liquidity (CR) and the AI information disclosed on the website. For H2(b), the value of the Pearson coefficient of −0.0541 and p-value of 0.7518 indicates a very weak negative correlation between ROA and information about AI disclosed on the website. However, the high p-value (0.7518) suggests that this correlation is not statistically significant. But for H2(c), the Pearson coefficient value of 0.1595 and p-value of 0.0579 indicate a moderate positive correlation of ROE with the disclosed AI information. The recorded value (0.0579) is slightly above the standard significance threshold of 0.05, implying that this correlation is not significant at a conventional significance level. Consequently, the results revealed a significant correlation between CR and AI information, indicating that companies with higher liquidity may be more likely to disclose their adoption of AI technologies on their websites. Furthermore, the proposed model was evaluated to examine the link between financial indicators and the average disclosure score for the use of artificial intelligence (Di_IA). The R-squared (R²) value reveals that approximately 7.53% of the variation in the average AI usage score is explained by the variability in CR, ROA, and ROE. R² represents the proportion in Di_IA, described by the models’ independent variables. A low R² indicates that the model does not satisfactorily explain the variance in AI scores. The p-value recorded (0.2578) indicates that the model as a whole is not significant. This suggests that the independent variables (CR, ROA, and ROE) do not contribute significantly to explaining the variation in Di_IA. The high p-value for the F-test and non-significant coefficients suggest that financial indicators CR, ROA, and ROE are not key predictors of AI scores. This means that there is no solid evidence that these financial proxies have a significant impact on the AI score. In conclusion, the model as a whole does not provide a significant explanation for the variation in the AI use score.

Statistical testing by logistic regression was followed using the Python vs. 3 program. A logistic regression analysis was performed with Machine Learning (ML) as the dependent variable and CR as the independent variable. ML in this context is treated as a binary variable, indicating whether or not a company uses ML technologies (where 1 indicates the use and 0 indicates absence). It was used as a CR as a predictor to see if there is a link between a company’s liquidity level and the use of ML technology. The logistic regression results show the coefficients of the model, constant (Intercept) −3.0685, CR 0.0777, and the p-value for CR is 0.115, which is higher than the usual threshold of 0.05. This suggests that there is no statistically significant association between CR and the use of ML technologies in this dataset. The value of the pseudo-R-squared at 0.1557 denotes a moderate fit of the model to the dataset. Interpretation of the coefficient for CR implies that for each unit increase in current liquidity, the odds of a company using ML increase by 0.0777, although this association is not statistically significant. The next indicator for which logistic regression was tested is RPA. The results reveal the constant (Intercept) −1.5066, CR 0.0491, and the p-value for CR registers the value 0.230, higher than the conventional level of 0.05, indicating no statistically significant association between CR and implementation of RPA technology in the companies in this data set, and the pseudo-R-square value (0.03697) suggests that the model does not explain much of the variation in RPA use. The interpretation of the coefficient for CR is that an increase in current liquidity is associated with a modest increase in the odds of a company using RPA, although this association is not statistically significant. Furthermore, the correlation between CR and SAP was also analyzed, as this is the second technology often used by companies. The results show that the constant coefficients (Intercept) record the value −1.1483, CR 0.0371, and the p-value for CR is 0.302, which is above the standard threshold of 0.05. This implies no statistically significant association between CR and SAP technology used in this data set. The pseudo-R-squared value is 0.02085, which is quite small and indicates that the model explains a very small portion of the variation in SAP technology adoption. Interpretation of the coefficient for CR implies that an increase in current liquidity is associated with a slight increase in the odds of a company using SAP, although again, this association is not statistically significant.

