Introduction

With the continuous development of the global economy and the acceleration of internationalization, the position of corporate compliance management in modern corporate governance has become increasingly important. Compliance management not only relates to the legal and operational risks of enterprises, but also directly affects their reputation and market competitiveness. In recent years, with the rapid development of information technology, especially the application of big data and artificial intelligence technology, enterprise compliance management is facing unprecedented opportunities and challenges1. Third-party supervision, as an independent and professional compliance management tool, has played an important role in improving compliance management level and preventing compliance risks. However, with the increasingly complex regulatory environment and diversified business operations of enterprises, how to improve the efficiency and accuracy of third-party supervision has become an important issue facing enterprises and regulatory agencies2,3.

In practical operation, third-party supervision workflows often face a series of problems. Firstly, there is the issue of data quality. Due to the diverse sources and inconsistent formats of data, data cleaning and standardization have become time-consuming and labor-intensive tasks, directly affecting subsequent analysis and decision-making. Secondly, existing regulatory methods and tools rely heavily on manual operations, making it difficult to meet the needs of large-scale data analysis and real-time monitoring. Furthermore, the difficulty in adjusting algorithm parameters greatly limits the improvement of regulatory efficiency. Many traditional algorithms often require a large amount of parameter adjustment and optimization when dealing with complex and changing regulatory tasks, which not only increases workload but also reduces the practicality and flexibility of the algorithm.

Based on this, this article applies the artificial intelligence ACO algorithm to optimize the compliance third-party supervision process of the involved enterprises. The basic principles and application advantages of ACO are introduced, and its application solutions are described in conjunction with specific requirements, including problem modeling, parameter initialization, path selection, pheromone updates, dynamic risk assessment and adjustment, resource allocation optimization, and other aspects31. The effectiveness of algorithm optimization is evaluated through simulation experiments, focusing on indicators such as data quality, model performance, scheme effectiveness, and supervision effectiveness. By discussing the experimental results, the advantages and disadvantages of ACO algorithm in optimizing compliance supervision processes are summarized, and improvement suggestions are proposed to provide efficient and practical optimization solutions for the involved enterprises.

Related work

In the study of the third-party supervision workflow for compliance of the involved enterprises, researchers face multiple challenges. Low data quality is a major issue. Due to the fact that the data of the companies involved in the case often involves multiple different sources, issues such as inconsistent data formats, missing data, and noisy data make data cleaning and preprocessing complicated. In addition, adjusting algorithm parameters is also a major challenge. Traditional optimization algorithms often require a significant amount of computing resources and time to adjust parameters when processing large-scale, high-dimensional data4. Scholars in different fields have utilized advanced optimization algorithms to calculate and adjust parameters for large-scale and high-dimensional data (Shen et al.,2023; Hafiz & Saeed6,7,. For example, in order to enhance the intelligence level of digital financial data processing, compliance risk review, and prevention and control processes, scholars such as Giacalone et al.8 utilized technologies such as big data and blockchain artificial intelligence to form a compliance model of “prevention-screening-response” chain, further improving third-party supervision channels; to further investigate the emerging cybersecurity risks faced by elderly care organizations in an increasingly digital environment, scholars such as Abrahams et al.9 investigated the role of third-party vendors in applying potential vulnerabilities, providing comprehensive and actionable strategies for elderly care organizations. However, these issues have not been effectively addressed in the research on the compliance third-party supervision workflow of the involved enterprises.

Many scholars have proposed different methods to optimize the third-party supervision workflow in response to the above issues. To maximize the safety and effectiveness of artificial intelligence and machine learning in healthcare, Sara et al.10 provided effective recommendations for third-party supervision agencies through a shift in perspective, from a product perspective to a system perspective. In response to the problems in daily safety inspections of university laboratories, scholars such as Sun et al.11provided a detailed introduction to the mechanism construction, qualification certification, and workflow of third-party security checks, effectively improving the level of safety inspections in university laboratories. Scholars have studied and utilized machine learning techniques12,13,14and blockchain technology15,16,17 to improve data quality, which has to some extent solved the problems of data processing and algorithm optimization.

