Over the past few decades, public financial management (PFM) has regularly integrated new technologies, notably by adopting increasingly sophisticated Financial Management Information Systems (FMIS). As technology has advanced and the quality of data has improved, PFM organisations have adopted new technological approaches, such as through data analytics, business intelligence tools and robotic process automation (RPA). These have been the building blocks for the application of AI, which has generally been used to enhance and deepen existing capabilities, with AI systems and approaches being integrated into existing products and processes. In PFM, AI so far has generally been a continuation of the field’s technological evolution rather than a technological revolution.
Governing with Artificial Intelligence
Current state of play
The current and envisioned uses of AI in PFM are mostly related to improving existing processes rather than fully re-envisioning and redesigning them (e.g. by removing any human intervention) or creating entirely new processes. AI is currently used as an assistant that provides high speed and low transaction cost in automating small, often mundane tasks for civil servants, and as an advisor that analyses historical or real-time data to forecast future events or behaviours to support civil servants’ own analyses.
In this context, AI approaches, especially ML, have current applications for PFM in macroeconomic and macrofiscal forecasting1 and supporting spending decisions; budget planning and monitoring; financial management, reporting and oversight; and engagement with external stakeholders.
Improving forecasting
AI can help address challenges with traditional financial and economic forecasting methods by enhancing the accuracy and timeliness of predictions, with AI systems able to outperform traditional economic forecasting models (Jung, Patnam and Ter-Martirosyan, 2018[5]).
AI’s predictive capacities are also seen as an opportunity to develop nowcasting – identifying changes in near real time and extrapolating potential near-term futures. Nowcasting takes into account the very recent past and present to forecast the very near future state of economic indicators that are typically only determined after a delay and are subject to revision, such as gross domestic product (GDP) or inflation. Central banks of different countries are exploring the adoptions of AI systems to provide more accurate and anticipated predictions than traditional time series models, even when using unstructured data, such as New Zealand (Richardson, van Florenstein Mulder and Vehbi, 2019[6]), France (Blanchet and Coueffe, 2020[7]), and Peru (Tenorio and Perez, 2023[8]).
The Swedish National Financial Management Authority (ESV) has developed an innovative GDP forecasting application leveraging explainable ML to enhance both accuracy and transparency in economic predictions. The model outperforms Sweden’s pre-pandemic official forecasts and addresses a key limitation of traditional AI-based forecasting — its opacity — by visualising the impact of variables over time. This tool enables policymakers, researchers, and the public to generate reliable projections while maintaining interpretability, supporting data-driven decision making. The initiative showcases how AI can be integrated into macroeconomic forecasting while ensuring accountability and trust.
Facilitating spending decisions
To facilitate decision-making in PFM, AI is backed by technological advancements that have become mainstream. These include big data for analysing vast quantities of information from multiple sources, data analytics tools to drill down into specific financial categories and beneficiaries and evaluate the effectiveness of expenditures based on trends and patterns, and data visualisations to allow effective communication of complex information.
AI can be used to build on these foundations by identifying trends and patterns and grouping data points based on similarity or shared characteristics. For spending decisions specifically, AI can analyse historical budget execution data to identify underspending or overspending patterns, predict future spending needs based on key parameters (e.g. demographic changes), and evaluate programme effectiveness by linking expenditure data with outcome metrics. This offers opportunities to accelerate and improve analyses in ways that can automate or augment the work of humans. ML techniques that leverage unstructured data open opportunities for combining datasets previously unused for such exercises (Box 5.8).
In 2022, South Korea developed and implemented dBrain+, an advanced financial management information system that leverages AI to analyse real-time economic, fiscal and financial data, optimising risk assessment and decision-making in public finance. Its key modules, the Korea Fiscal Information System (KFIS) and Korea Risk Assessment and Horizon Scanning (KORAHS), use AI-driven analytics to detect financial risks and support data-informed policy decisions. By centralising all national financial operations — from budgeting and fund management to debt oversight and performance evaluation — dBrain+ enhances efficiency, transparency and predictive capabilities across central and local governments.
A key strength of dBrain+ is its integration with 63 systems from 46 institutions, including the National Tax Service, Public Procurement Service and the Bank of Korea, enabling seamless coordination on contracts, tax collection and fund transfers. AI-powered analysis of this data in real time improves budget execution, accelerates financial reporting and supports the identification of risks, enabling better decision-making on fiscal policies and public spending. By providing tailored tools for different users — including civil servants, researchers and external stakeholders — dBrain+ strengthens accountability and modernises South Korea’s approach to AI-driven fiscal governance.
Supporting budget planning and monitoring
AI has the capacity to support budget planning and monitoring processes by providing outputs that support the formulation of accurate expenditure baselines and costing of new policies. For instance, Australia's Department of Veteran’s Affairs developed predictive systems and tools to help simulate future financial impacts of the policy decisions. These include the annual fiscal expenditure for each beneficiary as well as their average years on benefits, which are used for costings, budget estimates and policy evaluation (Australian Government, 2020[10]).
Another promising domain of application for AI is the identification, monitoring and mitigation of fiscal risks through the analysis of large data sets. Government fiscal risks can arise due to a variety of causes, including unsustainable spending or investment levels, which need to be identified early to take preventive action. AI can support the identification of such risks, as seen in Box 5.9. In another example, Indonesia uses a system called AI Financial Advisor (AIFA) to process unstructured financial and performance data to provide analytics on subnational governments’ fiscal performance in real time (Wisesa, 2023[11]).