In conclusion, logistic regression models for both ML, RPA, and SAP showed no significant association with CR in this dataset. This does not rule out the possibility of a relationship but suggests that if such a relationship exists, it is not strongly captured by the current data or may require analysis that includes multiple explanatory variables. The logistic regression analysis for SAP, using CR, ROA, ROE, average number of employees, and turnover as explanatory variables, involved transforming the NAIC code, a categorical variable, into several dummy variables for use in the model. The distribution of SAP usage (Figure 5) indicates an imbalance in the dataset, with 40 companies not using the SAP system, while 15 companies do. This imbalance can affect the logistic regression model’s performance, potentially leading to a biased model that favors the majority class.

To address this imbalance, several strategies can be employed, such as class rebalancing, which involves using resampling techniques to balance the class distribution and adjusting the weights of classes in the logistic regression model, thereby allowing the model to give more weight to the minority class. Correlations between explanatory variables and SAP use are also examined to see if there is any significant relationship that could influence the model. Subsequently, the logistic regression analysis was redone, taking these considerations into account. The correlation matrix does not provide a clear picture of the relationships between the explanatory variables and the use of SAP. However, examining the values of individual correlations reveals that positive correlations exist between SAP usage and the average number of employees, net turnover, and net profit. However, the correlations are relatively weak, indicating that none of these variables has a strong influence on a company’s likelihood of using SAP. Given these observations and the imbalance between classes, the logistic regression model was rebuilt, adjusting the weights of the classes in the model to attempt to compensate for the imbalance. The logistic regression model’s results concerning the classification ratio indicate that the accuracy for class “0” (non-SAP) remains 0%, and for class “1” (SAP), it is 29%. Recall for class “1” (SAP) is 100%, meaning that the model identified all SAP instances but failed to correctly identify non-SAP instances. The F1 score for class “1” is 45%, and the overall accuracy is 29%. The confusion matrix reveals that all 12 non-SAP instances were misclassified as SAP, and the five SAP instances were correctly classified. In conclusion, these results indicate that even after adjusting the class weights, the model fails to correctly classify non-SAP instances. This suggests that a different approach to data modeling or a more thorough examination of the data may be required. Thus, possible directions would involve exploring other statistical or Machine Learning models that may be more appropriate for this type of data, revising and possibly expanding the set of explanatory variables to improve the model’s ability to differentiate between companies that use and those that do not use SAP. We chose to continue the analysis with Random Forest, which is a robust model that performs well on datasets with class imbalance. Random Forest constructs various decision trees to improve accuracy and control overfitting according to Lei et al. [68]. The results show that the accuracy for class “0” (non-SAP) is 69%, which is a significant improvement over the logistic regression model. Class “1” (SAP) has 0% accuracy, indicating that the model failed to correctly classify any SAP instance. Recall for class “0” is 92%, while for class “1”, it is 0%. The overall accuracy of the model is 65%. The confusion matrix presented in Figure 6 shows that 11 of the 12 non-SAP instances were correctly classified, 1 non-SAP instance was misclassified as SAP, and all five SAP instances were misclassified as non-SAP.

Therefore, Random Forest is much more efficient in classifying correctly non-SAP instances but fails to correctly identify SAP instances. This may be due to the small number of SAP instances in the dataset, which makes it difficult for the model to learn the relevant features of this class. The empirical results showed that hypothesis H2 is partially validated, indicating a positive relationship between transparency of AI use and current liquidity, suggesting that companies that disclose more information about AI use may have higher current liquidity and also a statistically significant negative relationship between the AI information disclosure score and the net result of the exercise, indicating that greater transparency regarding AI is associated with a lower net result, contrary to the hypothesis of a positive correlation (

Table 4. Empirical results of the relationship between financial performance and artificial intelligence disclosure.

