Applying Behavioural Science in the Italian Public Administration

10. Future directions for BPA

Abstract

This chapter describes how artificial intelligence is reshaping behavioural public administration. AI is already affecting how civil servants review evidence, manage people and make decisions. These AI tools may improve civil servant’s productivity and consistency; but they also introduce new behavioural risks such as automation and selective adherence. Behavioural research is needed to develop better practices within public administrations that help civil servants use AI safely and effectively.

Artificial intelligence

Among the most promising advances in behavioural public administration is artificial intelligence (AI), which civil servants are already weaving into the work of administration. AI is a powerful and rapidly evolving technology, with many of the latest efforts leveraging Large Language Models (LLMs). AI has the potential to improve government productivity, responsiveness and accountability (OECD, 2025[1]; 2024[2]; [3]) (OECD, 2025[4]; 2024[5]; [3]). As reflected in the objectives of Italy’s Piano Triennale per l'informatica della Pubblica Amministrazione (Three Year Plan for the Digitalisation of the Public Sector) (AGID, 2024[6]), governments can benefit from AI by streamlining bureaucratic processes, improving civil servants’ AI knowledge and using AI to enhance policymaking. However, government is expected to manage the risks of AI as well as harness its benefits. This is why OECD countries and others are guiding their public administrations to use AI in a trustworthy manner. For example, Italy’s (2025[7]) Agenzia per l’Italia Digitale has released a draft Bozza di Linee Guida per l’Adozione di IA nella Pubblica Amministrazione (Guidelines for the Responsible Use of AI in the Public Sector) to involve stakeholders in setting effective approaches and safeguards. Governments are already using AI within the themes of this report as well as surfacing new, untested opportunities to apply AI.

Governments are using AI for behavioural public administration

Civil servants and researchers around the world are already building AI tools to support policymaking, public services and internal processes. The OECD (2025[1]) catalogued these use cases in its report Governing with Artificial Intelligence: The state of play and way forward in core government functions. Use cases described in that report support the themes of behavioural public administration, including:

• Decision-making prompts with AI. For example, AI can remove noise – unjustified variation in human decision making – leading to more predictable and accurate decisions (Du, 2023[8]). The potential for AI systems to assist in making data-driven decisions is leading to its adoption across organisations, including within the public sector. AI can identify and address errors and noise that distort human judgment across government (Mills, Costa and Sunstein, 2023[9]).

• Evidence-informed policy that harnesses AI for research. For example, World Bank evaluators have used AI to double the number of studies they can review to inform programme decisions that affect government operations (Bohni Nielsen, Mazzeo Rinaldi and Petersson, 2024[10]). Some governments also use Polis, an open-source civic engagement tool, to understand citizens’ views on many topics (OECD, 2025[1]). This helps civil servants overcome the Illusion of Similarity – where people overestimate how representative their own views are of the population – by providing fast, accessible evidence of citizens’ views.

• People management that involves AI as an assessor. In Singapore, a government agency used AI to process over 3 000 job applications, saving 177 days of staff time (OECD, 2025[1]). In the UK, His Majesty’s Revenue and Customs Agency used an AI-based recruitment platform to automate recruitment tasks for junior roles, such as assessing simple questions against a competency framework. In Sweden, a municipal government developed an AI bot to perform blind interviews, removing superfluous details like their age, sex and appearance. AI assessors may help mitigate in-group favouritism – the tendency for people to prefer others who share traits with themselves.

• Sludge can be identified with the help of AI. Current methods for finding sludge rely on Sludge Audits – a structured assessments to identify, prevent, and reduce unnecessary frictions (OECD, 2024[11]). Public administrations may benefit from conducting sludge audits on the use of AI tools, enabling them to address the barriers that may limit AI adoption. Sludge audits may also explore using AI as a tool within the sludge audit itself to rapidly review online processes and construct personas for users who may find the experience especially burdensome.

Box 10.1. Case study: Using AI to Support Sludge Reduction in People Management in Canada

AI tools, such as LLMs and chatbots, have the potential to support Sludge Audits by mapping user journeys, identifying friction points, and streamlining processes. This use case was piloted by a team within the Treasury Board of Canada Secretariat to improve the employee Transfer Out process. The purpose of the Transfer Out is to facilitate a smooth transition for civil servants as they move from one Government of Canada department or agency to another, managing the administrative, financial, and information management aspects of their departure.

Using the Sludge Audit methodology, the team learned that the Transfer Out experience across key actors typically requires between 67 to 78 different steps across 10 key behaviour stages or phases, over 4 hours of active time completing the process, and 88 hours of waiting. Responsibilities during the Transfer Out were at times unclear, instructions for what to do were hard to find or confusing, and users did not always have access to the systems they needed to complete each step. Managers often played a key role in the Transfer Out process but at times lacked the capacity or knowledge to fully support their employee.

To better support both employees and managers, the research team proposed developing an AI chatbot to promote more timely and efficient information retrieval and better decision-making. They also suggested that a chatbot would also help alleviate administrative burden for HR staffing teams while offering a long-term solution for managing much of their involvement in the Transfer Out process.

In addition to providing critical insights, the project team shared that Sludge Audits are also a useful tool to establish benchmarks for key metrics that could assess the impact of this type of AI solution, including task completion times, manual intervention rates, and user satisfaction. They are now in the process of formulating an accompanying framework for assessing the impacts of AI interventions in a people management context.

Source: Based on information provided by the Treasury Board of Canada Secretariat.

