Machine Learning in Procurement with a View to Equity

1. Introduction

Thanks to the use of digital information systems, organizations have stored data about their business processes. The large amounts of data collected can be used for many purposes, from monitoring progress and assessing risks to setting targets and making comparisons with other contexts (especially when these data are cross-checked with open data from other realities).

This chapter discusses the Italian case of tender notices published by public administrations.

Public procurement in Italy concerns contracts stipulated by the Public Administration with economic operators for the supply of goods, services, or works. The award process takes place according to precise rules, mainly regulated by the Procurement Code.¹

The process of awarding a public contract follows these main phases:

Tender notice: The Public Administration publishes a tender notice that defines the project and the requirements for participation.

Presentation of offers (or bids): Interested economic operators present their offer, which includes technical and economic proposals.

Evaluation of offers: A special commission evaluates the offers according to specific criteria (lowest price or most economically advantageous offer).

Awarding: The contract is awarded to the economic operator who submitted the best offer, compliant with the established criteria.

Execution of the contract: The winning operator executes the contract, respecting the agreed conditions.

The entire process is subject to principles of transparency, competition, and non-discrimination to guarantee fairness and correctness.

Tender notices are published in the Official Gazette and a few data on them is sent by the public administrations to the regional public expenditure observatories or directly to the National Anti-Corruption Agency (ANAC). ANAC was set up by the Italian government in 2010 with the aim of monitoring and reducing public expenditure. The ANAC database on tenders is public and can be queried online [1].

The accumulated big data, consisting of more than seven million notices, on the expenditure of tenders of all the Italian public administrations, whether they are works, supplies, or services, is an extremely valuable resource. It provides data and evidence to help the data analysis expert formulate and test the validity of the hypothesis, from the awarding process to the litigation in front of the Regional Administrative Courts, to the execution of the contract, testing, and so on.

The tenders in the ANAC database are highly diverse, covering works, supplies, and services, and each tender has its specific type and sector. A tender may fall under public works, services like consultancy, or supply contracts. These tenders are significant as they determine the allocation of public funds and play a key role in economic development. Understanding the tendering process is critical to ensuring transparency, competitiveness, and fairness. For instance, consultancy tenders, which are the focus of this chapter, can be prone to variability and require close monitoring to prevent cost overruns. The data collected on these tenders, such as the CIG identifier, tender type, and award amounts, serve as a foundation for our study. There are two types of motivations for the focus of the chapter on the analysis of counseling tenders. The first one is Simpson’s paradox.² According to this principle, a trend or correlation among variables appears in heterogeneous groups of data but it could disappear when the groups are merged. Therefore, we would like to reduce the risk of obtaining erroneous results by restricting the data analysis to a single group composed of more homogeneous types of tenders. The second is the awareness of the fact that the phenomenon of money laundering could be facilitated by the existence of a consultancy cost that is not easy to quantify or limit directly. Finally, we leverage a recent trend in research on Artificial Intelligence (AI) known as “Trustworthy AI” [2] in which ethical values and human rights are taken into account when AI tools are developed and used in society. Among the guaranteed rights, there is the right to obtain explanations for the outcome of a (technological) process, accountability such as the right to find responsibilities for every mistake of the AI tools, equity interpreted as the guarantee that everyone receives equal treatment regardless of ethnic group, the provenance, or other sensible features. In other terms, fairness and non-discrimination have a central role and should be used to evaluate the good performance of a human or technological process in which AI is involved [3].

The situation is particularly delicate in the case of tenders by public administrations (PA) involving the transfer of public funds. On the one hand, there is the right of every citizen of the territory to receive public money as a reward for the commitment and professionalism involved in a project for public administration. On the other hand, there are different contexts in which public administrations and economic operators operate in their territories. Not all situations are comparable, especially if we take into account the different economic backgrounds that exist in the Italian territory. The south, the so-called “Mezzogiorno”, is more disadvantaged, but also suffers more from the presence of criminal organizations originating from these regions. Criminals try to exploit and extort money from the legal and public administrations. As a result, southern Italy may suffer more from coexistence with criminal organizations and be suspected of involvement with them. This chapter studies in an objective way the available data on the tendering process on the Italian territory to measure with the lens of data analysts and tools for measuring AI fairness [3] if inequalities and asymmetries exist. Knowledge about the existence of the phenomenon is the first step to correct it.