Models	M1 ROA	M2 ROE	M3 CR	M4 NR (mil.lei)	M5 CR	M6 CR	M7 NR .(mil.lei)	M8 NR .(mil.lei)	M9 NR (mil.lei)	M10 NR mil.lei)
Medium company	−0.078 **	−0.041	27.90 *	−5602.96 ***	41.58 ***	28.68 *	−5460.65 ***	−5530.44 ***	−5618.05 ***	−5857.26 ***
Small company	−0.147 ***	−0.136 **	42.11 **	−6485.49 ***	47.58 ***	42.27 **	−6455.42 ***	−6385.64 ***	−6475.43 ***	−6944.06 ***
Industry type (ref = health and social care)
Extractive industry	0.06 ***	0.079	40.28 **	393.34	23.43	36.32 **	337.22	407.01	494.50 *
Manufacturing industry	0.07 ***	0.07 **	6.83 **	−119.13 *	1.79 ***	0.758	24.79 **	−83.53	42.505 ***
Energy production and supply	0.049	0.209 ***	16.57	466.88 *	6.127	9.91	583.13 **	533.30 *	651.83 **
Constructions	−0.017	0.133 **	8.30 **	−132.93 *	1.958 ***	-	39.44 ***	−80.17	39.44 ***
Wholesale and retail trade	0.046	0.067 ***	1.59	−27.48	−7.96	−4.32	1.249	−78.49	119.51
Transportation and storage	−0.011	0.117 *	9.59 *	−273.15 *	2.778	2.84	−171.93	−226.75	−93.23
Hotels and restaurants	−0.05	0.072	12.98 **	−173.14 *	6.637 *	6.63 *	−0.768	−120.38 *	−0.768
Information and communications	0.125	−0.139	−13.13 *	342.77 **	−25.94 *	−7.73 *	579.88*	−1.97 ***	175.42 **
Professional, scientific activities	−0.028	0.029	4.54 **	−72.51 *	1.369 ***	1.369 ***	-	13.67 ***	-
AI aggregated disclosure score	−0.185	0.209	19.05 *	−517.12 **	25.50 *					−428.87 **
AI aggregated disclosure score * information and communications										363.87 *
SAP								−119.61 *
RPA						7.29 *			−177.40 **
NLP							−581.86 *
Constant	0.212	0.078	−47.22 **	6660.25 ***	−46.34 ***	−41.03 **	6457.80 ***	6507.64 ***	6477.81 ***	7009.47 ***
Observations	55	55	55	55	55	55	55	55	55	55
S.E. of Reg.	0.40	0.204	6.73	146.80	6.01	6.67	148.36	157.97	149.01	200.92
F-test	6.89 **	4.45 ***	4.82 ***	189.94 ***	6.93 ***	4.95 ***	185.90 ***	163.54 ***	184.24 ***	297.30 ***
R2 adj.	19.3	14.05	0.459	0.976	0.568	0.467	0.976	0.973	0.976	0.956
Jarque–Bera test	5202.82 (0.00)	1583.71 (0.00)	63.48 (0.00)	4.571 (0.101)	27.47 (0.00)	63.81 (0.00)	135.79 (0.00)	20.43 (0.00)	5.06 (0.079)	115.51 (0.00)
Breusch–Godfrey serial correlation LM Test: Null hypothesis: no serial correlation at up to 1 lag	0.86 (0.356)	3.52 (0.06)	2.93 (0.09)	1.162 (0.28)	2.171 (0.14)	2.67 (0.11)	2.72 (0.106)	3.25 (0.078)	0.92 (0.34)	1.95 (0.168)
White heteroskedasticity test	0.057 (1.00)	0.087 (1.00)	84.13 (0.00)	16.11 (0.00)	37.40 (0.00)	91.81 (0.00)	7.81 (0.00)	25.17 (0.00)	23.77 (0.00)	105.73 (0.00)

Note: ***, **, * mean statistically significant at 1%, 5% and 10%; () represents the probability. Source: own work.

). The results could not prove the existence of a statistically significant relationship between ROA or ROE and the degree of disclosure of information regarding AI. Thus, companies that are more transparent about their use of AI could do so as part of a marketing or branding strategy, suggesting innovation and technological advancement. This can attract investment or improve access to financing, reflected in better current liquidity. At the same time, significant investments in AI technology can lead to large upfront expenses, including research and development, the acquisition of new technology, and employee training, which could temporarily reduce the bottom line for the year. These investments, while they may reduce short-term profit, are often necessary for long-term growth and innovation.

