This chapter maps the main trends in the deployment and anticipated future developments of AI in Italian financial markets and the broader financial sector. It is primarily based on an OECD survey of the Italian financial industry conducted in Q2 2025, which received 450 responses. Additional resources include a supervisory questionnaire directed at Italian financial authorities, as well as additional public sector and industry engagement through OECD-led roundtables, workshops, bilateral meetings, and desk research. The chapter presents key findings on the current deployment of AI in Italian financial markets, the adoption of governance frameworks for AI technologies and the main regulatory and non-regulatory constraints perceived by industry participants.
Artificial Intelligence in Italian Financial Markets
Abstract
This chapter presents and analyses information gathered by the OECD regarding the deployment of AI in Italian financial markets. It covers the entirety of financial market activity in Italy, including the underlying infrastructure and the overall value chain – covering both primary and secondary financial markets. Beyond the focus on financial market activity, the report also covers wider finance activity including banking and insurance.
The analysis is based on the results of the OECD survey of the Italian financial industry, which was developed in consultation with Banca d’Italia and the European Commission’s Reform and Investment Task Force (SG REFORM) and in collaboration with the other Italian competent authorities. A total of 450 responses were received, representing a 49% response rate across banking, asset management, insurance, pension funds and other financial-market-related sectors (see Annex A for the survey methodology).1 Additional insights were obtained through virtual bilateral meetings and online roundtables with selected market participants from banking, insurance, asset management and pension funds (see Annex B for a list of participants).
The conceptual structure is developed on the basis of the structure of the industry survey, which focussed on several key dimensions, namely: profiles of survey respondents (including technological capabilities and volume of investment into AI), current uses of AI (including AI use cases and benefits of AI use), future use of AI, governance frameworks (including operational risks and cyber-threats) and constraints to the wider use of AI (including regulatory and non-regulatory factors). The same structure is followed in the sub-headings of the chapter, reflecting the key themes of the survey (see Annex A).
Furthermore, the conceptual structure of the report is guided by relevant OECD standards, notably the OECD AI Principles (OECD, 2019[1]), the G20/OECD High-Level Principles on Financial Consumer Protection (FCP Principles) (OECD, 2022[2]), the OECD Recommendation on Financial Literacy (OECD, 2020[3]) and the G20/OECD Principles of Corporate Governance (OECD, 2023[4]), while also incorporating key principles and standards established by other international organisations and bodies. Table 1.1 presents the structure of the report by chapter and section.
1.2.1. Deployment of AI solutions across the Italian financial system
Overall, 39% of survey respondents currently deploy AI in their activities. Financial market participants such as asset managers, wealth managers and investment advisors have adoption rates near 30%. Securities investment firms show lower adoption at 22%. No broker-dealer or investment adviser reported using AI, although these categories have smaller sample sizes. Among major sectors, insurance shows the highest AI deployment at 70% of respondents. Banks have a 59% adoption rate, while for pension funds it stands at 10% (Figure 1.1). AI penetration in Italian financial markets is likely to be higher than sector-level figures, as institutions operating across multiple sectors often include financial market activities. Similarly, AI deployment by asset managers may relate to back-office or administrative functions rather than investment processes.
Note: Percentages are calculated based on the number of respondents per sector (shown in brackets on the Y axis). This question was mandatory, with 450 responses received. Although most responding mutual banks use AI, only approximately 11% of the licensed institutions in this category answered the survey (Federcasse, 2025[5]).
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
The analysis includes all main sectors of financial activity (e.g. banks, asset managers, insurers and pension funds), to provide a comprehensive view of the finance industry. The number of financial sector entities that responded to the survey represents 49% of the total number of supervised financial sector entities (including foreign firms with activity in Italy supervised by Banca d’Italia) (Figure 1.2).
National surveys across OECD countries also show widespread AI deployment in finance. A survey in Switzerland with 400 respondents found that 50% of institutions use AI or have initial applications in development, while a further 25% intend to adopt it within three years (FINMA, 2025[6]). In the United Kingdom, a survey with 118 respondents revealed that 75% of financial firms deploy AI (BoE/FCA, 2024[7]). A Bank of Japan study, involving 155 financial institutions, reported an AI usage rate hovering around 60% (Bank of Japan, 2024[8]). A survey in Finland with 83 respondents showed that 73% either already use AI solutions or plan to implement them within the next two years (FIN-FSA, 2025[9]). In France, a survey with 100 respondents showed that 90% use or plan to use AI in the short term, while 54% of respondents reported deploying AI use cases in production (AMF, 2026[10]). Asian financial firms are also testing and deploying AI with varying breadth and intensity (OECD, 2025[11]).
Note: This chart shows the share of firms in each sector that responded to the OECD project survey, among the list of entities provided by Banca d’Italia and other Italian financial authorities. Banca d’Italia’s list of financial firms includes both Italian firms and foreign firms active in Italy. Financial market infrastructure is given a 100% rate as the four firms active in Italy all responded to the survey. The “overall share of respondents” reflects the number of survey respondents, as a share of the total validated and consolidated mailing list of 917 entities.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.2.2. Types of AI use cases in the Italian financial system
The survey offered 52 purposes, aggregated into 23 macro‑areas for figures and analysis. To enhance clarity, these macro‑areas were further divided into 17 applications and 6 methods. Applications represent AI outputs that directly support business outcomes, such as human resources, sales, modelling and post-trade. Methods represent AI outputs used across processes that indirectly lead to business outcomes, such as coding or data analysis. This section begins with macro‑area analysis, followed by a deeper review of individual purposes. Annex A provides details on grouping methods and purposes.
Data analytics and output generation is the most common method for current and planned AI use (Figure 1.3). Respondents reported 379 current uses and 761 planned uses over the next three years. This reflects the broad applicability of data analytics across sectors and business models. Other methods, such as translation and coding, also show strong adoption and are expected to grow steadily over the next three years.
Fraud detection and prevention is the most widely used application and is expected to remain dominant through 2028. Financial market-specific applications receiving the most responses include asset allocation and trading strategies. Asset allocation covers portfolio management, investment research, robo‑advice, financial advice and market analysis. Trading strategies include underwriting, IPOs, algorithmic trading, hedging, predictive analytics, forecasting and market-making. Respondents most frequently reported using AI for predictive analytics (21%), market analysis (17%) and investment research (11%). Few respondents reported current or planned AI use for new product development or post-trade processing, including P&L and reconciliations.
These results suggest AI is primarily used for business functions outside core financial market activities. This may reflect the respondent population, which includes institutions with limited market exposure, and a cautious approach to AI adoption. Firms appear to prioritise established applications such as data analysis and fraud prevention, while riskier, less tested purposes remain secondary.