The secondary objective of this chapter is to demonstrate that the ANAC database is a valuable resource. When analyzed, the available data can be used to improve the tendering process. Two directions are followed. The first uses predictive AI models trained on the ANAC data, to predict whether a contract awarded by a tender will be involved in a variation. This prediction could help public administrations avoid the variation, which often leads to delays and increased costs. In the second direction, a tender recommendation system is being developed on the ANAC database. This is a valuable resource for both the PA and economic operators, who can use it to search for similar cases in the past, similar competitors who could be possible partners for similar projects, and obtain useful information on costs, time, constraints, award procedures, etc. Moreover, recommendation systems help to increase the competitiveness and diversity of the bidding pool by analyzing the companies that typically bid on specific tenders by helping to suggest new companies that have not participated before but might be interested in bidding.

The remainder of this chapter is organized as follows: Section 2 introduces the related works; Section 3 describes the case study; Section 4 describes the study on equity of awarding tenders on the territory; Section 5 provides insights about predicting the presence of a variation in the procurement; Section 6 outlines the recommendation system and the obtained results. Finally, Section 7 concludes the chapter.

2. Related work

In the context of Machine Learning (ML) in procurement, there is a growing focus on leveraging AI technologies both at the National [4] and European level [5]. One study emphasizes the importance of integrating ethical and sustainable sourcing practices into supply chain management, highlighting how ML and big data analytics can improve decision-making processes [6]. These technologies enable more socially responsible procurement by ensuring transparency, reducing biases, and fostering fair supplier selection and contract management [7].

Another perspective is offered through research on supply chain resilience, which examines how AI and ML can enhance information exchange among supply chain partners. This approach reduces risks and supports socially equitable outcomes by promoting transparency and fairness in procurement processes. AI technologies are seen as key to driving social equity by improving supplier diversity and advocating for fair labor practices across supply chains [8, 9].

A systematic literature review of social procurement in the construction and infrastructure sectors explores the evolution of procurement practices to include social value creation. This shift focuses on using government initiatives to mandate social outcomes in procurement contracts. The review identifies barriers and enablers to social procurement and provides strategies for overcoming challenges, highlighting the potential of procurement practices to deliver social benefits and promote equity [10, 11].

Additionally, the role of AI in procurement is explored in terms of its potential to strengthen diversity and inclusion. Data analytics and ML are highlighted as tools to identify opportunities to engage diverse suppliers and ensure equitable procurement processes [12]. This approach aligns with global trends in utilizing technology to support socially equitable procurement practices and underscores the importance of integrating equity considerations into ML models used in public procurement [13, 14].

Recommendation systems are increasingly utilized in the legal domain to swiftly retrieve relevant documents for specific cases. In the study of Dhanani et al. [15], a graph clustering method is proposed to group referentially similar judgments and identify semantically relevant ones within these clusters. In the study of Nai et al. [16], numerical vectors known as sentence embeddings [17] were trained using BERT (Bidirectional Encoder Representations from Transformer) [18] to build an abstract and general representation of the semantic content of contract descriptions. Input sentences were taken from the brief descriptions in natural texts of procurement in the ANAC database, resulting in 768-dimensional vectors. Subsequently, for an individual procurement case, the most similar and relevant ones in the rest of the database were searched using SBERT [19] and LaBSE [20]; similarly, in the study of Nai et al. [21], the performance of a recommendation system based on the embeddings provided by two different commercial models was compared.

Bert LaBSE model is a language-agnostic Bidirectional Transformer Sentence Embedding [20]. A transformer has one or more encoders, used to represent in a latent model the context of the sentence, while the decoder parts are usually applied to generate sentences from the abstract representation and to change the language in automatic translation. BERT connects the encoder and decoder through a token-based attention mechanism. In BERT, consecutive sentences from the input are separated into tokens, transformed into numerical vectors from which the system performs different tasks such as learning the words’ context through random masked words and learning if sentences are consecutive (or related in a question-answering task).