Transparency in AI usage can improve the company’s market perception, thereby increasing its attractiveness to investors and potentially enhancing access to capital, which is reflected in improved liquidity. However, if investments in AI do not yet generate a quick enough return or are perceived as risky, this could negatively impact short-term profits, ultimately affecting the accounting net result. Companies in the early stages of AI adoption may experience higher costs that outweigh the short-term benefits, which can explain the lower bottom line. As technologies and their applications mature, the impact on profitability is expected to become increasingly positive. Increased liquidity suggests that the market is responding positively to the long-term potential of AI adoption, even if the benefits are not yet reflected in the company’s net income. Transparency in the use of AI can indicate a commitment to innovation and continuous improvement, which can lead to long-term benefits, even if the company incurs higher costs in the short term. Investors and stakeholders may value this commitment, which explains the improved liquidity, while the positive effects on the bottom line may only become apparent in the longer term. The impact of transparency in the use of AI on financial performance can vary significantly across economic sectors, depending on the degree to which AI technologies are essential to the specific activities of that sector. In some sectors, AI adoption is a vital strategic necessity for maintaining competitiveness, justifying the initial investments and explanations for the increased liquidity and short-term impact on net income.

To understand the impact of specific artificial intelligence (AI) technologies, such as SAP (Software Application Programming), RPA (Robotic Process Automation), and NLP (Natural Language Processing), on financial performance indicators, regression coefficients will be analyzed; the impact on current liquidity or net result, a negative impact (−119.61 *) on financial performance, could indicate high initial implementation and adaptation costs. These initial investments may temporarily affect profitability but are often justified by the long-term benefits in efficiency and scalability. RPA (Robotic Process Automation). Impact on current liquidity: 7.29 * (significant at the 10% level), suggesting that RPA implementation has a positive impact on current liquidity, likely due to the operational efficiency and short-term cost reductions enabled by process automation. Impact on net income: −177.40 ** (significant at the 5% level), which may indicate that although RPA improves efficiency and can increase liquidity by reducing operational expenses, the initial implementation or integration costs may negatively impact short-term profits. NLP (Natural Language Processing). Impact on net income: −581.86 * (significant at the 10% level), suggesting that investments in NLP technologies may have a negative influence on accounting results in the short term. This may reflect the significant costs associated with developing or acquiring advanced NLP solutions and integrating them into business operations, which outweigh the immediate benefits in terms of profit.

Implementing AI technologies often requires substantial upfront investments in hardware, software, and employee training. These costs can negatively impact short-term profits, even though they promise long-term efficiencies and cost savings, reflected in improved current liquidity. Adopting AI technologies may involve initial challenges related to integration with existing systems and adapting business processes, which may lead to unexpected costs or temporary disruptions. As companies overcome initial hurdles and begin to fully exploit the capabilities of AI technologies, increased operational efficiency can lead to cost reductions, improved cash flow, and ultimately better liquidity. Investments in cutting-edge technologies, such as AI, can enhance a company’s market perception and attractiveness to investors while also opening up new business opportunities, even if the direct financial benefits are not immediate. In conclusion, while the adoption of AI technologies may have a short-term negative impact on the bottom line due to initial implementation and integration costs, the long-term benefits in operational efficiency and innovation can improve a company’s current liquidity and overall financial position.