Across OECD countries, financial institutions increasingly use AI for customer relations, marketing, process automation, and back-office operations, aiming to boost productivity, cut costs, and strengthen risk management. In the United States, AI spans nearly all financial functions (U.S. Department of Treasury, 2024[12]), while UK firms focus on internal process optimisation, cybersecurity, and fraud detection (BoE/FCA, 2024[7]). Swedish and Finnish surveys highlight summarisation, translation, and automation (Finansinspektionen, 2024[13]; FIN-FSA, 2025[9]), whereas Japanese firms prioritises customer relations and targeting (Bank of Japan, 2024[8]). Dutch banks report AI use for creditworthiness and fraud prevention (DeNederlandscheBank/AFM, 2024[14]). Similarly, respondents to the French survey indicated internal applications, such as productivity tools and internal assistants as the most common use cases (86% of respondents), with no specific application to financial activities (AMF, 2026[10]). IOSCO notes rapid advancements and growing interest, with applications in trading, customer interaction, and internal operations (IOSCO, 2025[15]). IMF analysis shows current AI trends in capital markets follow machine learning practices, with broader applications expected in the medium term (IMF, 2024[16]). Firms also report AI use in robo‑advising, algo-trading, research, sentiment analysis, and compliance monitoring, alongside experimentation in pre‑ and post-trading activities (OECD, 2024[17]).
Note: This figure shows the number of times respondents reported using (or expecting to use) AI for various purposes, classified into “Methods” (Panel A) and “Applications” (Panel B). The survey question listed 52 “purposes” for AI use and was multi-choice. Panel B only shows the ten most commonly reported Applications. See Annex A for more information on the methodology used. Values for 2024 are based on 174 respondents, and values for the next three years are based on 360 respondents. The question on current purposes was mandatory, while the question for future purposes was not.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.2.3. Breakdown of AI use cases by financial sector
Among financial market respondents using AI, asset managers are strongly experimenting and deploying this technology. They reported nearly 1 000 AI use cases in experimentation or production, second only to banks, which reported over 2 500 use cases. Banks report more use cases in production and almost as many in development or experimentation as all other sectors combined. Banks also lead in average use cases per respondent, with insurance and reinsurance companies also reporting high averages.
Overall, aggregate use cases in production exceed those in development or experimentation by nearly 1 000. This difference is most pronounced in banking, while other sectors report similar numbers in both categories. This may indicate that banks invest more in viable use cases or are further ahead in AI implementation than other sectors.
The partition of number of AI use cases in experimentation per company size highlights the importance of banks and insurers in AI testing in Italy. These types of firms are strongly represented among large firms, and they have the highest volume of AI use cases in development, with 54%. Small firms report more participation in use cases than medium-sized firms. This higher representation of small firms reflects their greater presence among survey respondents, with 105 small firms compared to only 65 medium-sized firms (Figure 1.4). It is important to note that there are an additional 1 059 use cases in experimentation which were reported by firms that did not disclose their headcount.
Concerning use cases in production, the OECD survey indicated 400 cases (47%) deployed by large firms, 117 (14%) by medium-sized ones, 321 (38%) by small firms, and 7 (1%) by micro enterprises, with another 1 939 use cases reported by companies that didn’t disclose their number of employees. Similar shares were found in the survey in France, where 51% of entities using AI were classified as large firms, while among the group that reported a lack of AI uses, 57% were small businesses and 28% were micro‑enterprises (AMF, 2026[10]).
Note: This figure shows the aggregated number of use cases reported by respondents, categorised by the size of responding firms. Firm size is calculated based on the number of employees disclosed by the responding firm. Firms that did not disclose the number of employees are not included in the figure. This question was mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.2.4. Role of third-party AI models and services
Firms face a strategic choice between purchasing AI models or developing proprietary ones. Large players often adopt a hybrid approach, building in-house models for competitive areas while relying on third-party solutions elsewhere. Third-party based AI is most common in data analytics and content generation, with strong current and expected use. Other frequent purposes include internal process optimisation, text generation, internal GPTs, fraud detection, translation and customer support. Expected growth is highest for fraud detection, asset allocation and targeted marketing. Specialised financial market tasks, such as post-trade processing or trading P&L, show limited reliance on third-party AI, reflecting a preference for internal development in high-risk areas.
Cloud services are the main method of AI deployment, reported by 74% of respondents. The fact that almost three-quarters of respondents depend on third party cloud services highlights the importance of robust AI governance frameworks. Other frequently used third-party services include GPAI model implementation (39%) and model implementation support (25%). Data acquisition services are used by 16%, while 9% of firms operate AI use cases without any third-party involvement (Figure 1.5, Panel A).
The most widely reported third-party provider was cited in 29% of responses. The second, third and fourth providers are reported by 19%, 10% and 8%, respectively (Figure 1.5, Panel B). A comparison may be drawn with similar results found in the survey conducted by the French AMF, where the top three AI service providers were non-European players, as reported by 33%, 13% and 11% of respondents (AMF, 2026[10]).
Firms adopt different approaches when selecting providers, influenced by company policy and size. Some evaluate existing cloud partners’ AI offerings while avoiding exclusive dependence, maintaining dialogue with multiple vendors to compare services. Certain companies disclosed in project bilateral meetings that they do not prioritise Italian data centres, assuming all EU-based centres meet sovereignty requirements.2
Larger firms often rely on Big Tech providers, complemented by smaller local vendors. Big entities establish backups for every engaged cloud provider, enabling migration if switching becomes necessary. One mid-sized company stated its choice depends on shared objectives and responsiveness. It prefers start-ups over Big Tech, believing smaller firms react faster and may provide source code when partnerships end.
During OECD consultations, global firms reported using vendor-provided tools, open-source solutions, and pre‑trained models in certain cases. As major vendors roll out new products, firms increasingly integrate and customise these tools to meet specific operational needs. Large vendors remain the preferred choice because their enterprise‑level products include features addressing privacy and security concerns. These solutions also offer integration within existing systems, making them attractive for financial institutions.
Note: Percentages in Panel A are calculated based on 174 firms declaring the use of AI. Respondents could select multiple answers. This question was non-mandatory. Panel B counts the number of references to each of the top four providers, as well as the references to all other providers combined. Respondents were asked to specify the top three service providers. A total of 288 responses were provided. This figure makes no distinction between the first, second and third provider reported by respondents.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.2.5. Use of general-purpose AI models by financial institutions
Generative AI (GenAI) models3 are the most reported type currently in use or in development, cited by 84% of firms disclosing information in this section (Figure 1.6). This dominance likely reflects respondents’ high level of experimentation with language‑based GenAI models and the fragmentation of traditional non-GenAI models across several categories in this analysis. Machine learning, particularly classification and clustering techniques, follows with 56% of respondents, while natural language processing and text mining are reported by 48%.
