Currently, AI-enabled solutions in the public sector are integrated in a fragmented way that does not allow them to be easily transferable to other public organisations, highlighting the need for methodologies that streamline AI development process. For this reason, regulatory bodies have established standards that define what characteristics AI systems need to meet. Meanwhile, recent research initiatives call for comprehensive methods that ofer guidance on how to embed requirement aspects into the main AI development stages. However, existing approaches ofer high-level recommendations overlooking the dificulties in translating them into lowlevel implementations in order to bring together procedural and technical approaches. Moreover, the level of granularity of AI lifecycle stages does not facilitate the modularity of AI systems, which would enable the easy integration of new implementation to meet system requirements. In this paper we enhance the AI lifecycle pipeline with explicit operations that aid in the modular development and reuse of AI systems. Furthermore, we propose a structured framework that enables the operationalisation of system-wide objectives throughout the various phases of the AI lifecycle, while systematically leveraging existing toolkits. To illustrate the framework's adoption, we present a proof-of-concept example to showcase its practical application.
Artificial Intelligence (AI) is experiencing rapid growth and demonstrates significant impact across
various sectors, including public administration. AI capabilities improve decision-making processes,
enhance operational eficiency, and foster public service delivery [
Reaching this objective, however, remains a challenging problem. On the one side it is necessary to deal with the intrinsic complexity of delivering AI-based solutions reated to continuosuly evolving data, the unpredictable nature of AI models leading to variable outcomes, technological complexity. On the other side public administration sector to a much greater extent has face such problems as the lack of expertize and skill gaps, organizational implementation challenges, and insuficient architectural support on designing, deploying and monitoring AI-based systems.
Moreover, the development and deployment of AI products in public administration requires
considering regulatory compliance and AI trustworthiness aspects[
Current state-of-the-art approaches address these challenges from two perspectives. First,
international AI regulatory authorities focus on establishing guidelines [
Second, diferent solutions to support the complementary activities for AI product development and execution. These solutions, in the form of software libraries and framework may help in diferent areas, dealing with data and model quality, security and fairness, addressing diferent AI domains and applications. Still, their usage is often fragmented or too specialized; their relation to the regulatory settings and requirements is not often clear. Such clarity would help AI practitioners, and also non-AI experts, to seamlessly incorporate these tools into the lifecycle of an AI product in a modular and reusable manner, enabling both new and existing systems to be easily assembled according to initial or additional requirements.
To address these issues, MLOps practices should be enhanced to align more explicitly with systemwide requirements. Although extensive tools and frameworks have emerged on implementing individual ML pipeline components, developing production systems that incorporate them requires a more holistic approach which is still missing in the literature. It is essential to carefully plan the necessary AI lifecycle components and determine how they interconnect to support the system’s overall functional and performance goals.
In this paper, we propose a revision of the reference ML operations (MLOps) life-cycle [
The paper is structured as follows. Section 2 provides some background information about the concepts that our paper relies on. In Section 3 we outline the framework components and its final representation. Section 4 describes a proof-of-concept on how to use the framework, and finally conclusions are drawn in Section 5.
Reusability of AI solutions in the public sector is an important aspect as highlighted in recent guidelines
2. Various countries have adopted the practice of sharing software within the public sector to transform
their administrative procedures. Initiatives like the E-Government Action Plan 2016-20202 3, and
Sharing and Reuse of IT solutions Framework 4 encourage member nations to reuse information and
solutions. This approach aims to reduce costs, and improve the development of digital services by
reusing existing components or entire solutions in a more transparent and organised way [
Therefore, addressing the bespoke development of AI solutions in the public sector requires
responsible development practices [
To this end, on one hand, a range of procedural tools to regulate AI [
Although advancements are being made in creating regulatory, procedural, and technological tools,
integrating trustworthiness, ethical considerations, and service-level requirements into the AI
development process remains a challenge. This is being addressed by developing frameworks that assist in
converting high-level requirements into practical steps. Frameworks like Z-inspection [
Research eforts to integrate trustworthiness into the development of AI systems often treat each
aspect of trustworthiness separately. Although the TOP methodology [
In sum, there is no existing framework or approach that ofers a strategy to align actionable AI lifecycle operations with all the corresponding requirements an AI system must adhere to [19]. Moreover, there is a low level of automation of requirements implementation and verification due to a lack of standardization from a lifecycle perspective. There is still no structured and modular approach to leveraging existing technical tools which address these requirements in a fragmented way.