Hypothesis H3 focused on ESG performance moderated by AI, which states that there is an interaction between ESG scores and AI disclosure in terms of the impact on companies’ financial performance. Companies with better ESG scores and AI disclosure on their websites are likely to have superior financial performance compared to those with lower ESG scores or those that do not disclose AI. This hypothesis is based on the assertion that sustainability and corporate governance initiatives, combined with technological innovations such as AI, can lead to competitive advantage and, ultimately, better financial performance. It tests whether ESG and AI act as complementary forces in generating value for the company. The analysis aimed to estimate unifactorial regression models in which the complementarity of sustainability and corporate governance initiatives, combined with technological innovations such as AI, was assessed, which can lead to better financial performance. The estimates targeted three proxies for the ESG components, namely ESG risk score, ESG exposure, and ESG management, at the horizon of 2022. The models could only be specified in unifactorial form, given the data limitations (there were only 13 companies that had ESG records), and the results should be viewed with due caution, given these drastic limitations on the data. Starting from the more general formalization of the hypothesis, which included both ESG scores and the IA disclosure variable, as well as their potential interactions, the models were narrowed down to include only the interactions between DI_IAi and the ESG proxies as follows:

Model for ESG score:

$${{\mathrm{F}\mathrm{i}\mathrm{n}}_{\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{m}}}_{\mathrm{i}}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1\left({{\mathrm{D}\mathrm{i}}_{\mathrm{I}\mathrm{A}}}_{\mathrm{i}}\times {{\mathrm{E}\mathrm{S}\mathrm{G}}_{\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}}_{\mathrm{i}}\right)+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(2)

Model for ESG exposure:

$${\mathrm{F}\mathrm{i}\mathrm{n}\_\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{m}}_{\mathrm{i}}={\mathsf{\beta }}_0+{\mathsf{\beta }}_2\left({\mathrm{D}\mathrm{i}\_\mathrm{I}\mathrm{A}}_{\mathrm{i}}\times {\mathrm{E}\mathrm{S}\mathrm{G}\_\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}}_{\mathrm{i}}\right)+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(3)

Model for ESG management:

$${\mathrm{F}\mathrm{i}\mathrm{n}\_\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{m}}_{\mathrm{i}}={\mathsf{\beta }}_0+{\mathsf{\beta }}_3\left({\mathrm{D}\mathrm{i}\_\mathrm{I}\mathrm{A}}_{\mathrm{i}}\times {\mathrm{E}\mathrm{S}\mathrm{G}\_\mathrm{M}\mathrm{a}\mathrm{n}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}}_{\mathrm{i}}\right)+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(4)

Di_IAi×ESG_Score_i is an interaction term showing how ESG scores moderate the impact of IA disclosure on financial performance, and εi is the error term for company i. Considering that financial performance was alternatively evaluated for four proxies (ROA, ROE, CR, and NR for the year), twelve models were run, with three alternative models being run for each performance proxy for one ESG proxy (ESG score, ESG exposure, and ESG management). The empirical results revealed a significant impact at the 10% significance level only in the case of a single model, namely the impact of the interaction between ESG exposure and aggregate AI disclosure score on the financial performance of some companies (

Table 5. Empirical results of the interactions between ESGs and AI technologies on the net result.

Variables	Coefficient	Std. Error	t-Statistic	Prob.
C	−330.80	563.97	−0.60	0.56
ESG_EXPOSURE * AGGREGATE SCORE_AI	116.50	67.84	1.72	0.10
R-squared	0.39	Mean dependent var		839.60
Adjusted R-squared	0.33	S. D. dependent var		1898.47
S.E. of regression	1554.14	Akaike info criterion		17.68
Sum squared resid	26,568,929.75	Schwarz criterion		17.76
Log likelihood	−112.89	Hannan–Quinn criter.		17.66
F-statistic	6.91	Durbin–Watson stat		1.65
Prob (F-statistic)	0.02	Wald F-statistic		2.95
Prob (Wald F-statistic)	0.10

* Source: own work.

). The model was estimated assuming Huber–White corrected errors to deal with heteroskedasticity.