GenAI also predominates in Finland, France and Sweden. The use of GPAI was the most widely reported category in surveys conducted in Finland and Sweden, followed by machine learning and rule‑based systems (FIN-FSA, 2025[9]; Finansinspektionen, 2024[13]). Similarly, in France the use of GenAI was reported by 52% of respondents, followed by NLP and supervised learning (AMF, 2026[10]). The spike in the use of GPAI is relatively recent. As the Swedish survey shows, all GPAI use cases were taken into production between2 022 and 2024 (Finansinspektionen, 2024[13]).
Other disclosed AI model types include machine and deep learning for regression and forecasting (37%), logic and knowledge‑based approaches (25%), agentic AI (18%) and rule‑based expert systems (17%). In the “other” category, respondents listed supervised and deep learning models such as gradient boosting, random forests, deep neural networks and vision transformers. They also mentioned reinforcement learning, generative adversarial networks for weather nowcasting and synthetic data, large language models for document classification and clustering, and tools like Google Vision API, Microsoft Copilot and ChatGPT.
Note: Percentages in this figure are calculated in relation to 174 firms that answered this question. This question was mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
Respondents deploying GPAI models mainly use non-customised licensed models, such as commercial LLM licences, accounting for 73% of responses. Tailored GPAI models for specific use cases represent 23%, while fine‑tuned models using proprietary data account for 19% (Figure 1.7).
Answers in the “other” category include tools from popular vendors, such as models on public cloud with private networks, private cloud with secured networks, proprietary on-premise models and private accounts. Some respondents report not using any third-party proprietary general-purpose AI models.
Many firms block open and freely accessible GPAI tools due to security concerns. Alongside technical blocking, companies monitor user access to such resources regularly. Respondents also roll out tools aligned with company data protection policies. For example, one firm uses a tool under a contract tailored to company-specific terms and conditions. This is supported by company-wide communications instructing employees on safe tool usage without disclosing confidential information.
Note: Percentages in this figure are calculated in relation to 135 firms that answered this question. The question was not mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
Data analysis and content generation are the highest-ranked method for GPAI use, both currently and in future plans (Figure 1.8, Panel B). Respondents report 215 current uses and 405 planned uses over the next three years. Translation is also popular, with 36 current mentions and 93 planned, followed by coding with 31 current and 68 planned uses. Macro‑areas with the fewest responses include new product development, post-trade processing, credit underwriting and pricing, profit and loss calculations in trading, reconciliations and pension contribution collection.
The outlook for GPAI applications in financial markets is significant. Although numbers remained modest in 2024, respondents indicate strong intentions for 2025‑2028. Asset allocation receives 111 mentions for planned GPAI use, while trading strategies receives 54. Respondents in asset allocation show greater enthusiasm for GPAI than for AI, ranking it as the most important GPAI application. Fraud detection and prevention remains a popular application for traditional AI (Figure 1.8, Panel A).
Globally, firms currently use or plan to use GPAI for internal operations, communication, and risk management, while fully automated end-to‑end applications remain in development (OECD, 2023[18]). Early applications include summarisation, translation, and context-sensitive information retrieval, mainly for internal, lower-risk purposes rather than customer-facing ones (IOSCO, 2025[15]). Authorities increasingly monitor GPAI use, as reflected in national surveys: 91% of Swiss AI users employ generative tools like chatbots (FINMA, 2025[6]); 78% of US firms and 74% of Finnish respondents report GPAI deployment (U.S. Department of Treasury, 2024[12]; FIN-FSA, 2025[9]). In Japan, 60% of institutions use GPAI for summarisation, proofreading, translation, and operations, with performance rated as meeting or exceeding expectations in all business fields except for “search for information on internal rules” category (Bank of Japan, 2024[8]).
Note: This figure shows the total number of times respondents reported using (or expecting to use) GPAI for various purpose, classified into “Applications” (Panel A) and “Methods” (Panel B). The survey question listed 52 “purposes” for AI use and was multi-choice. Panel B only shows the 10 most commonly reported Methods (total of 18 Methods). See Annex A for more information on the methodology used. Values for 2024 are based on 174 respondents, and values for the next three years are based on 360 respondents.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
In total, GPAI use cases in development or experimentation total 1 042, compared to 713 in production (Figure 1.9). A relatively high proportion of respondents report only one or two GPAI use cases in production, development or experimentation. A slight majority have one or two GPAI use cases. Additionally, 48% of respondents report three or more GPAI use cases in development or experimentation, compared to 40% in production. Notably, five respondents report over 70 GPAI use cases in development or experimentation, while only one reports such a number in production.
Industry feedback indicates firms reduce GPAI use cases when moving from development to production compared to non-GPAI models. This may reflect market enthusiasm for GPAI, its higher risk profile, and its status as a newer technology still in experimentation. Firms also highlight challenges in progressing from proof of concept to full deployment. Several disclosed that initial enthusiasm during proof-of- concept phases often leads to perceived underwhelming returns at deployment, resulting in fewer production use cases than originally planned.
Note: This figure shows the share of responding firms based on the number of GPAI use cases they have in development and experimentation (Panel A), and in production (Panel B). Respondents are grouped by number of use cases (1‑2, 3‑10, 11‑50 and over 50). 139 firms reported at least 1 use case in development and experimentation, and 145 firms reported at least one use case in production. Respondents who reported no use cases are not included in the figure.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.2.6. The benefits of AI deployment
Italian financial institutions already report tangible benefits from deploying AI. Three‑quarters of firms using AI report operational efficiencies, while 62% mention productivity gains and 45% highlight improved internal processes. Respondents also cite coding efficiencies (33%), cost reductions (30%) and better translation (29%). Front-office benefits include improved decision making (49%), new analytical insights (22%), predictive analytics and marketing or sales improvements (both 17%). Respondents see faster and less costly operational tasks as the most noticeable benefit of AI. These appear easily attainable at the current AI stage for most finance companies, requiring little or no tailored models or expensive data training (Figure 1.10).
Financial market-specific benefits remain rare. Fewer than 3% of respondents report improvements in portfolio allocation, trading strategies, or execution efficiency. This contrasts with frequent AI use in asset allocation and trading strategies, suggesting a lack of technology maturity for these purposes. Some benefits may have been categorised under broader operational gains, as AI often supports isolated steps rather than full integration. Competitive advantage for financial market participants may also arise from internal efficiencies, freeing resources for tailored AI models. Improved decision making, reported by half of firms, could extend to market activities as trust in AI outcomes grows. Additionally, as resources are freed from routine internal tasks, such as administrative processes, they may be allocated to developing AI models tailored to financial market needs.