This paper proposes a framework, dedicated to implement a structured approach for requirements implementation, ensuring ethical functioning and adherence to service level agreements. Furthermore, 5https://framework.frictionlessdata.io/ 6https://fairlearn.org/main/auto_examples/plot_credit_loan_decisions.html#sphx-glr-auto-examples-plot-credit-loan-decisions-py 7https://github.com/Trusted-AI/AIF360 8https://holisticai.readthedocs.io/en/latest/getting_started/index.html 9https://huggingface.co/docs/bitsandbytes/main/en/index 10https://www.evidentlyai.com/ 11https://docs.seldon.io/projects/alibi-detect/en/stable/ 12https://huggingface.co/docs/transformers/en/index 13https://www.tensorflow.org/responsible_ai/model_card_toolkit/guide 14https://github.com/mostly-ai/mostlyai 15https://docs.giskard.ai/en/latest/getting_started/index.html it combines enhanced AI development operations with system requirements into a unified framework to systematically integrate existing technical tools that address some aspects in isolation toward achieving system-wide objectives.
The proposed framework aims at modeling the operations involved in the development, deployment, monitoring and maintenance of an AI product in a structured way. Specifically, we represent the AI life-cycle phases, the horizontal axis, as a minimal set of explicit operations having their specific semantics and characteristics. Note that in a specific AI solution, each operation may have multiple implementation variants or even be omitted within the corresponding phase. We then refine these operations across various cross-cutting aspects, vertical axis, related to the established requirements that the AI system must satisfy and consider, such as quality, fairness, privacy, and others. In this way, we make the implementation of these aspects explicit in the life-cycle and provide a way to map them to concrete solutions and tools.
In this section, we explain the significance and role of both axes along with their respective components, in Section 3.1 and Section 3.2. Finally, in Section 3.3 we present the layout of the proposed framework which hosts the procedures that the AI solution implements.
Building on earlier research on AI development lifecycle [
rithm to fit the best model JSON Evaluation data
after applying a function reports
the model torch, keras
for serving it, matching evaluation parameters
Type
In defining the operations semantics, we first specified the list of artifacts that each operation expects as input and produces as output, which are summarized in Table 2. In AI systems, managing these artifacts, produced by various procedures, enables automation, supports quality assurance, improves reproducibility, and facilitates auditability.
Furthermore, to facilitate the development and reuse of an AI solution or its components, it is essential to ensure that the system is suficiently modular. Modularity contributes also to easily ’apply’ system requirements and in assessing whether these requirements are met. We achieve modularity by dividing the three main stages of the AI development lifecycle into 15 explicit operations, as shown in Figure 1. These operations serve as the fundamental unit of development and deployment of an AI system and may be revisited and executed iteratively in varying orders [30].
The first stage of the AI lifecycle is composed of four well-defined operations related to data preparation such as, data profiling, data validation, data preprocessing and data documentation. Data preprocessing in particular, may involve many code implementations that apply trasformations to data, including data cleaning, bias mitigation, or data augmentation. During data augmentation, the initial dataset undergoes modifications to increase its volume and diversity. It is commonly used when
[26], Risk Card [27], AI Product Card [28], AI Cards [29]
Stage n o i t a r a p e r P a t a D g n i l l e d o
M training data quantity is not suficient, when some features contain bias or are not fairly distributed among diferent subgorups, or to address privacy concerns by avoiding the use of real data that contains sensitive information.
Then, the modeling stage includes the following operations: feature engineering, model training, model evaluation, model validation, and model and AI product documentation. Note that each operation may have multiple implementations within an AI solution, or diferent procedures addressing various system requirements may be grouped under the same operation category. Despite difering implementations, all variants within a given operation category share the same semantics, they receive inputs of the same kind and produce consistent type of outputs. This uniformity supports the modular design principle inherent in our proposed framework. Also, this means that each implementation of an operation in the AI pipeline functions as a small piece of the broader AI system, collectively contributing to the system’s overarching objectives.