As regards equity in public procurement, van Dijk and Wilke [22] examined the effect of different interests on public-good provision by adopting the equity theory. The latter aims at determining whether the distribution of resources is fair taking into account the contributions (or costs) and the benefits (or rewards) for each individual in an organization. Decarolis and Giorgiantonio [23] employ ML in studying the effectiveness of indicators used in police investigations on corruption (called red flags) in public tenders for roadworks. The survey [24] provides an overview of the many control measures for fairness that researchers in AI proposed to monitor the fairness of a process. The monitored process might be assisted by AI and ML tools and its fairness needs to be verified with respect to the membership to a group (that expresses a sensitive feature) for some of the involved actors. The measures for fairness are guided by different concepts: (a) parity-based metrics that compare the predictive positive rates across groups. An instance is statistical independence between the predicted score and the group; (b) Confusion matrix-based metrics that compare groups by taking into account the potential differences across groups. An instance is separation that evaluates the equal opportunity of comparable individuals from the different groups; (c) Calibration-based metrics that compare groups by the probability that the predicted scores are emitted. An instance of this metric is sufficiency with tests for fairness. In Section 4.1, we apply the above three concepts to assess the fairness of the participation of economic operators in public tenders and the subsequent award of contracts, taking into account their different origins: north of Italy, center of Italy, or “Mezzogiorno” (south or islands).

3. Case study: Tenders of Italian anti-corruption authority: ANAC

The primary dataset pertains to ANAC, the Italian government authority collecting public tenders. A dedicated section on the ANAC website [1] allows users to access data in a standard format, enabling category selection and downloading in compressed files. In the ANAC Open Data catalog, the central dataset concerns the creation of a Tender Notice, organized by year and month. Besides the tender dataset, four additional significant datasets are available: the list of Contracting Authorities (CA) that have issued a tender, the list of tenders that have received an Award, the Economic Operators (EO) who have won a tender, and the activities associated with a tender post-award (e.g., contract-start, contract-end, subcontract, etc.). Each of these four datasets is briefly described below. CAs³ are public bodies or entities acting on behalf of public institutions, responsible for procuring goods, services, or works through tenders. They oversee and manage the entire public tender process. CAs are categorized into three types: Central (e.g., ministries), Regional and Local (e.g., municipalities), and other entities (e.g., hospitals). Awards include a list of awarded tenders, with details such as the final awarded amount, the award date, and the EOs receiving the tender. EOs can be individual enterprises, artisans, partnerships or corporations, cooperatives, etc. A 10-character alphanumeric variable called CIG identifies each tender and its related activities. CAs and EOs are recognized by their tax code (an alphanumeric string), and CAs also have a unique ISTAT code. Finally, the Variants dataset contains information on any variation authorized during execution from the original contract, as a result of unforeseen circumstances. Figure 1 presents a portion of the Tender Notice and Award where some tenders have not been awarded to an EO (with DATE_AWARD and TENDER_AWARD fields left empty).

4. Descriptive tasks and fairness study

When analyzing a phenomenon from data, it is suggested to create descriptive models (such as geographical heat maps, histograms, etc.) that allow a first look at the phenomenon with a summary of an immediate interpretation. Figure 2 summarizes the total amount spent on counseling tenders in the studied period (2016–2019). The outlier over the other regions, whose expenses are instead more similar, is the region Lazio, in the central macro-area. The capital of Italy, Rome, is located in this region and this could be a justification for the higher number of counseling contracts for the government of the country.

In Figure 3, we show the result of a predictive model (linear regression) that we trained to predict, as a target, the fraction of tenders for counseling in the same period on the whole Italian territory. The regression model was trained from the predictive features that are macroeconomic variables of the economic richness of the regions. These variables are shown in Table 2 with the coefficients in the regression model.

These economic variables were chosen by the financial and economic institutions in Italy, like Banca d’Italia and Sistema Conti Pubblici Territoriali [25]. Data on these variables were downloaded from ISTAT.⁹ To train the regression, we expressed the value of each variable for each region as a fraction of the global value for the Italian territory, so that the ranges of the variables are comparable and none of them can dominate the others in the regression model.

The results of the regression in terms of root mean squared errors (RMSE) are good because RMSE accounts for 0.3%. The intuitive meaning of using linear regression to predict the presence of contracts awarded by economic operators in the regions, based on socio-economic data of the regions, is that the presence of awarded contracts is proportional to the economic wealth of the people living in the region and to the number of economic operators. The coefficients of the regression agree with this interpretation (see right column of Table 2). Among the variables, the most important ones are those with a higher and positive coefficient, that are internal consumption, GDP, and number of economic operators while those that are negatively correlated with the target have a large negative coefficient: expenses for consumption on the territory by families and expenses for consumption of the public administrations.