These gains are expected to grow, with OECD analysis showing that the finance sector could achieve 12% productivity growth over the next decade across G7 economies from AI adoption (Filippucci et al., 2025[19]). The results highlight variations in AI-related labour productivity gains across the G7, with the United States, the United Kingdom and Germany presenting the highest expected productivity gains. Such a scenario may be influenced by the composition of national economies, with countries that have a higher concentration of AI-exposed industries experiencing greater productivity gains (Filippucci et al., 2025[19]). UK financial sector participants most frequently identify benefits such as data and analytical insights, AML monitoring, fraud prevention, and cybersecurity (BoE/FCA, 2024[7]). In Finland, respondents emphasise improvements in internal processes, enhanced customer experience and support, alongside significant cost reductions, as the most valuable gains (FIN-FSA, 2025[9]). Respondents to the French survey reported data analysis, cost reduction and internal process improvement among the main benefits of AI deployment (AMF, 2026[10]).
The discrepancy between usage frequency and perceived benefits may indicate very early stages of technological adoption for these purposes. Another explanation could be that respondents classify more efficient or less costly issuing, trading, clearing or settling under generic categories such as operational efficiencies and productivity gains. AI may also support specific steps of issuance, trading, clearing or settlement without full integration across the process, contributing to perceived benefits relevant to financial markets.
Note: Percentages in this figure are calculated based on 174 firms declaring the use of AI. Responses are broken down into the top 10 benefits (Panel A) and other benefits (Panel B). Benefits reported by fewer than 5% of respondents are not shown. This question was non-mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.2.7. Use of open-source models and components
Among Italian firms deploying AI, 39% do not use any free or open-source components.4 For those that do, open-source models show slightly higher penetration than ML development libraries, both near 35% of respondents (Figure 1.11). The most reported open-source components in the “other” category were OpenAI tokens. Several companies disclosed that although they experiment with open-source components, deployment limitations arise from the need for substantial internal expertise to ensure efficient implementation. Firms also remain cautious, citing the high level of risk involved. At the same time, certain actors expect significant reliance on open-source components in the future.
Note: Percentages in this figure are calculated in relation to 174 firms declaring the use of AI. Respondents could select multiple answers. This question was mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
Among Italian firms responding to the question on training data types, 63% use internal data for training or fine‑tuning AI applications. Additionally, 44% rely on publicly available data. A further 16% disclose using third-party licensed data, while 13% employ acquired datasets for training or fine‑tuning purposes (Figure 1.12). In comparison, the French survey found that 40% of AI tools use only internal company data, while the rest takes into account public data, licensed commercial data or a combination of approaches (AMF, 2026[10]).
During OECD consultations, global firms reported that internal AI projects face significant obstacles due to restrictions on data usage, including licensing, ownership, and contractual limitations. Several firms disclosed abandoning proof-of-concept projects because extending data usage licences proved too costly. Other challenges involve obtaining agreements from every stakeholder before developing new applications – a process repeated for each separate project. One firm described a lengthy consent process for an AI application using bond issuance data with fragmented ownership across providers in the value chain, ultimately leading to project abandonment. Firms also stressed the need for efficient internal data governance structures to optimise data utilisation for AI development and manage third-party provider involvement effectively.
Note: Percentages in this figure are calculated based on the 139 firms that answered this question. This question was not mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.2.8. Challenges and approaches to explainability of AI models
Across jurisdictions, explainability remains a persistent governance challenge for financial firms and supervisors, especially with complex and non‑deterministic models, as opacity can undermine accountability and consumer trust (OECD, 2024[17]; IOSCO, 2025[15]). Firms and authorities increasingly require at least partially explainable outcomes in addition to keeping “human‑in‑the‑loop,” while insurers often favour simpler, more interpretable models in higher‑impact use cases (OECD, 2024[17]; EIOPA, 2024[20]). Governance frameworks now elevate explainability as a focal control area. For example, participants in the Swiss survey highlighted that explainability is a key component of governance frameworks, alongside data protection, IT and cybersecurity (FINMA, 2025[6]). Firms also report struggling to balance certainty and explainability with black‑box performance and call for clearer sector‑specific guidance on what must be communicated to customers (OECD, 2024[17]). In response, many adopt harm‑mitigation and impact‑based assessments, tighter data‑quality and provenance controls, and cross‑functional governance to manage non‑determinism and deployment risks (IOSCO, 2025[15]).
Survey results in Italy show high variation of explainability methods in use, indicating similar adoption across firms. Each of the explainability methods presented to respondents was chosen by a comparable share of participants, namely ex-ante techniques (49%), followed by ex-post model-agnostic techniques (46%), ex-post model specific techniques (42%) and in-process techniques (40%) (Figure 1.13, Panel A).
Most respondents (54%) use only one explainability method, highlighting the high degree of heterogeneity involved. Twenty-six per cent deploy two methods, and 9% use three. Eleven per cent employ all four types, representing firms with the highest transparency internally or to the market (Figure 1.13, Panel B).
For ex-ante techniques, respondents report implementing measures during design to improve explainability. These include clear and understandable features, prompt engineering, targeted training and supervised learning. For ex-post techniques, Shapley values (SHAP) are most frequently mentioned, while manual review and combined approaches appear less often.
One company disclosed prioritising inherently interpretable models whenever feasible. When complex “black-box” models such as deep learning networks or gradient boosting machines are required, the firm applies post-hoc explanation techniques such as SHAP, LIME or attention mechanism analysis. This example illustrates how firms combine multiple techniques to enhance AI model explainability.
Note: Percentages in this figure are calculated based on 81 firms that answered this question. This question was not mandatory. Respondents could select multiple answers. Panel A shows the number of respondents that selected each explainability method, while Panel B shows the number of respondents that selected one, two, three or four of the methods in Panel A.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.2.9. Level of action autonomy across AI models
Among Italian entities providing input on action autonomy, 75% operate under no‑action autonomy, meaning “human support”. A low-action autonomy or “human-in-the-loop” is observed in 35% of firms, while 16% report medium-action autonomy or “human-on-the-loop”. Only 8% of respondents indicate their AI use cases operate with high-action autonomy, described as “human-out-of-the-loop” (Figure 1.14). Companies deploy AI use cases with significant human involvement, consistent with their risk management frameworks.
Across jurisdictions, firms largely limit AI action autonomy to augmentation with human‑in‑the‑loop, reserving fully automated decision‑making for narrow, lower‑risk tasks; end‑to‑end autonomy remains mostly in development for financial‑market uses (OECD, 2023[18]; IOSCO, 2025[15]). “Human in the loop” was the most reported security measure in the French survey (AMF, 2026[10]). In capital markets, autonomy appears in pockets such as adaptive algorithmic trading that can identify and execute trades without human intervention, yet asset‑management processes still use AI to inform choices rather than decide, and fully AI‑based investment processes are marginal (OECD, 2021[21]; ESMA, 2023[22]; 2025[23]). Banking deployment is cautious, with GPAI in pilots or sandboxes focussed on productivity rather than autonomous decisions (EBA, 2024[24]). Insurance firms emphasise interpretable models under human oversight, with high‑impact uses requiring Management/Executive Board approval, reinforcing constrained autonomy (EIOPA, 2024[20]). Supervisory use of GPAI is supportive, not decisive; decision‑making remains human‑led (Prenio, 2025[25]). Governance guidance stresses defined human intervention points for high‑risk scenarios (e.g. credit scoring, algorithmic trading) and harm‑mitigation/impact assessments to manage non‑determinism- further curbing autonomous operation (Crisanto et al., 2024[26]; IOSCO, 2025[15]).