Although some progress has been made in the above research, there are still many problems that have not been effectively solved. The existing methods still lack adaptability in dealing with dynamically changing regulatory demands. In large-scale data environments, the computational efficiency and real-time performance of algorithms are still not high. In addition, most existing methods rely on experience18,19,20. In these aspects, ACO has unique advantages21,22,23. This article utilizes ACO to flexibly adjust workflow and optimize implementation and supervision strategies in a dynamic environment, in order to make up for the shortcomings of existing research and further enhance the effectiveness of third-party supervision workflow for compliance of the involved enterprises.

Improvement of Third-party supervision workflow for compliance of involved enterprises based on artificial intelligence ACO

Construction of ACO

ACO is a heuristic algorithm that simulates the foraging behavior of ants in nature, used to solve the combinatorial optimization problem in the compliance third-party supervision workflow of involved enterprises. The core idea of ACO is to guide ants to find the optimal path through the accumulation and volatilization of pheromones24,25,26. The ACO structure is shown in Fig. 1:

Fig. 1
figure 1

ACO structure diagram.

Full size image

From the ACO structure diagram in Fig. 1, it can be seen that the structure of ACO includes four modules: initialization, solution construction, local search, and updating pheromone, and achieves optimization through continuous iteration27,28,29.

In the initialization phase, the number of ants is generally set to N; the pheromone volatilization coefficient is set to α; the initial value of pheromone is set to \(\:{\pi\:}_{0}\); the number of iterations is set to \(\:{I}_{max}\). The pheromone matrix is initialized as \(\:{{\Sigma\:}}_{ij}\), usually a small positive number.

In the ant colony optimization process, each ant chooses the next step based on the pheromone concentration and heuristic information. The probability of choosing the next step is determined by the pheromone strength and heuristic information as follows:

$$\:{P}_{ij}^{k}\left(t\right)=\frac{{\left[{\pi\:}_{ij}\right(t\left)\right]}^{\beta\:}{\cdot\:\left[{\tau\:}_{ij}\right]}^{\gamma\:}}{{\sum\:}_{m\in\:{\mathcal{D}}_{i}^{n}}{\left[{\pi\:}_{il}\right(t\left)\right]}^{\beta\:}{\cdot\:\left[{\tau\:}_{il}\right]}^{\gamma\:}}$$
(1)

In Formula (1), \(\:{\pi\:}_{ij}\left(t\right)\) represents the concentration of pheromones on edge (i, j), which is an pheromone matrix; \(\:{\tau\:}_{ij}\) is heuristic information; \(\:\beta\:\) and \(\:\gamma\:\) are parameters that control the relative importance of pheromone and heuristic information; \(\:{\mathcal{D}}_{i}^{n}\) is the next set of positions that ant m can choose from at position i.

In the local search stage, the constructed solution is locally optimized and then enters the updating pheromone stage. The updated pheromone matrix is \(\:{\pi\:}_{ij}(t+1)\):

$$\:{\pi\:}_{ij}\left(t+1\right)=\left(1-\alpha\:\right)\bullet\:{\pi\:}_{ij}\left(t\right)+\partial\:{\pi\:}_{ij}\left(t\right)$$
(2)

Among them, \(\:\partial\:{\pi\:}_{ij}\left(t\right)\) is the amount of pheromone added on edge (i, j) at iteration time t.

$$\:\partial\:{\pi\:}_{ij}\left(t\right)={\sum\:}_{m=1}^{N}\partial\:{\pi\:}_{ij}^{m}\left(t\right)$$
(3)

In ACO, the update of pheromones32; Huo Feng et al., 2024; Chari et al.34, is completed by the amount of pheromones left by each ant on the path. The amount of pheromones \(\:\partial\:{\pi\:}_{ij}^{m}\left(t\right)\) left by each ant m on the edge (i, j) is:

$$\:\partial\:{\pi\:}_{ij}^{m}\left(t\right)=\frac{{\Omega\:}}{{\psi\:}_{m}}$$
(4)

Among them, \(\:{\Omega\:}\) is a constant related to the scale of the problem, and \(\:{\psi\:}_{m}\) represents the total length or total cost traveled by ant m.