The total amounts for counseling tenders for each region are shown in Table 3 together with the proportion of tenders predicted by the regression model. If the predicted value of the fraction of tenders awarded is negative, it means that for those regions, the estimated value is higher than the actual value. This means that in these regions, we would expect proportionally more tenders to be awarded: this is the case for many regions in the Mezzogiorno. In Section 4.1, we go deeper into the analysis by considering the fairness of the distribution of tenders in the three macro-areas mentioned.

5. Predictive tasks

In the predictive tasks, several ML models were employed to evaluate their performance and accuracy in handling complex datasets. The models included Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), XGBoost (XGB), and Naive Bayes (NB), each offering distinct advantages and trade-offs explored in this research.

LR, a linear model, is often used for binary classification tasks. It is straightforward and interpretable, making it a suitable baseline model. However, its simplicity limits its effectiveness in capturing complex relationships in high-dimensional or non-linear data [27].

DT, a non-linear model, splits data based on feature importance, making it more flexible than LR. DT is easy to interpret and can handle both numerical and categorical data. However, it tends to overfit the training data if not properly tuned, leading to reduced generalization on unseen data [28].

RF builds on the decision tree model by creating an ensemble of trees, improving robustness, and reducing overfitting. It performs well on complex datasets due to its ability to capture non-linear patterns and interactions between features. RF also provides feature importance scores, making it useful for interpretability in large datasets [29].

XGB, an advanced gradient boosting algorithm, often outperforms other models in terms of accuracy, especially with complex datasets. It builds trees sequentially, optimizing for errors made by previous trees, which allows for a more refined learning process. XGBoost is highly efficient in handling large-scale data but requires careful tuning of hyperparameters to achieve optimal performance [30].

NB is a probabilistic model based on Bayes’ theorem, and while it performs well in certain text classification tasks or cases where feature independence can be assumed, it is generally less effective with complex, non-linear datasets. Its main advantage is its computational efficiency, which makes it useful for large datasets with simple relationships [31].

In particular, the described experiment aims to identify procurement that has undergone variations. Variations could be used by economic operators as a “kind of strategy” to gain back part of the lot amount that they originally lowered to have more chances of being awarded the contract. Thus, variations impact the final contract price and time and alter the budget allocations. Specifically, funds allocated to a variant are no longer available for other activities, thus unbalancing the overall expenditure. Therefore, procurement for services (S), namely consultancy services, was selected for this experiment. To keep the dataset balanced in the consultancy-related variants, procurement with the same CPV divisions (73 and 85) was chosen 4 years before COVID-19 (2016–2019); the resulting dataset thus consists of 25,732 tenders.

The dataset contains procurement marked by whether they include a variant (label = 1) or not (label = 0). The two classes are inherently imbalanced, with one possibly more prevalent than the other. The dataset was balanced by sampling equal numbers from each class to address this. This ensures that the ML models do not favor the more frequent class, thus providing a fair evaluation. The balanced data was randomly divided into training, validation, and test sets, with an 80–10–10 split, where the test set was reserved for unseen data to evaluate model performance on new, unobserved examples. Features were standardized and encoded to improve the performance of certain models, ensuring they operate effectively across varying scales of input data. In detail, based on the dataset described in Table 1, several preprocessing steps were applied to prepare the data for ML models. Numerical fields, such as the Tender amount and Awarded amount, were normalized to bring all values within a standard range, ensuring that no single variable dominated the learning process due to differences in scale. Categorical fields like Tender type were encoded using one-hot encoding, which converts the categories into binary vectors for compatibility with ML algorithms. For boolean fields, such as the Framework Agreement and Subcontracting admitted, a boolean encoding was used, where 1 indicates true and 0 indicates false. These preprocessing steps are standard in ML to ensure that the data is in a format suitable for training, reducing bias, and improving the accuracy of the models [32]. As described at the beginning of this section, various models were trained: LR, DT, RF, XGB, and NB. These models were chosen for their diverse capabilities, from simple interpretations to handling complex, non-linear relationships. Hyperparameter tuning was conducted on the models using the Hyperopt library to optimize their performance and identify the best combination of parameters for accurately predicting procurement affected by variants. The models were evaluated using cross-validation [33] to test their generalization capabilities. This method ensures that models perform well across different subsets of data, reducing the risk of overfitting the training set. Model performance was assessed using Accuracy [34], Precision [34], Recall [34], and F1-Score.¹¹ These metrics provide insights into how well each model can identify procurement with variations, which is crucial for understanding their budgetary impacts. The data and scripts for this part of the experiments are publicly available.¹²