Note: Percentages in this figure are calculated based on 142 firms that answered this question. This question was not mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets. The four levels of action autonomy are based on the OECD Framework for the Classification of AI systems (2022[27]), https://www.oecd.org/en/publications/oecd-framework-for-the-classification-of-ai-systems_cb6d9eca-en.html.
Only 10% of firms using AI reported employing secondary or challenger models. Among this small group, one‑third are banks. Firms stated that secondary or challenger models serve multiple purposes. They evaluate outputs from primary models, enhance performance, cross-check accuracy, and support specific tasks. They also validate models, act as backups if primary models fail, assist during development, complement staff controls, and help identify the most effective model for a given application. Respondents not using secondary or challenger models usually cited limited AI adoption within their organisation, implementation at group level, or outsourcing AI solutions to external providers. Some respondents rely solely on commercially available AI systems, including ChatGPT, Microsoft Copilot, Leonardo.ai, and other externally developed and managed tools.
1.3.1. Governance structures and AI oversight by functions
Governance rules form part of the existing financial sector regulations that also apply to the deployment of AI technologies. Several respondents to an OECD survey conducted in 2024 noted that market participants’ willingness to adopt AI often depends on the robustness of governance frameworks, for example in the case of risks related to the use of AI in credit decision making. Survey respondents frequently identified weak governance as a key risk area (OECD, 2024[17]).
Survey results indicate that the differences among the top choices of AI governance structures are relatively minor, suggesting the absence of a single preferred framework. Additionally, rather than relying on a single structure, a large number of respondents tend to combine multiple tools and mechanisms as part of their overall governance structure for AI. Of the 450 survey respondents, 27% rely on AI strategy, guidelines, principles and/or codes of conduct as the governance framework for their use of AI applications (Figure 1.15). A relatively similar share of respondents disclosed the use of cyber-security and operational risk frameworks and model governance frameworks under human oversight (24% and 23% of companies, respectively). One in five respondents use data governance frameworks, centred around concepts of privacy, stewardship and ownership. Other governance frameworks, controls or processes are put in place by 19% of respondents. Notably, establishment of an explicit AI governance framework was disclosed by 16% of respondents.
As respondents could select multiple options, many appear to combine various governance approaches depending on the specific area, for example, a group-level AI strategy complemented by cybersecurity and data governance‑specific frameworks. This may be explained by the wide range of implications that the use of AI has at the operational and risk management levels for the firms.
Among the respondents with no governance framework, it was noted that existing non-AI specific controls, such as IT policies and codes of conduct, are currently deemed sufficient. In several cases, governance arrangements are established at the group level rather than at firm level, while in others, relevant measures are still under development or consideration.
Governance structures vary across institutions. For example, larger groups opt for global monitoring systems featuring a mix of centralised and local governance, to ensure checks and controls are accounted for along the AI model development lifecycle. Stakeholders interviewed by the OECD acknowledged that the success of the implementation of such governance frameworks will depend on fostering adequate corporate culture under a strong leadership.5
Note: The figure shows the share of all question respondents (450), and the share of respondents categorised as financial market participants (124), who chose each answer. The financial market participants include asset managers, securities investment firms, investment advisors, wealth managers, broker-dealers, central securities depositories, central counterparties or exchange/multilateral trading facility. Respondents could choose multiple answers. This question was mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
Out of 174 respondents using AI, 52% reported having dedicated data science teams within their companies for AI development and/or deployment. Most of these firms operate AI teams only at company group level, representing 25% of respondents (Figure 1.16). Fifteen per cent maintain teams exclusively at firm level, while 8% have teams at both levels. AI teams typically consist of one to five employees. Some firms adopt a federated structure, where a central AI team allocates resources to smaller teams for tailored use cases.
Firms cited several reasons for not creating dedicated AI teams. These include organisational constraints linked to size and budget, reliance on external providers, and ongoing AI exploration at group level rather than firm level.
Note: Percentages are calculated using the number of firms that declared using AI and answered this question (161). Not every firm that declared using AI answered this question, which was non-mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
Across OECD countries, firms have adopted layered governance structures that integrate multiple frameworks, controls and processes to oversee AI deployment. For example, in the United Kingdom and Switzerland, most firms reported establishing dedicated governance strategies for AI (BoE/FCA, 2024[7]; FINMA, 2025[6]). In Finland, the majority indicated having an AI strategy, as well as AI-related ethical guidelines and codes of conduct. Notably, 94% of surveyed Finnish firms reported implementing an IT risk management system, which in most cases also encompassed AI (FIN-FSA, 2025[9]). In Switzerland, governance frameworks often focus on specific areas such as data protection, explainability, IT, cybersecurity, and risk management (FINMA, 2025[6]). In France, 72% of respondents reported implementation of internal AI-specific governance policy, mostly encompassing rules for the use of AI and data protection measures (AMF, 2026[10]).
Out of all survey respondents, 31% of firms disclosed assigning responsibility for AI-driven outputs within their AI accountability frameworks to business area users, while 26% of companies place it with executive leadership. A similar share (16%) of firms designated developers and other accountable persons from outside the firm as the responsible functions. Data science teams were the last commonly designated function category- (14% of respondents) (Figure 1.17). However, it is important to bear in mind that only a quarter of survey respondents reported having established dedicated AI teams. Among respondents who did not designate any specific responsible function, the most common explanations included the absence of current AI use cases, full outsourcing of AI models, or the allocation of responsibility at the group level.
Note: Percentages in this figure are calculated in relation to all 450 respondents to the survey. Respondents could choose multiple answers. This question was mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
Accountability mechanisms are also becoming more widespread in other OECD jurisdictions. In the United Kingdom, 84% of firms designated an accountable person for AI outputs, most often within the executive leadership functions (BoE/FCA, 2024[7]). Similarly, Finnish firms reported assigning responsibility for AI-related outputs to managerial and director-level positions (FIN-FSA, 2025[9]).