Finally, the operationalisation stage consists of operations such as, model deployment, model monitoring, production data monitoring, system monitoring, pre-inference monitoring, and post-inference monitoring. Specifically, model deployment involves getting the model ready for the serving phase, enabling it to be accessed and utilized for inference, generation, classification, or the particular task it was designed for. It is crucial to note that the result of deployment is a function that encapsulates the model along with its parameters and settings required for its operation. These parameters are identical to those used during the model evaluation and validation stages, ensuring that the deployed model is the same one that was previously assessed. Table 3 summarizes the minimal list of explicit operations that make up the stages of the AI development lifecycle together with their input and output artifacts we defined previously.
The second axis of the framework captures all requirements that an AI-enabled solution must satisfy,
driven by the specific needs of the public sector [
We have delineated six fundamental requirements that a basic AI solution must satisfy and facilitate for implementation. These represent the minimal core criteria for evaluating not only the model itself but also key performance indicators (KPIs) such as training data quality, inference response time or number of resources utilized. We recommend including aspects such as quality, fairness, optimization, adaptability, transparency, and privacy, as these are among the most relevant considerations highlighted in the OECD’s report on the use of AI applications by governments [31]. To minimize the overhead that the framework’s comprehensiveness may impose on simple or small-scale products, it is advisable to focus on a single requirement dimension or to restrict the operation categories to only those that are essential.
Quality. Among these aspects, quality encompasses data integrity, model performance, system and pipeline eficiency. A critical step is assessing the available data, including its volume, quality, and formats, to determine viable machine learning approaches. This analysis also helps stakeholders decide whether additional data collection is necessary to achieve the desired model performance or data augmentation techniques are convenient to apply. Fairness. The second aspect under consideration remains one of the most critically important topics in the field of machine learning. Within the research community, discussions typically center on two key aspects: (1) quantifying a model’s fairness and (2) implementing interventions, during data pre-processing, model training [32], or post-processing, to improve fairness metrics [33]. Diferent philosophical interpretations of fairness, such as equality vs. equity are operationalized through distinct fairness measures and mitigation operations. Optimization. To minimize time and resource utilization during both the training and deployment phases, it is essential to integrate optimization strategies into the design of AI systems. This approach enables the systematic incorporation of performance-aware management principles throughout the machine learning lifecycle, ensuring alignment with predefined performance metrics and operational eficiency standards. Adaptability. Another important consideration is the adaptability of the AI system when it is reused by other public organizations seeking to reproduce it for similar use cases. Privacy. Among the various domains of responsible engineering, privacy has seen the most significant regulatory developments in recent years, with numerous jurisdictions enacting dedicated legislation. Transparency. An equally significant factor when developing AI-based solutions is the traceability and the communication of descriptive and prescriptive aspects that are relevant to diferent AI stakeholders [ 34].
The design of the proposed framework, is shown in Figure 2. The framework serves as an integrator of existing open-source technical tools and libraries. It organizes implementation procedures, denoted as <placeholder_procedure>, into a matrix considering both the AI development stages and the envisioned requirements. Earlier, we introduced the components that define the horizontal and vertical dimensions that coordinate the end-to-end development and deployment of an AI system.
The framework has a dual approach towards AI implementation and validation, which makes it accessible to both AI developers and compliance experts. On the one hand, it helps to explicitly and transparently declare and implement all requirements by leveraging existing tools and libraries. On the other hand, by embedding requirements into every operation of the AI lifecycle, our framework ensures that quality assurance, transparency, fairness and other aspects are built-in and evidence-based attributes of AI systems, rather than an afterthought. Moreover, it addresses the limited reusability of AI solutions in the public sector by responsibly engineering them. Finally, the framework plays a crucial role in implementing requirements by providing a systematic and structured approach to ensure that AI systems are developed ethically and comply with established service level agreements.