Compliance third-party supervision workflow of the involved enterprise

The third-party supervision workflow for compliance of the involved enterprises aims to ensure that their experience activities comply with laws, regulations, and internal policies, thereby reducing legal risks and violations. The general third-party supervision workflow for compliance of involved enterprises includes preliminary assessment and investigation stage, data collection and analysis stage, compliance risk assessment stage, formulation of compliance plan stage, implementation and supervision stage, and reporting and feedback stage30.

In the preliminary assessment and investigation stage, basic information and relevant documents of the enterprise are collected, and feedback information from management and employees is obtained through face-to-face interviews or questionnaire surveys to preliminarily identify compliance risk points of the enterprise. In the data collection and analysis stage, financial data, transaction records, employee information, etc., are collected, and the collected dataset is subjected to data cleaning and feature selection. In the compliance risk assessment phase, the compliance risk assessment model is used to identify potential compliance risk points in the operation of the enterprise and evaluate existing compliance control measures. In the stage of formulating compliance plans, targeted compliance improvement plans are developed based on the results of risk assessments, and improvement measures and plans are proposed. In the implementation and supervision phase, a supervision mechanism is established; compliance plans are implemented; the implementation of compliance plans is regularly inspected and evaluated. In the reporting and feedback phase, compliance reports are generated; evaluation results and improvement measures are summarized; compliance status is reported to management and relevant departments.

ACO optimization of third-party supervision workflow for compliance of involved enterprises

In the traditional third-party supervision workflow for compliance of involved enterprises, it is often difficult to select effective data features during the data collection and analysis stage, which often leads to problems such as high data dimensions and weak correlation of data features. There are also cases of inaccurate model results during the compliance risk assessment phase. In addition, during the implementation and supervision phase, as well as the reporting and feedback phase, the traditional third-party supervision workflow for compliance of the involved enterprises is often difficult to develop a reasonable implementation and supervision plan. The use of ACO can effectively solve these problems, and ACO can penetrate into various stages of the compliance third-party supervision workflow of the involved enterprises. Figure 2 shows the workflow diagram of third-party supervision compliance for the involved enterprise after ACO optimization:

Fig. 2
figure 2

Workflow diagram of third-party supervision compliance for the involved enterprise after ACO optimization.

Full size image

Figure 2 clearly illustrates how ACO has penetrated into the compliance third-party supervision workflow of the involved enterprises. Based on traditional workflows, the application of ACO has almost optimized all aspects of traditional workflows, reflecting the importance of different information through the definition of heuristic information matrices in the initial evaluation and investigation stages. In the data collection and analysis phase, the heuristic information matrix (Tehrani et al., 2024; Tema et al.,2024) is defined as the correlation between features, and ACO is used to select the optimal data cleaning and feature selection strategy. ACO is utilized to optimize the parameters of the risk assessment model during the compliance risk assessment phase; ACO is utilized to optimize the combination of compliance plans and monitoring strategies and plans during the development, implementation, and monitoring phases. During the reporting and feedback phase, automated reports are generated.

Selection of evaluation Indicator

In the compliance third-party supervision workflow of the involved enterprises, the indicators selected in this article include four aspects, namely data quality, model performance, scheme effectiveness, and supervision effectiveness.

The evaluation indicators for data quality include data accuracy indicators and data integrity indicators. Mean Square Error (MSE) is used to evaluate the accuracy of cleaned data. The formula for data integrity indicators is:

$$\:{\Gamma\:}=\frac{\sum\:_{j=1}^{q}\xi\:({x}_{ij}=\text{N}\text{A})}{p\times\:q}$$
(5)

Among them, \(\:\xi\:({x}_{ij}=\text{N}\text{A})\) is the indicator function. When \(\:{x}_{ij}\) is a missing value, that is, \(\:{x}_{ij}=\text{N}\text{A}\), \(\:\xi\:\left({x}_{ij}=\text{N}\text{A}\right)=1\), otherwise it is 0; p and q respectively refer to the number of samples and the number of features.