The interaction term is close to statistical significance, proving that there may be a relationship between ESG exposure and companies’ financial performance when considering AI disclosure. However, because the p-value is exactly 0.10, this result could only be considered as having a trend, rather than being firmly significant, in a more conservative analysis. Although the model explains a moderate proportion of the variation in financial performance, a fairly large percentage of the variation remains unexplained by the model. The model as a whole is significant, as indicated by the F-test and its associated p-value, which suggests that the variables included collectively have a significant impact on financial performance. It is important to note that while this model may reveal trends, its interpretation should be performed with caution, and it would be useful to compare it with other models and statistical tests to confirm these results.

Table 6. Statistical hypothesis testing summary.

Hypotheses	Statistical Outcomes
H1(a). There is a significant positive correlation between the average disclosure score of AI technologies and company size	Validated
H1(b). There is a significant positive correlation between the average AI technology disclosure score and industry typology	Not validated
H1(c). There is a significant positive correlation between the average disclosure score of AI technologies and the average number of employees	Not validated
H2(a). There is a significant positive correlation between the current liquidity (CR) indicator and AI disclosure	Validated
H2(b). There is a significant positive correlation between the return on assets (ROA) indicator and AI disclosure	Not validated
H2(c). There is a significant positive correlation between the return on equity (ROE) indicator and AI disclosure	Not validated
H3. There is an interaction between ESG performance scores and AI disclosure in terms of the impact on companies’ financial performance	Partially validated

exhibits a summary of the results obtained from the statistical testing of the hypotheses

The practical implications of key findings mainly lie in the need to integrate artificial intelligence into the operational corporate strategy and increase the transparency of information on advanced and intelligent technologies adopted in published annual reports, as well as through other communication channels. These implications target managers who can find other solutions to improve external reporting and transparency to increase the relevance and credibility of information but also regulators to review reporting and communication requirements adapted to the practical capacity of companies to meet them.

5. Conclusions

This study’s results revealed that company size may influence transparency regarding AI use; however, the industry is not a strong predictor of AI disclosure policy, which may imply either uniformity of disclosure practices across sectors or a possible lack of industry-specific regulations to guide these practices. Also, workforce size is not a major determinant of AI disclosure. The results of the analysis also highlighted different aspects depending on the specific financial indicators considered. Thus, the results of testing the correlation between current liquidity (CR) and AI information presented on the website indicate that companies with better current liquidity tend to be more active in adopting and disclosing AI technologies, suggesting that financial resources can facilitate investment in innovative technologies. Following the analysis, both the quantity and the quality of the AI information disclosed influence the quality of financial reporting and also the performance of listed companies. Contrary to expectations based on previous research [25], the results showed a very weak, albeit statistically insignificant, negative correlation between ROA and AI information. This suggests that, in this data set, return on assets is not a clear determinant of the degree to which AI information is disclosed on company websites. There is a tendency for companies with better financial performance, as measured by ROE, to be associated with a better presentation of AI technologies, although this association is not strong enough to be considered statistically significant. The results are in line with those of Babina et al. [36] and Przegalinska et al. [37], although there are also differences that are likely due to the research context. Meanwhile, the interplay analysis leads to the conclusion that a relationship between ESG exposure and corporate business performance can be demonstrated when considering AI disclosure, although more in-depth and niche analyses would be necessary for robust results. The results are in line with those highlighted by Zhang [49] and Tian and Shi [50].

Consequently, the results reveal that although there are indications of a relationship between financial performance and AI disclosure, this relationship varies by specific financial indicators and is not always statistically significant. The findings highlighted the need for a more nuanced and multidimensional approach to understanding the relationship between financial performance and AI adoption in corporate practices in depth and detail. Managerial implications relate to improving communication skills regarding the disclosure of AI information in the annual reports. Key limits of the work are found in the small data size and the setup of the statistical analysis models, aspects that can be reviewed in future research studies. The overall conclusion that emerges highlights the need for a holistic approach to researching the impact of AI in the corporate environment, encompassing both financial and sustainability factors. Research findings lead us to the conclusion that we should further investigate this topic, aiming to discover the underlying mechanisms that influence the adoption and implementation of AI technologies, as well as the way companies communicate about these technologies.