1.3.2. Risk management, operational resilience and cybersecurity
Half of the respondents reported the use of human oversight (human-in-the‑loop) as a key safeguard to manage the risks of unintended AI activity (Figure 1.18). Other safeguards were significantly less commonly identified – incident management and reporting mechanisms were cited by 25% of respondents, output and input restrictions by 22%, and the use of back-up systems and failure monitoring and ampersand alert systems by 20%. Firms also cited data integrity checks/secure data pipelines (14%), as well as alert systems for flagging unusual/unexpected behaviour (14%) as part of their risk management frameworks. To address ethical risks, firms aim to ensure the fairness of AI models. Particularly in use cases involving external stakeholders (e.g. customers), some firms avoid the use of sensitive data in AI model development, such as gender or ethnicity.
Note: Percentages in this figure are calculated in relation to all 450 respondents to the survey. Respondents could choose multiple answers. This question was mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
Almost half (46%) of respondents have not implemented any specific safeguards to address emerging AI-specific cyber threats (Figure 1.19). While 23% of respondents have conducted initial assessments, these have not translated into any dedicated safeguards. A share of 21% of firms are operating with basic safeguards, while acknowledging a potential need for future improvements. Meanwhile, for 12% of companies, the comprehensive measures set in place are assessed as covering most of the identified threats. Only 3% of firms have implemented fully robust and continuously updated measures concerning AI-specific risks.
Note: Percentages in this figure are calculated based on all 450 respondents to the survey. Respondents could choose multiple answers. This question was mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.3.3. Level of understanding of AI technologies and talent needs
Operational level understanding of AI technologies was most widely reported for all operational staff, while complete understanding was most commonly identified for senior management. Among the operational staff, the majority were assessed as having either a limited understanding (29%) or only basic operational usage (28%) of the AI tools. For middle management, 32% of respondents indicated a partial understanding, while 26% are perceived to have a limited understanding. Although senior management, followed by board members, are perceived as those more knowledgeable in AI across different companies, such knowledge is still incipient. Senior management was reported to have a partial understanding by 33% of firms, and a limited one by 22% of respondents. Most respondents indicated that board-level comprehension is either partial (32%) or limited (26%) (Figure 1.20).
The identification of senior management and board members as the most knowledgeable function may be unexpected as these functions do not typically involve day-to-day technical aspects of AI operations and the use of AI tools. Notably, in the French survey, the level of understanding was also indicated to be the most developed at more senior levels (AMF, 2026[10]). This trend can be attributed to several factors. First, a quarter of respondents declared that executive leadership is the function responsible for AI outputs in their firms, ranking second after business area users, which could lead to enhanced exposure to AI technologies. Second, as seen throughout the questionnaire and disclosed by firms in bilateral meetings, the level of enthusiasm of firms regarding the opportunities presented by AI technologies may translate into a higher level of engagement by upper management than with other technologies, particularly in terms of guiding the development of company-wide AI strategies. A third factor may be that the interpretation of what “understanding” of AI technologies means is likely to differ depending on the function. As senior management may not deal with AI in day-to-day business operations, they may not be exposed to the technical aspects of such tools to the same extent than data science teams developing and deploying AI models. This could contribute to an enhanced perception of the understanding of AI technologies compared with all operational staff, whose interpretation of what a complete understanding of AI implies may be different.
Note: Percentages in this figure are calculated based on all 450 respondents to the survey. Respondents had to choose one level of understanding per function level. This question was mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
In this context, it is worth noting that out of 174 respondents using AI, 52% reported having dedicated teams within their companies for AI development and/or deployment. Thus, the establishment of dedicated AI teams is not yet widespread in the Italian financial sector (Figure 1.16).
Plans to provide AI-related training to employees in the future have been reported by 44% of respondents. Currently, 41% of responding firms have offered such training at the general awareness level. In contrast, only 14% have provided specific advanced training for AI developers and data scientists. Notably, 28% of companies indicated that no AI-specific training has been provided to date (Figure 1.21). Since most firms either do not offer AI training or provide only general awareness-level sessions, it follows that AI training tailored to specific areas, such as particular financial market activities, is not yet widely available.
Note: Percentages in this figure are calculated based on all 450 respondents to the survey. Respondents could select multiple answers. This question was mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.4.1. Regulatory constraints
Survey respondents were asked to select from a list of 26 factors that constitute regulatory constraints to the adoption of AI use cases within their firms, and to classify them as either a large or small constraint. The 26 factors included in the survey were classified into eight categories in Figure 1.22 below. More details on the survey methodology may be found in Annex A.
Regulatory uncertainty and potential misalignment of rules are the most cited constraints to wider deployment of AI in finance in Italy, although they were mostly classified as small barriers. These were followed by regulations related to data protection and intellectual property, which were described by a significant portion of respondents as major barriers to broader AI adoption. Many respondents also highlighted operational and third-party regulations, as well as rules concerning governance, fairness, and market conduct.
Within the category relating to “clarity and alignment of regulations”, which was most often selected by respondents, one in five respondents cited a lack of regulatory clarity as either a large or small constraint to wider AI adoption. A significant share also identified the absence of supervisory guidance as a key limitation, with 8% of all respondents seeing it as a large constraint. Many firms reported challenges in defining what qualifies as an AI model, leading to confusion about the obligations associated with such a classification. Other concerns, generally viewed as small constraints, included conflicts with existing sector-specific rules, potential ex-post regulatory interventions, and a lack of alignment across jurisdictions.
Respondents underlined the role that the implementation of the AI Act will play in shaping their approach to AI deployment. The AI Act may be overlapping with the existing regulatory framework (e.g. CRR and DORA), which requires firms to review and integrate AI considerations into existing structures, increasing the cost and complexity of such compliance.
Note: Panel A: Absolute values calculated as sums of categories of constraints specified in Annex A. Respondents could select multiple answers. This question was non-mandatory; Panel B: Percentages were calculated based on all 450 respondents to the survey. This chart takes into account all factors under the “concerns related to clarity and alignment of regulations” category. Further information on survey methodology may be found in the Annex A. Respondents could choose multiple answers. This question was non-mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
Throughout the OECD consultations, certain global finance industry players identified a fragmentation of regulations as one of the key constraints to scaling up AI adoption. The regulatory areas of concern mostly related to data and privacy, followed by obligations related to consumer outcomes, especially with respect to ensuring fairness, explainability and ethical dimensions of AI use. An increasing use of third-party models and data is also an area of concern, where more regulatory guidance could be helpful.
Anecdotical examples from survey respondents show concerns that the plurality of bodies overseeing the implementation of the AI Act will lead to a highly bureaucratic implementation burden and possible conflicting guidance. There were also anecdotical concerns about ambiguities concerning definitions and scope of obligations. This uncertainty makes it challenging for firms to plan long-term AI strategies and assess the risks of their AI systems. Some survey respondents also expressed concerns over the risk of future ex-post interventions by regulators, once the AI Act and related guidelines are fully enforced.