The evaluation metrics for model performance include accuracy, recall, F1 score, and AUC-ROC (Area Under the Curve–Receiver Operating Characteristic) (Sayed et al.,2024).

The evaluation indicators for the effectiveness of the plan include: Cost-Effectiveness Ratio (CER) and Net Present Value (NPV). CER is used to evaluate the cost and benefit of a compliant plan and is the ratio of total cost to total benefit; NPV is used to evaluate the long-term benefits and costs of implementing and monitoring programs, and the formula is:

$$\:\text{N}\text{P}\text{V}=\sum\:_{y=0}^{Y}\frac{{\eta\:}_{y}-{C}_{y}}{(1+r{)}^{t}}$$
(6)

Among them, \(\:{\eta\:}_{y}\) is the benefit in year y; \(\:{C}_{y}\) is the cost for year y; r is the discount rate; Y is the time period of the plan.

The evaluation indicators for the effectiveness of supervision include Supervision Coverage Rate (SCR) and Compliance Rate Change (CRC). SCR mainly evaluates the coverage of supervision measures, which represents the proportion of the total number of stations that are actually supervised in all areas or departments that need to be supervised; CRC is mainly used to measure the effectiveness of supervisory measures; compliance rate represents the proportion of items that comply with regulations among all inspection items.

Simulation cases

Experimental preparation

This experiment uses simulated real datasets to verify the effectiveness of ACO algorithm, including four types of data: corporate financial data, corporate operational data, legal and regulatory data, and historical violation record data. Enterprise financial data package includes balance sheet, income statement, and advanced flow, etc.; business data covers production volume, sales volume, inventory level, etc.; legal and regulatory data involves relevant legal and regulatory texts and compliance requirements; historical violation record data includes the company’s past violation records and processing results. The four types of data each provide different perspectives and information, which can better facilitate subsequent experiments. The collected partial data is shown in Table 1:

Table 1 Part of experimental data content.
Full size table

Table 1 shows some of the data collected in this simulation experiment. Specifically, the simulation experiment dataset collects financial data, operational data, compliance data, and historical violation record data from 50 companies, covering the years 2015–2020. The data includes total assets, total liabilities, revenue, production volume, sales volume, inventory level, compliance status, and violation records. The experiment divides the collected dataset into training and test sets in an 8:2 ratio.

The simulation experiment is conducted on a high-performance server, with hardware configuration including a CPU of 8 or more cores, a memory core of 32GB or more, and storage space of 1 TB or more. The software environment adopts Linux operating system, and programming language is Python. Algorithm implementation and optimization are carried out using NumPy, Pandas, and Scikit-learn, and MySQL database is used for data storage and processing in data management.

The main purpose of this simulation experiment is to verify the effectiveness of the artificial intelligence ACO algorithm in optimizing the third-party supervision workflow of the involved enterprises, and to verify whether the artificial intelligence ACO algorithm can solve the problems of low data quality and difficulty in adjusting algorithm parameters in the current regulatory process.

Data preprocessing and feature extraction

In the data preprocessing section, it is necessary to handle missing values and outliers in the collected data and standardize the data format. Different methods are used to deal with different problems. For parts with fewer missing values, the mean is used to fill in; Mean filling uses the mean to fill the missing values, which can effectively avoid losing too much data. for parts with a large number of missing values, linear interpolation is used to fill in the missing data, The linear interpolation method fills the missing values ​​based on the linear relationship between adjacent data points, effectively ensuring the continuity of the data. as deleting the data affects the overall feature extraction. In addition, the linear difference method is also used to replace outliers. Figure 3 shows the matrix scatter plot of various data items in the dataset.

Fig. 3
figure 3

Scatter plot of various data matrices.