5.1 Predictive tasks: Results

The cross-validation metrics provide a robust evaluation of model performance by accounting for variability in the data, as highlighted in Figure 4 and Table 5. Among the models analyzed, XGB demonstrates superior performance, with the highest cross-validated accuracy of 0.965 and F1-Score of 0.965. This indicates high precision in its predictions and a strong balance between precision and recall, with precision and recall values of 0.983 and 0.947, respectively, making it an ideal choice for datasets where both false positives and false negatives are costly. The RF model also performs well, with an accuracy of 0.947 and an F1-Score of 0.948. Although slightly lower than XGB, it still offers a commendable trade-off between recall and precision, reflected in its precision of 0.957. This suggests that RF is reliable in applications where precision is prioritized, albeit with slightly less balance than XGB. The DT model showcases a moderate performance, with an accuracy of 0.953 and an F1-Score of 0.954. Despite having lower overall precision than the RF and XGB models, it achieves a high precision of 0.965, indicating its effectiveness in scenarios where correct positive predictions are more critical than capturing all potential positive instances. In contrast, LR and NB present substantially lower cross-validation metrics. LR achieves an accuracy of 0.879 and an F1-Score of 0.874, reflecting its limited ability to capture the complexities in the data compared to ensemble methods. Similarly, NB records the lowest performance across the board, with an accuracy of 0.736 and an F1-Score of 0.683, indicating that its assumptions may not align well with the underlying data distribution. Overall, the cross-validation metrics emphasize the robustness of ensemble methods, particularly XGB and RF, in delivering consistent and reliable predictions across varying datasets.

Model	Accuracy	F1-Score	Precision	Recall
XGBoost	0.965	0.965	0.983	0.947
Random Forest	0.947	0.948	0.957	0.940
Decision Tree	0.953	0.954	0.965	0.941
Logistic Regression	0.879	0.874	0.916	0.837
Naive Bayes	0.736	0.683	0.859	0.567

Table 5.

Predictive models results: (a) Precision, (b) Recall, (c) F1-Score, (d) Accuracy. In bold the best model, underlined the second.

The recommendations for stakeholders are to carefully monitor the incidence of variations in total costs and contract completion times. For policymakers, it is a matter of proposing regulatory corrections, but also of carefully monitoring the overall impact of variations on public expenditure. Again, the big data available in the ANAC database can help, as it can be used to model the general behavior of economic operators in proposing and executing tenders, and compare them specifically for the frequency of variation requests. Data analysts and machine learning tools applied to the big data of public contracts could detect economic operators that represent anomalies in proposing several variations in a recurring manner and, on the contrary, propose best practices and examples of virtuous behavior.

6. Recommendation systems

By filtering the database for a description of CPV codes related to “consultancy” tenders, we obtained 72 CPV codes. In particular, the CPV codes used for filtering the ANAC database are mainly from these divisions based on the first two digits: Table 6 shows the list of CPV codes retrieved by using the recommendation system on the description of codes. Results show that there are 25 categories of CPV codes starting with the 71XXX000-Y division which are related to counseling.

6.1 Item retrieval and precision score

The recommendation system trained on the original ANAC database was used to obtain the top 10 tenders most similar to each of the “consultancy” tenders used as a query example.

For item retrieval, we used a Bert LaBSE model (Language-agnostic Bidirectional Transformer Sentence Embedding) [20]. Embedding [18] is a transformative approach to natural language that has become popular with deep artificial neural networks trained on large volumes of documents. It is used to represent the semantic content of documents written in natural language. As it is the case of language-agnostic models, documents might be written even in many languages. Embeddings were employed to treat the short description of procurement provided with the tender subject field in the ANAC database. Embeddings of the tender subjects represent, in a concise yet quite general and powerful way, the topic referred to in the tender subject. Embeddings are numerical vector representations that capture the semantic essence of the data. Cosine similarity was later used to find similar items.