The following proof-of-concept example demonstrates how the framework can be instantiated with tools and modular implementations aligned with both specific requirements and AI development pipeline. We assume that each tool employed to address specific requirements is used to carry out a particular operation, with its inputs and outputs explicitly defined. This way, the operations create an abstraction layer that provide a level of uniformity to the employed toolkits. Additionally, the framework can be implemented alongside an existing cloud-native MLOps or AI platform, such as Google Vertex AI 16 or DigitalHub AI platform 17. To enable such integration, the implementation procedures must be adapted to comply with the governance and tool orchestration mechanisms required by the target platform.
This use case presents a classification model designed to identify the most suitable candidates for hiring
a new employee [35]. Figure 3 illustrates the populated framework. First, we focus on defining the
requirement dimensions that the AI-enabled solution should meet. Specifically, the stakeholders define
the list of the principles that the AI product should comply with based on the output of the evaluation
reports [
In this paper, we described the necessary operations that explicitly operationalize AI system requirements, contributing to implement responsible, modular, and reusable AI systems. Researchers and industry professionals in AI have been actively working on developing tools and methods that aid in transitioning experimental models to production deployment. By addressing the gaps in current frameworks that bring together procedural and technical approaches, our proposed enhanced lifecycle efectively facilitates the alignment between the implementation and the specific requirements of the AI system. Although existing libraries provide solutions for components of AI systems, they often fail to clearly specify the requirement aspect they address. Some approaches only partially cover some aspects, lacking a holistic perspective on the overall AI system. The clear structure of our framework explicitly connects each particular implementation to its relevant operation category and the specific requirement aspect it addresses. Moreover, the framework has a dual approach towards AI development lifecycle and requirements characteristics, enabling automation of AI product development and also verification. The framework contributes to the reuse of AI products by ofering a systematic methodology for developing them by transparently and explicitly describing and assembling the components of an AI system. Although the framework enables non-technical AI specialists to adopt AI easily and responsibly, especially in areas with low level of AI adoption like the public sector, a potential limitation is the necessity to establish procedures for conflicting requirement dimensions, such as balancing accuracy with fairness. This issue can be addressed earlier in the AI product’s design phase through a proactive approach, which aids in resolving ethical dilemmas in AI and also guides the implementation and evaluation phases.
The authors acknowledge the support of the “AIxPA - Artificial Intelligence in the Public Administration system” area of Flagship Project - Progetto Bandiera PNC-A.1.3 “Digitalizzazione della pubblica amministrazione della Provincia Autonoma di Trento”.
During the preparation of this work, the author(s) used ChatGPT and Paperpal in order to: Grammar and spelling check, paraphrase and reword. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. https://www.sciencedirect.com/science/article/pii/S0164121221001643. doi:https://doi.org/10. 1016/j.jss.2021.111067. [16] J. Baker-Brunnbauer, TAII Framework. In Trustworthy Artificial Intelligence Implementation:
Introduction to the TAII Framework, 2022. [17] G. Stettinger, P. Weissensteiner, S. Khastgir, Trustworthiness Assurance Assessment for High-Risk