Despite these constraints, most respondents do not see any major conflicts with existing sector rules or regulatory requirements. Instead, firms call for more co‑ordinated and proportionate regulatory guidance, tailored to the particular use cases of the financial sector. The absence of detailed supervisory guidelines increases uncertainty about how existing and future regulations will be interpreted and enforced. This is amplified in the case of cross-border operations, where inconsistent regulatory approaches may pose additional compliance issues at group level.
Constraints related to regulatory compliance are particularly burdensome for smaller firms, which claim that although they can navigate these regulations, their efforts are hindered by a lack of resources to invest in dedicated AI governance structures or training. As a result, some firms postpone adoption, rely on group-level solutions, or outsource AI development entirely.
Within the second most commonly selected category relating to “operational and third-party related resilience”, 16% of survey respondents identified cyber-resilience obligations as large constraints, while third-party risk management frameworks and rules on operational resilience were identified as such by approximately 10% of all respondents (Figure 1.23).
Certain firms disclosed that they are currently facing issues with the implementation of the Digital Operational Resilience Act (DORA) as they find it challenging to map all relevant ICT actors, given the level of complexity of their operating system which often spread across different company group levels (EU, 2022[28]).
Note: Percentages in this figure are calculated based on all 450 respondents to the survey. This chart takes into account all factors under the “operational and third party-related resilience” category. Further information on survey methodology may be found in Annex A. Respondents could choose multiple answers. This question was non-mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
1.4.2. Non-regulatory constraints
The survey asked respondents to select from a list of 39 potential non-regulatory constraints to the further deployment of AI, which were classified into eight categories in Figure 1.24 below. More details on the survey methodology may be found in Annex A.
Organisational, skills, and cultural factors were the most frequently cited category of non-regulatory constraints, although they were mostly classified as small barriers. These were followed by data-related constraints, which are evenly split as either large or small constraints. Other identified barriers include operational and business risks, and cost-related constraints. Less commonly mentioned were concerns related to ethics, compliance, and liability, as well as issues concerning market integrity and consumer or investor protection (Figure 1.24).
Non-regulatory constraints highlighted by industry in meetings include concerns related to data and skills gaps, third-party dependence, and the need for strategic alignment among stakeholders. Other major constraints include issues related to ensuring data and model implementation security, as well as challenges with change management within the firm.
Note: Absolute values calculated as sums of categories of constraints specified in Annex A. Respondents could select multiple answers. This question was non-mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
In the organisational, skills and cultural constraints category, a quarter of all survey respondents referred to limitations stemming from talent acquisition, reflecting difficulties in recruiting for the development and deployment of AI tools. Many firms report not adopting AI due to the current lack of use cases in their line of business, as identified by 16% of respondents. Internal factors such as limited AI expertise, unclear integration strategies, and organisational resistance to change also impede progress. Adoption is often hindered by competing priorities within the organisation, which affect the allocation of resources to AI projects (Figure 1.25). Concerns related to ethics and liability, as well as market integrity and consumer/investor protection issues, were less prominent in the reporting by supervised entities, possibly given the safeguards provided for by applicable regulation and the provisions of the AI Act.
Organisational and skills-related constraints may be acutely felt by financial market participants as the area demands substantive knowledge, combined with technical AI skills, which may be difficult to find. Furthermore, establishing internal use cases, for example in settlement or post-trade processing, may be challenging due to the current limited use cases at the global level.
As explained in the qualitative input, institutions often lack the necessary technical and analytical skills to deploy and monitor AI effectively, requiring training and cultural transformation. Sometimes company structure is not suited for facilitating collaboration across departments, such as IT, compliance, and risk, which is essential for the effective deployment of AI models. Some local companies are also limited in their AI exploration as they wait for the group-level assessment and issuance of guidance on whether and how to develop AI use cases.
Industry participants also disclosed in the project meetings that many organisations simply lack compelling AI use cases for their lines of business, meaning that introducing AI into already efficient operations could result in new problems instead of the expected performance enhancement. Interconnectivity between systems is perceived as a major issue, along with difficulties identifying the most effective use cases. Many firms disclose that rationalising AI deployment and investment is often difficult for the purpose of internal decision making. Concerns also exist around capital risk due to over-reliance on AI without adequate oversight.
Furthermore, as firms differ in levels of investment and cultural readiness, AI adoption is uneven. Major organisations have already shifted to cloud-based platforms to accommodate AI workloads, while certain entities rely on legacy IT for varying reasons, among them cyber-security. Other constraints to the deployment of AI from the company side may stem from the difficulty with quantifying the return on investment for AI-related project budget-planning. The issue of talent is also often raised, as more traditional institutions struggle with recruiting data scientists and retaining those that may be dissatisfied with the slow pace of incorporation of their projects.
Note: Percentages in this figure are calculated in relation to all 450 respondents to the survey. This chart takes into account all factors under the “organisational, skills and cultural constraints” category. Further information on survey methodology may be found in Annex A. Respondents could choose multiple answers. This question was non-mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
Data-related constraints refer to challenges in gaining access to high-quality data, as well as in managing such data within internal processes. A third of survey respondents identified data quality as a constraint to AI adoption, while access to data and lack of data management processes are also frequently identified. For large non-regulatory constraints, firms most frequently indicated data quality issues, particularly concerning their accuracy and consistency, as identified by 22% of all respondents. Firms also identified other data challenges, including obtaining access to data (12% of respondents). Access to data was the most widely cited small constraint (15% of respondents). Other widely cited small constraints referred to dealing with data sets and issues with data format, shortcomings in internal data processes or lack of data management processes (Figure 1.26).
High-quality, unbiased and up-to-date data is essential for effective AI implementation, yet respondents frequently cited issues with data availability and consistency in their qualitative input. As reported by Italian and global firms, it is a particular challenge to find data specific to financial market activities. Complications arise as all treatment of personal data requires identification and assignment of the related responsibilities. For data that is not public, there is a need to obtain the agreement of users for its use, which raises the level of complexity of the process. Firms struggle with a lack of adequate internal data management processes with structured frameworks for data handling, crucial for meaningful AI integration, such as the input of validated and up-to-date information. Companies also expressed the view that sometimes the level of data governance maturity varies across the group or business lines, which presents an obstacle for the successful scaling up of AI at the company level.
Note: Percentages in this figure are calculated based on all 450 respondents to the survey. This chart takes into account all factors under the “data-related constraints” category. Further information on survey methodology may be found in the Annex. Respondents could choose multiple answers. This question was non-mandatory.
Source: 2025 OECD Survey on the use of AI in Italian financial markets.
References
[10] AMF (2026), The Use of AI by Financial Market Participants in France, https://www.amf-france.org/sites/institutionnel/files/private/2026-02/202602-rapport-amf-ia-et-acteurs-des-marches-financiers-en.pdf (accessed on 5 February 2026).