To retrieve tenders related to the users’ interest from ANAC-approved tenders, a query in natural language can be submitted by the user and processed to retrieve the top k tenders more similar to the description provided by the query. For example, Table 7 shows the top 10 most similar tenders related to the query “legal counseling”.¹³ Cosine similarity [35], a mathematical metric used to measure the similarity between two vectors in a multi-dimensional space, is then used to compare these embeddings. It involves calculating the cosine of the angle between the embeddings, providing a measure of how similar they are, based on orientation rather than magnitude. Table 7 shows the results obtained by the user when a query is made. The system, when processed the user’s query and transformed it into an embedding using the deep neural network of Bert LaBSE, uses cosine similarity to measure the relevance of tenders in its database to the query. This ensures that the most similar and relevant tenders are returned, enhancing the efficiency and effectiveness of the search. However, these results can be explained by looking at the texts of the notices taken into consideration. Clearly, for the example provided by the query “legal counseling” for the recommendation system, it is easy to recognize that the sets of tenders shown in Table 7 whose descriptions are “LEGAL SERVICES” and “LEGAL ASSISTANCE SERVICES” are similar to the query and to each other. These are assigned a cosine similarity of 0.99 and the two calls appear to be, each other, among the top 10 of their neighbors. If we consider as true positives (TP) the neighbors who are also of the “consultancy” type, we can calculate the precision@10 using the following formula:

Query	Cig	Top k Tenders subject	Code CPV	Accuracy
1	69665502ED	LEGAL SERVICES	79,111,000-5	TP
2	8,018,638,568	LEGAL SERVICES	79,111,000-5	TP
3	76317249DE	LEGAL SERVICES	79,111,000-5	TP
4	8,139,176,483	LEGAL ASSISTANCE SERVICES	79,110,000-8	TP
5	78756296A5	LEGAL ASSISTANCE SERVICES	79,110,000-8	TP
6	8139089CB5	LEGAL ASSISTANCE SERVICES	79,110,000-8	TP
7	7993481D2F	LEGAL ASSISTANCE SERVICES	79,110,000-8	TP
8	7,912,174,489	LEGAL ASSISTANCE SERVICES	79,110,000-8	TP
9	79122199AA	LEGAL ASSISTANCE SERVICES	79,110,000-8	TP
10	7856558CC0	LEGAL ASSISTANCE SERVICES	79,110,000-8	TP

Table 7.

Top k tenders retrieved by the recommendation system for the query “LEGAL SERVICES” with k=10.

Precision@10=Relevant  tenders  in top 10 recommendations10=TP10E2

We tested the recommendation system for 100 queries randomly selected from the database with CPV in Table 6. We also filled the database with irrelevant examples randomly selected from the rest of the ANAC database to give the system also examples of irrelevant cases for the query. Their number is equal to the number of cases with CPV in Table 6. The obtained results of the recommendation system are that precision@10 values are, on average, equal to 0.85 considering the list of the top 10 most similar results to the queries. Results indicate that more than half of the recommended tenders out of the top 10 are of the “consultancy” type, and are similar to the tender under investigation (query tender). Table 8 and the side picture show the results of precision for the tender recommendation system for random queries with CPV in Table 6. It results that 74.15% of queries have an extremely high precision score (equal to 0.9) while there is almost 8% of queries whose recommended tenders were less similar (with the lowest precision scores, in the two ranges 0.20–0.40 and 0.40–0.60).

Table 8.

Distribution of precision scores for the tender recommendation system for random queries with CPV in Table 6.

The results allow to make claims for stakeholders: they should spread the possibility of using machine learning tools and recommendation systems, making the search for cases more effective and speeding up the bidding and awarding processes. If contracting authorities had a complete collection of cases similar to their own at the moment they were drawing up their tender, they could prepare their tender with the most appropriate constraints. As a result, we would expect to see a lower number of no bids. Similarly, if economic operators had a collection of similar cases to their own, they would propose more appropriate bids and see a higher success rate in their bidding, with a consequent reduction in effort and resources.

7. Conclusions

The case of tenders issued by public administrations in Italy in the field of consultancy over a period of 4 years is considered. A full analysis of the data in the ANAC database was carried out. A descriptive analysis of the data was applied in the form of a geographical heat map, which highlighted that Lazio was the Italian region with the highest number of tenders in consultancy.