Full size image

The distribution of the matrix scatter plot in Fig. 3 can roughly infer whether there are obvious data missing or outliers in the dataset. If there are significant gaps in the matrix scatter plot, it indicates that there are missing values in some data in the dataset; if there are some data in the matrix scatter plot with large intervals from other data, it indicates that there are outliers in some data in the dataset. It can be seen that there is no significant gap between a particular point and other points, that is, there are no obvious outliers in the various data in the dataset. Moreover, it can be preliminarily judged from the matrix scatter plot that there are no obvious gaps in each data item, that is, there is no significant missing value in any data type.

Data feature extraction is performed after data cleaning. The compliance status, severity of violations, and resolution status of violations in the dataset are encoded. For compliance status and violation resolution status, non-compliance and compliance are encoded as 0 and 1 respectively. Similarly, unresolved and resolved are encoded as 0 and 1, respectively. The low, medium, and high levels of violation severity are encoded as 0, 1, and 2, respectively. After encoding is completed, data from different sources are integrated together for standardization processing. The standardized partial data is shown in Table 2:

Table 2 Part of the standardized data.
Full size table

Table 2 only shows part of the standardized data. It can be clearly seen that the standardized data not only eliminates the problem of different dimensions, but also merges data of different categories together.

After completing the data preprocessing and feature extraction of the dataset, a high-quality dataset suitable for further analysis is obtained.

ACO optimization

80% of the processed dataset is used as the training set, and the remaining 20% is used as the test set. ACO is used to optimize the compliance third-party supervision workflow of the involved enterprises. The initial parameter settings and parameter value ranges of ACO are shown in Table 3.

Table 3 ACO parameter configuration table.
Full size table

The initial parameter configuration and corresponding value range of ACO can be obtained from the configuration table in Table 3. Among them, the range of ant count is 10 to 100, with an initial ant count of 20; the range of values for the pheromone importance factor is 0.5 to 3, with an initial pheromone importance factor of 1; the value range of heuristic information importance factor is 1 to 5, with an initial heuristic information importance factor of 2; the range of values for pheromone volatilization coefficient is 0.1 to 0.9, with an initial pheromone volatilization coefficient of 0.5; the range of values for pheromone intensity is 50 to 200, with an initial pheromone intensity of 100; the maximum number of iterations ranges from 50 to 500, with an initial maximum of 100 iterations.

The loss value data obtained by inputting the training set data into the ACO model is shown in Fig. 4:

Fig. 4
figure 4

Loss value after 100 iterations of the training set.

Full size image

Figure 4 shows the loss value of the training set after 100 iterations. It can be seen that in these 100 iterations, the initial loss value of the training set is about 100, and it drops to nearly 10 at the 100 th iteration. From a trend perspective, the loss value shows a clear downward trend. It is worth noting that there are certain fluctuations in some iterations due to the influence of random noise, but the overall trend continues to decrease. However, from the perspective of iteration, the loss value has been showing a downward trend throughout 100 iterations and has not flattened, so it is necessary to adjust the parameters of ACO.

After several adjustments, the parameters of the ACO are adjusted to 50 for the number of ants, 1.5 for the pheromone importance factor, 2.5 for the heuristic information importance factor, 0.3 for the pheromone volatility coefficient, 150 for the pheromone strength, and 200 for the maximum number of iterations. The loss value of the training set after 200 iterations is shown in Fig. 5:

Fig. 5
figure 5

Loss value after 200 iterations of the training set.

Full size image

Figure 5 shows that after 200 iterations, the loss value drops from 100 to nearly 5. This also indicates that by increasing the number of iterations of ACO, the loss value can be further optimized to achieve better results. Moreover, it is worth noting that in the 190 th iteration and the following ten iterations, the loss value only decreases by 2.27, and the maximum difference in loss value among these iterations is only 3.3, indicating that the downward trend of the model’s loss value tends to be gentle and the model is close to convergence. Increasing the number of iterations can cause overfitting.