[8] Bank of Japan (2024), Use and Risk Management of Generative AI by Japanese Financial Institutions -Based on the results of questionnaire survey, https://www.boj.or.jp/en/research/brp/fsr/data/fsrb241029.pdf (accessed on 19 December 2025).
[7] BoE/FCA (2024), Artificial intelligence in UK financial services - 2024, https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024 (accessed on 19 December 2025).
[26] Crisanto, J. et al. (2024), Regulating AI in the financial sector: recent developments and main challenges, BIS, https://www.bis.org/fsi/publ/insights63.htm (accessed on 10 December 2025).
[14] DeNederlandscheBank/AFM (2024), The impact of AI on the financial sector and supervision, https://www.dnb.nl/media/gepbbikm/ai.pdf (accessed on 19 December 2025).
[24] EBA (2024), Risk Assessment Report - November 2024 (Special topic – Artificial intelligence), https://www.eba.europa.eu/publications-and-media/publications/risk-assessment-report-november-2024 (accessed on 19 December 2025).
[20] EIOPA (2024), Report on the digitalisation of the European insurance sector, https://www.eiopa.europa.eu/publications/eiopas-report-digitalisation-european-insurance-sector_en (accessed on 19 December 2025).
[23] ESMA (2025), Artificial intelligence in EU investment funds: adoption, strategies and portfolio exposures, https://doi.org/10.2856/6000433.
[22] ESMA (2023), Artificial intelligence in EU securities markets, https://doi.org/10.2856/851487.
[30] EU (2024), Regulation (EU) 2024/1689 (AI Act), https://eur-lex.europa.eu/eli/reg/2024/1689/oj (accessed on 28 November 2024).
[28] EU (2022), Regulation (EU) 2022/2554 of the European Parliament and of the Council of 14 December 2022 on digital operational resilience for the financial sector and amending Regulations (EC) No 1060/2009, (EU) No 648/2012, (EU) No 600/2014, (EU) No 909/2014 and (EU) 2016/1011 [DORA Regulation], https://eur-lex.europa.eu/eli/reg/2022/2554/oj (accessed on 20 May 2025).
[5] Federcasse (2025), L’organizzazione del Credito Cooperativo - Credito Cooperativo, https://www.creditocooperativo.it/page/il-credito-cooperativo/lorganizzazione-del-credito-cooperativo (accessed on 13 October 2025).
[19] Filippucci, F. et al. (2025), “Macroeconomic productivity gains from Artificial Intelligence in G7 economies”, OECD Artificial Intelligence Papers, No. 41, OECD Publishing, Paris, https://doi.org/10.1787/a5319ab5-en.
[13] Finansinspektionen (2024), AI in the Swedish financial sector, https://www.fi.se/contentassets/084ebc13d6364a28a87a37c9a557ec9c/report-ai-swedish-financial-sector.pdf (accessed on 15 September 2025).
[9] FIN-FSA (2025), The use of artificial intelligence by financial sector actors, https://www.finanssivalvonta.fi/globalassets/en/publications/supervision-releases/2025/the-use-of-artificial-intelligence-by-financial-sector-actors_en_.pdf (accessed on 15 September 2025).
[6] FINMA (2025), FINMA survey: artificial intelligence gaining traction at Swiss financial institutions, https://www.finma.ch/en/news/2025/04/20250424-mm-umfrage-ki/#:~:text=Around%2050%25%20of%20institutions%20have,management%20and%20enterprise%20risk%20management. (accessed on 9 September 2025).
[16] IMF (2024), “Chapter 3: Advances in Artificial Intelligence: Implications for capital market activities.”, in Global Financial Stability Report: Steadying the Course: Uncertainty, Artificial Intelligence, and Financial Stability, https://www.imf.org/-/media/Files/Publications/GFSR/2024/October/English/ch3.ashx (accessed on 19 December 2025).
[15] IOSCO (2025), Artificial Intelligence in Capital Markets: Use Cases, Risks, and Challenges, https://www.iosco.org/library/pubdocs/pdf/IOSCOPD788.pdf (accessed on 19 December 2025).
[29] OECD (2025), “AI openness: A primer for policymakers”, OECD Artificial Intelligence Papers, https://doi.org/10.1787/02f73362-en.
[11] OECD (2025), “Artificial intelligence in Asia’s financial sector: A review of country policies”, OECD Artificial Intelligence Papers, No. 50, OECD Publishing, Paris, https://doi.org/10.1787/3385bbd8-en.
[17] OECD (2024), “Regulatory approaches to Artificial Intelligence in finance”, OECD Artificial Intelligence Papers, No. 24, OECD Publishing, Paris, https://doi.org/10.1787/f1498c02-en.
[4] OECD (2023), G20/OECD Principles of Corporate Governance 2023, OECD Publishing, Paris, https://doi.org/10.1787/ed750b30-en.
[18] OECD (2023), “Generative artificial intelligence in finance”, OECD Artificial Intelligence Papers, No. 9, OECD Publishing, Paris, https://doi.org/10.1787/ac7149cc-en.
[2] OECD (2022), G20/OECD High-Level Principles on Financial Consumer Protection 2022, OECD Publishing, Paris, https://doi.org/10.1787/48cc3df0-en.
[27] OECD (2022), “OECD Framework for the Classification of AI systems”, OECD Digital Economy Papers, No. 323, OECD Publishing, Paris, https://doi.org/10.1787/cb6d9eca-en.
[21] OECD (2021), Artificial Intelligence, Machine Learning and Big Data in Finance: Opportunities, Challenges and Implications for Policy Makers, OECD Publishing, Paris, https://doi.org/10.1787/98e761e7-en.
[3] OECD (2020), Recommendation of the Council on Financial Literacy, https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0461 (accessed on 18 December 2025).
[1] OECD (2019), Artificial Intelligence in Society, OECD Publishing, Paris, https://doi.org/10.1787/eedfee77-en.
[25] Prenio, J. (2025), “Starting with the basics: a stocktake of gen AI applications in supervision”, FSI Briefs, No. 26, BIS, https://www.bis.org/fsi/fsibriefs26.htm (accessed on 19 December 2025).
[12] U.S. Department of Treasury (2024), Artificial Intelligence in Financial Services, https://home.treasury.gov/system/files/136/Artificial-Intelligence-in-Financial-Services.pdf (accessed on 19 December 2025).
← 1. The OECD survey was shared with over 900 institutions across the Italian financial sector and accessed by 731 institutions, receiving 450 responses.
← 2. Input to project bilateral meetings conducted in 2025.
← 3. GPAI models (also known as foundation models in some jurisdictions) include Generative AI models, such as LLMs. A GPAI model means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market (EU, 2024[30]).
← 4. Terms “free” and “open-source” are not necessarily interchangeable, as “open-source” does not necessitate free of charge, see AI openness: A primer for policymakers (OECD, 2025[29]).
← 5. Input to project bilateral meetings conducted in 2025.