More research is needed to use AI effectively

AI is a rapidly developing technology whose use-cases are still being explored. There are use-cases where AI aligns with the theory of behavioural public administration but has not been evaluated in practice. Researchers have built AI tools that could support civil servants but have not yet been used. Civil servants may also need to monitor for and address emerging risks as AI is increasingly used in government. These future needs and opportunities include:

• Policy design. Civil servants may develop new cognitive errors and distortions if they use AI to research, develop and design policy. For example, civil servants may develop automation bias – a tendency to defer to AI automatically without adequate scrutiny, as well as selective adherence – a tendency to follow AI-outputs when they confirm civil servants’ pre-existing beliefs (Alon-Barkat and Busuioc, 2024[12]). This poses a risk of exacerbating existing cognitive errors and may need to be addressed as part of public administrations’ strategies for promoting better decision making.

• Evidence-informed policy, which may need to adapt to a future where research is increasingly conducted, analysed and synthesised with AI. For example, a study of Boston Consulting Group researchers found that those who used AI were, on average, 25% faster and completed 12% more tasks compared to a control group without AI; however, on tasks beyond the capabilities of AI, researchers who worked without AI assistance made fewer mistakes, suggesting that researchers who used AI were over-reliant on its help (Dell’Acqua et al., 2023[13]). Public administrations may need to develop clear guidelines for when to use, and not use, AI as well as how to validate AI outputs as part of evidence and evaluation.

• Recruitment can benefit from including an AI assessor in the process. It’s possible that AI may assess candidates more objectively than a human, but this is not guaranteed because AI’s judgement can suffer from skewed training data, resulting in adverse outcomes. AI outputs can also be influenced, helpfully or harmfully, by the prompts of its human operators. Therefore, it’s essential to evaluate AI’s performance against human assessors whenever it’s deployed in a recruitment workflow, and to ensure human-in-the-loop for all high-risk or high-impact use cases. One EU study testing how human oversight could reduce subjective, AI-based hiring assessments found that participants were as likely to follow the guidance as a fair vs unfair AI. Human oversight alone did not lead to fair outcomes (Gaudeul et al., 2025[14]).

Box 10.2. Case study: A promising toolkit to counter AI misuse in the UK

There is a risk that civil servants could potential overly defer judgement to or misuse AI systems in their work, potentially inadvertently. For example, they may develop automation bias – a tendency to defer to AI automatically without adequate scrutiny (Alon-Barkat and Busuioc, 2024[12]). To combat this risk, the UK Cabinet Office developed a toolkit to help teams use AI tools effectively. The toolkit advises, among other things, behaviourally informed practices to combat misuse and bias, including how to: form diverse groups to review and challenge AI outputs, identify hidden behavioural and organisational risks as well as prioritise and track them, create risks mitigation approaches, and monitor outcomes for bias.

The Cabinet Office tested the toolkit through an LLM called “Assist” used by more than 200 government teams. They found that risks such as staff blindly following AI-generated advice or misunderstanding its capabilities were common but preventable. Including behavioural questions into the design and deployment process could help teams address the different challenges before they cause issues. The toolkit offers an easy, replicable and low-cost way for public administrations to avoid unintended consequences and scale AI tools with greater safety and confidence.

Source: (Cabinet Office, 2025[15]).

References

[7] AGID (2025), Bozza di linee guida per l’adozione di IA nella pubblica amministrazione, https://www.agid.gov.it/sites/agid/files/2025-02/Linee_Guida_adozione_IA_nella_PA.pdf.

[6] AGID (2024), Piano Triennale per l’informatica nella PA, https://www.agid.gov.it/it/agenzia/piano-triennale.

[12] Alon-Barkat, S. and M. Busuioc (2024), “Public administration meets artificial intelligence:  Towards a meaningful behavioral research agenda on algorithmic decision-making in government”, Journal of Behavioral Public Administration, Vol. 7, https://doi.org/10.30636/jbpa.71.261.

[10] Bohni Nielsen, S., F. Mazzeo Rinaldi and G. Petersson (2024), Artificial Intelligence and Evaluation, Routledge, New York, https://doi.org/10.4324/9781003512493.

[15] Cabinet Office (2025), The Mitigating ‘Hidden’ AI Risks Toolkit, https://www.gov.uk/government/publications/a-human-centred-approach-to-scaling-and-de-risking-ai-tools/the-mitigating-hidden-ai-risks-toolkit-html.

[13] Dell’Acqua, F. et al. (2023), “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.4573321.

[8] Du, M. (2023), “Machine vs. human, who makes a better judgment on innovation? Take GPT-4 for example”, Frontiers in Artificial Intelligence, Vol. 6, https://doi.org/10.3389/frai.2023.1206516.

[14] Gaudeul, A. et al. (2025), The Impact of Human-AI Interaction on Discrimination, Publications Office of the European Union, https://doi.org/10.2760/0189570.

[9] Mills, S., S. Costa and C. Sunstein (2023), “The Opportunities and Costs of AI in Behavioural Science”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.4490597.

[1] OECD (2025), Governing with Artificial Intelligence: The state of play and way forward in core government functions.

[4] OECD (2025), Governing with Artificial Intelligence: The state of play and way forward in core government functions, OECD Publishing.

[2] OECD (2024), Artificial Intelligence Papers: Governing with Artificial Intelligence.

[11] OECD (2024), Fixing frictions: ’sludge audits’ around the world: How governments are using behavioural science to reduce psychological burdens in public services, OECD Publishing.

[3] OECD (2024), G7 Toolkit for artificial intelligence in the public sector, OECD Publishing, https://doi.org/10.1787/421c1244-en.

[5] OECD (2024), Governing with Artificial Intelligence: Are governments ready?, https://doi.org/10.1787/26324bc2-en.