AI-Based Systems, IEEE Access 12 (2024) 22718–22745. doi:10.1109/ACCESS.2024.3364387. [18] B. Li, P. Qi, B. Liu, S. Di, J. Liu, J. Pei, J. Yi, B. Zhou, Trustworthy AI: From principles to practices 55 (2023) 1–46. URL: https://dl.acm.org/doi/10.1145/3555803. doi:10.1145/3555803. [19] V. S. Barletta, D. Caivano, D. Gigante, A. Ragone, A rapid review of responsible ai frameworks: How to guide the development of ethical ai, in: Proceedings of the 27th International Conference on Evaluation and Assessment in Software Engineering, EASE ’23, ACM, 2023, p. 358–367. URL: http://dx.doi.org/10.1145/3593434.3593478. doi:10.1145/3593434.3593478. [20] OECD framework for the classification of AI systems (2022). [21] E. U. A. for Cybersecurity, M. Adamczyk, N. Polemi, I. Praça, K. Moulinos, A multilayer framework for good cybersecurity practices for AI – : security and resilience for smart health services and infrastructures, European Union Agency for Cybersecurity, 2023. doi:doi/10.2824/588830. [22] A. Singla, Machine learning operations (MLOps): Challenges and strategies 2 (2023) 333–340. URL: https://jklst.org/index.php/home/article/view/107. doi:10.60087/jklst.vol2.n3.p340. [23] M. Testi, M. Ballabio, E. Frontoni, G. Iannello, S. Moccia, P. Soda, G. Vessio, MLOps: A taxonomy and a methodology 10 (2022) 63606–63618. URL: https://ieeexplore.ieee.org/document/9792270/ ;jsessionid=84941428F3357304322A00024BA33EE0. doi:10.1109/ACCESS.2022.3181730. [24] M. Arnold, J. Boston, M. Desmond, E. Duesterwald, B. Elder, A. Murthi, J. Navratil, D. Reimer, Towards Automating the AI operations lifecycle, 2020. URL: http://arxiv.org/abs/2003.12808. doi:10. 48550/arXiv.2003.12808. [25] M. Mitchell, S. Wu, A. Zaldivar, P. Barnes, L. Vasserman, B. Hutchinson, E. Spitzer, I. D. Raji, T. Gebru, Model cards for model reporting, in: Proceedings of the Conference on Fairness, Accountability, and Transparency, ACM, 2019. URL: https://doi.org/10.1145%2F3287560.3287596. doi:10.1145/3287560.3287596. [26] M. Pushkarna, A. Zaldivar, O. Kjartansson, Data cards: Purposeful and transparent dataset documentation for responsible ai, 2022. arXiv:2204.01075. [27] L. Derczynski, H. R. Kirk, V. Balachandran, S. Kumar, Y. Tsvetkov, M. R. Leiser, S. Mohammad, Assessing Language Model Deployment with Risk Cards, 2023. URL: http://arxiv.org/abs/2303. 18190, arXiv:2303.18190 [cs]. [28] A. Celepija, A. Palmero Aprosio, B. Lepri, R. Kazhamiakin, AI product cards: a framework for code-bound formal documentation cards in the public administration, Data 38; Policy 7 (2025) e1. doi:10.1017/dap.2024.55. [29] D. Golpayegani, I. Hupont, C. Panigutti, H. J. Pandit, S. Schade, D. O’Sullivan, D. Lewis, AI cards: Towards an applied framework for machine-readable AI and risk documentation inspired by the EU AI act, 2025. URL: http://arxiv.org/abs/2406.18211. doi:10.48550/arXiv.2406.18211. arXiv:2406.18211 [cs], version: 1. [30] Understanding and Managing the AI lifecycle. AI Guide for Government (2023). URL: https: //coe.gsa.gov/coe/ai-guide-for-government/understanding-managing-ai-lifecycle/. [31] OECD, Governing with artificial intelligence: Are governments ready? (2024). doi: 10.1787/ 26324bc2- en. [32] M. Magnini, G. Ciatto, R. Calegari, A. Omicini, Enforcing fairness via constraint injection with
FaUCI (2024). [33] R. K. E. Bellamy, K. Dey, M. Hind, S. C. Hofman, S. Houde, K. Kannan, P. Lohia, J. Martino, S. Mehta, A. Mojsilovic, S. Nagar, K. N. Ramamurthy, J. Richards, D. Saha, P. Sattigeri, M. Singh, K. R. Varshney, Y. Zhang, AI fairness 360: An extensible toolkit for detecting, understanding, and mitigating unwanted algorithmic bias, 2018. URL: http://arxiv.org/abs/1810.01943. [34] F. Königstorfer, S. Thalmann, AI Documentation: A path to accountability, Journal of Responsible Technology 11 (2022) 100043. URL: https://linkinghub.elsevier.com/retrieve/pii/S2666659622000208. doi:10.1016/j.jrt.2022.100043. [35] C. Calluso, M. G. Devetag, Discrimination in the hiring process – state of the art and implications for policymakers, Equality, Diversity and Inclusion: An International Journal 43 (2024) 103–121. URL: http://www.emerald.com/edi/article/43/9/103-121/1214508. doi:10.1108/EDI- 10- 2023- 0340.