Model evaluation

In the model evaluation stage, a comprehensive evaluation of the third-party supervision workflow for compliance of the involved enterprises after optimizing the ACO algorithm is conducted. The data from the test set is input into ACO, and the evaluation indicators of the test set are used as the final evaluation result for ACO to optimize the third-party supervision workflow of the involved enterprise. By comparing the evaluation results of the training and test sets before and after using ACO optimization, the feasibility of ACO optimization in the compliance third-party supervision workflow of the involved enterprises is analyzed.

A comprehensive evaluation includes four aspects: data quality, model performance, scheme effectiveness, and supervision effectiveness. Figure 6 shows the evaluation results of data quality and model performance:

Fig. 6
figure 6

Results of data quality and model performance indicators.

Full size image

Figure 6 shows the performance metrics of the model in terms of data quality and model performance on the training and test sets before and after ACO optimization. Figure 6a shows the performance results of the model data quality before and after ACO optimization. Among them, the MSE and \(\:{\Gamma\:}\) of the training set before optimization are 0.04 and 0.03, respectively; the MSE and \(\:{\Gamma\:}\) of the training set after optimization are 0.02 and 0.02, respectively; the MSE and \(\:{\Gamma\:}\) of the test set before optimization are 0.05 and 0.035, respectively; the MSE and \(\:{\Gamma\:}\) of the test set after optimization are 0.03 and 0.025, respectively. It can be seen that both the training and test sets show a significant decrease in MSE and \(\:{\Gamma\:}\) after ACO optimization. Figure 6b shows the performance of the model before and after ACO optimization. Whether it is the training set or the test set, the accuracy, recall, F1 score, and AUC-ROC after ACO optimization are all better than before optimization. Therefore, it can be seen that, similar to the evaluation results of data quality, the model performance after ACO optimization is still better than before optimization in terms of model performance.

In order to better verify the optimization effect of ACO, this paper sets up different optimization algorithms to conduct comparative experiments on the performance indicators of the test set data quality and model performance.

Table 4 Optimization results of each optimization algorithm.
Full size table

Table 4 shows the optimization results of different algorithms on the experimental data test set. It can be seen that the indicators of GA (Genetic Algorithm) are 0.79, 0.78, 0.79 and 0.79 respectively; the precision and recall of PSO (Particle Swarm Optimization) algorithm are 0.77 and 0.79 respectively; the precision and recall of reinforcement learning algorithm RL (Reinforcement learning) are 0.76 and 0.77 respectively; It can be seen from the above data that ACO has the most obvious optimization effect on the test set. It optimizes other algorithms in various indicators, can better optimize model performance, and greatly improve model training efficiency.

Figure 7 shows the indicator results of the scheme effect and supervision effect.

Fig. 7
figure 7

Evaluation results of scheme and supervision effect. (a) Evaluation results of the scheme and supervision effect before ACO optimization. (b) Evaluation results of the scheme and supervision effect after ACO optimization. (c) Evaluation results of the scheme and supervision effect of the training set before and after ACO optimization. (d) Evaluation results of the scheme and supervision effect of the test set before and after ACO optimization.

Full size image

Figure 7 shows the scheme and supervision effect evaluation results of the model before and after ACO optimization. It can be seen that the CER, NPV, SCR, and CRC of the model optimized with ACO are better than those without ACO optimization. Specifically, the CERs of the training and test sets using ACO are 0.25 and 0.28, respectively, while the CER values without ACO are 0.35 and 0.38, respectively. In terms of NPV, the NPV values of the training and test sets using ACO are 0.41 (million US dollars) and 0.4 (million US dollars), respectively, while the NPV values without ACO are 0.35 (million US dollars) and 0.335 (million US dollars), respectively. In terms of SCR, the training and test sets using ACO have SCR values of 0.9 and 0.88, respectively, while those without ACO have SCR values of 0.8 and 0.75, respectively. In terms of CRC, the values for the training and test sets using ACO are 0.15 and 0.12, respectively, while the values without ACO are 0.1 and 0.08, respectively.