Ensuring ethical and fair practices in public procurement is essential, particularly in light of regional economic disparities. The proposed study leverages ML models to assess fairness in the awarding of contracts. By applying fairness tests (independence, separation, and sufficiency), the discussed study shows that contracts tend to be awarded more frequently in northern regions compared to southern ones, despite similar economic conditions. This raises concerns about equity and fairness in the allocation of public funds, making it crucial to implement AI systems that promote fairness and reduce biases. The obtained results confirm the persistence of the historical southern question in Italy, the socio-economic consequences of which are inequality and unequal opportunities for the population throughout the Italian territory. Future procurement processes should consider these findings to ensure fair opportunities for all regions.

As regards the predictive task, the experiments successfully identified procurement cases with variants, highlighting their effects on budget distribution. When a variant causes an increase in the final price, it alters expenditure allocations, potentially depriving other activities of necessary funds. The predictive tasks highlight that XGBoost (accuracy 96.5%) outperformed all other models, making it ideal for predicting tender contract variations where both false positives and negatives are costly. RF (accuracy 94.7%) also showed strong performance, while Logistic Regression and Naive Bayes were less effective, with lower metrics indicating their unsuitability for complex datasets. For public authorities, implementing XGBoost or Random Forest models can help prevent delays and cost overruns by accurately predicting variations in bids. Policymakers should focus on integrating these AI tools to ensure fair and efficient public procurement processes. Contractors can use the insights from these models to better structure their bids.

Another discussed task is to train recommendation systems on the description of the tender topics. The recent technology of transformers (SBERT) was used, which is very successful in grouping texts with similar content. This type of recommendation service could be extremely useful to speed up the search for interesting cases in a large volume of examples, as well as to find similar situations or economic operators working in the same domain. The recommendation system boosts bidders’ experience to access and design with more competitive techniques by providing insights into previous bidders and winners for similar tenders. The results obtained by the recommendation system trained on the same consultancy data show that the majority of the queries (74.15%) submitted to the system have a high precision score (higher than 0.90), which serves as an evaluation of the relevance of the answers within the top 10 results. If the remaining precision score ranges are considered (below 0.90) the percentage of queries drops to values similar to 4.49, this indicates that a very small number of queries resulted in less relevant outputs and shows that the performance of the system follows a skewed distribution, skewed toward the high precision scores. However, analyzing the characteristics of these lower-performing queries can reveal specific patterns or contexts where the system struggles.

The recommendation to stakeholders and policymakers is to enlarge and enforce the use of AI/ML tools to assist actors in the public contract bidding process and execution in order to effectively use the available big data and make the process more efficient in the use of resources. A recommendation system based on a Large Language Model (LLM) specifically trained on tenders and their related data (such as bid amount, successful bidders, and contracting authorities) can improve the tendering process for the various stakeholders in decision-making and improve the efficiency of the operations. It indicates the most relevant tenders by analyzing the supplier’s or company’s history and the expertise related to the area of interest. The LLM-based recommendation system can offer insights into risk assessment by keeping a track record of specific sectors of winning bidders and allowing the financial authorities to access the risk of investment and finances for specific companies. The future work is to employ the presented tools to discover anomalies, propose best practices, and recommend suitable actors in order to enlarge, especially in the Mezzogiorno, the economic operators’ participation in the public tenders.

As Artificial Intelligence continues to transform delicate sectors of society, from healthcare to finance, it is crucial to ensure transparency, trust, and acceptance of AI technology [2]. Unfortunately, the complexity of many AI models often renders them “black boxes” with limited interpretability. Future works consider integrating Explainable AI (XAI) in predictive and recommendation models.

Acknowledgments

The authors thank the Department of Management for supporting this research with legal domain experts in public tenders, particularly Prof. Gabriella Margherita Racca and Dr. Francesco Gorgerino for contributing to understanding the procurement domain.

Conflict of interest

The authors declare no conflict of interest.

Appendices and addenda must be cited in the main text (example: See Appendix A). The section containing them must be titled accordingly (“Appendices”, “Appendix A”, “Addendum”, “Nomenclature”, etc). An example of appendix/addendum/nomenclature is given below:

Abbreviations