Experimental discussion

This simulation experiment uses ACO to optimize the compliance third-party supervision workflow of the companies involved in the case, and compares the performance of the models using ACO and not using ACO on multiple key evaluation indicators. The experimental results show that after adding ACO optimization, the data quality, model performance, scheme effectiveness and supervision effectiveness are significantly better than those without ACO, further verifying the advantages of ACO in data cleaning, integrity maintenance, prediction classification, balancing cost-effectiveness and formulating regulatory measures. ACO’s optimization ability makes the supervision process more accurate and efficient, and provides a more intelligent solution for compliance management. However, the experiment also exposed certain limitations and issues worthy of further discussion. First, because the data of various indicators in the training set are generally higher than those in the test set, it reflects that the model may have a certain degree of overfitting during the training process, that is, it performs well on known data, but when facing unknown data in the test set, the generalization ability decreases. This shows that although ACO optimization has improved the effectiveness of the supervision process to a certain extent, it may still face problems of adaptability and stability when dealing with complex and dynamic real-world scenarios. Although the indicators on the test set are still at a good level, its continued stability and anti-interference ability in long-term practical applications have not been clearly evaluated. Future research should further optimize the ACO parameter adjustment strategy, introduce richer test data sets, and combine other intelligent optimization algorithms to improve the generalization and robustness of the model, ensuring that the ACO-optimized regulatory process remains efficient and reliable in a more complex corporate compliance environment.

Conclusions

This paper innovatively studies the enterprise compliance third-party supervision workflow model based on artificial intelligence ACO optimization, and proposes an optimization plan for the problems of low data quality and difficulty in adjusting algorithm parameters in the current compliance supervision and consolidation work. The structural characteristics of the ACO algorithm are analyzed in detail, and combined with the specific needs of compliance supervision, the supervision work can be operated more intelligently and efficiently. Through simulation experiments, the results show that the model with ACO optimization performs well in multiple key indicators, including improving data quality, optimizing model performance, enhancing the feasibility of the plan, and improving the effectiveness of supervision. Compared with traditional methods, ACO optimization can more accurately identify compliance risks, optimize data integration processes, and effectively reduce the complexity of supervision work, showing strong applicability and promotion value. This innovation not only provides more scientific technical support for corporate compliance management, but also lays an important foundation for the construction of future intelligent supervision models.

This paper still has certain limitations. (1) The amount of simulation data used in the experiment is relatively limited, and it fails to fully cover the supervision needs in larger and more complex scenarios, which may affect the generalization ability of the model in practical applications. (2) The data source is relatively single, lacking multi-dimensional data support from different types of enterprises, industries or regulatory agencies, which may limit the adaptability of the model. (3) The parameter adjustment of the ACO algorithm is still difficult. How to further optimize the parameter tuning method to improve the stability and applicability of the model is still a key issue that needs to be studied in the future. In order to further improve the practical value of the research, future work should focus on expanding the data scale, enriching the data source, and combining other intelligent optimization algorithms to continuously improve the intelligence level of the compliance supervision process and promote corporate compliance management to develop in a more accurate and efficient direction.

In order to further enhance the practical value of this study, future work should focus on expanding the data scale, enriching the data source, combining other intelligent optimization algorithms, gradually solving challenges such as scalability and parameter sensitivity, and promoting the improvement of the intelligence level of compliance supervision process. Through this process, corporate compliance management will develop in a more accurate and efficient direction, and ultimately provide a more intelligent, flexible and efficient solution for corporate compliance supervision. And In order to further improve the practical value of the research, future work should focus on expanding the data scale, enriching the data source, and combining other intelligent optimization algorithms, especially multimodal AI integration methods, to improve the intelligence level of the compliance supervision process. Multimodal AI integration will help to integrate different types of data (such as structured data, unstructured data, real-time data, etc.) to provide more comprehensive and accurate decision support. This can not only improve the accuracy and reliability of the supervision process, but also enhance its adaptability in more complex scenarios. Ultimately, it will promote corporate compliance management in a more accurate and efficient direction and provide more intelligent and scalable compliance supervision solutions for all types of companies.