For several years, France’s tax agency (DGFiP) has implemented an AI-enabled “warning system” that aims to identify municipalities with financial difficulties, provide them with financial advice and proactively support the implementation of corrective measures.
This warning system was based initially on an algorithm using historical tax and financial data to score municipalities. More recently, DGFiP developed a predictive AI system designed to identify municipalities’ financial difficulties earlier. The system was trained on data spanning four years to predict outcomes for the fifth year. The predictive system also relies on unsupervised clustering techniques to categorise municipalities with similar financial characteristics without predefined outcome examples.
In 2022, an experiment with the system covered 2 500 municipalities, with around 40% of these identified as facing financial difficulties. Of these, around 17% had not been detected by the previous algorithm. Additionally, around 35% of municipalities were identified with temporary, non-structural difficulties, highlighting the system’s capacity to differentiate between permanent and transient financial problems.
Automating management, reporting and oversight activities
PFM and reporting activities involve important but sometimes repetitive tasks that are particularly well-suited for automation. AI techniques such as natural language processing (NLP) can be used to analyse digital images to extract information from documents (e.g. vendor information), identify and classify documents (e.g. invoices), perform document comparison (e.g. compare invoice and vendor information), or identify trends and patterns (e.g. internal controls on payment requests).
In France, for example, the French tax agency (DGFiP) (2024[13]) has developed an AI-based tool as part of the regular internal control processes that “automatises the selection of payment requests to be controlled [and] optimises the workload and the quality of controls performed”. The Finnish Government Shared Services Centre for Finance and HR (Palkeet) established a Centre of Excellence for Robotic Process Automation (RPA). It focuses on developing and deploying RPA solutions across various financial and HR activities — such as the management of supplier information, balancing of accounting data and processing of financial transactions — and integrating AI into automation processes where complex decision-making or data processing are necessary (Palkeet, 2024[14]).
Brazil’s National Treasury (STN) is using AI to enhance fiscal transparency by classifying subnational government expenditures according to the international COFOG standard. Previously a manual and resource-intensive task, the adoption of AI — using ML models with Convolutional and Recurrent Neural Networks — has reduced classification time from 1 000 hours of human work time to just 8 hours, while achieving over 97% accuracy. This breakthrough led to the publication of the Expenditure by Function of the General Government report in 2024, a milestone in Brazil’s fiscal statistics.
Building on this success, STN is now expanding AI applications to new areas, including the classification of climate-related expenditures. In collaboration with the Inter-American Development Bank (IDB), Brazil is strengthening its capacity to assess the fiscal implications of climate change. By pioneering AI-driven public finance management, Brazil sets an example for other nations seeking to modernise fiscal statistics and enhance transparency in an increasingly complex economic landscape.
In addition to strengthening reporting, the development of targeted verifications to identify errors (e.g. improper payments) and fraud (e.g. identity theft) have become a major objective for many governments. This is especially true in the wake of the COVID-19 pandemic, which exposed vulnerabilities in several countries’ payment systems. AI’s capacity for identifying trends and patterns can help with this. For instance, during the pandemic, the Danish Business Authority developed AI-based controls for aid applications from companies for various support schemes (van Noordt and Tangi, 2023[15]).
Facilitating engagement with stakeholders and users
As discussed throughout this chapter and synthesised in Chapter 2, chatbots powered by NLP and language models are increasingly employed in government to directly provide services. As related to PFM, the United Arab Emirates (UAE) has developed U-Ask, a unified AI-powered chatbot for government services that can also be used to answer simple fiscal reporting questions2. In Mexico, the government has introduced an AI virtual assistance tool as part of its Intelligent Support Platform, designed to guide users through government programmes and supports (2023[16]). The tool provides information on benefits, eligibility and application processes for individuals, businesses and local governments, utilising a simple keyword search or personalised questionnaires to tailor the information to the user's profile.
Managing risks and challenges
Associated risks
Lack of transparency and explainability
Inadequate or skewed data in AI systems
Due to their black box nature, AI-based systems that have the best forecasting outputs represent both a step forward in accuracy and a step back in fiscal transparency (Jung, Patnam and Ter-Martirosyan, 2018[5]). This lack of transparency makes it difficult for governments to verify the decision-making processes of these models, which is crucial for accountability and regulatory compliance. Consequently, this limitation has prompted governments to prioritise the use of simpler AI systems to improve human modelling and sensitivity analysis.
While governments and PFM organisations have become increasingly adept at identifying AI-related risks and challenges, many are still developing comprehensive frameworks and practical approaches to manage these risks. Their efforts focus on establishing governance structures, building technical capacity and creating clear protocols for AI deployment in public financial systems.
Governments are also working to develop methods to “unbox” AI systems and make their reasoning more transparent, explainable and interpretable — all important conditions for using these systems in PFM. For instance, as discussed above (Box 5.7), the Swedish National Financial Management Authority (ESV) has developed an application for analysing the impact that each data variable has on the prediction of the “black-box models”, as part a wider work programme to integrate AI in the financial management of Swedish government (Boström et al., 2020[17]).
AI use cases show that ethical risks like incomplete or insufficient data and skewed algorithms can be significant in the field of PFM. AI can amplify patterns of inequality embedded in financial data, leading to financial exclusion of perceived high-risk individuals (Crisanto et al., 2024[18]). This phenomenon is observed in the banking industry, where biased credit distribution may perpetuate discriminatory lending practices (Bailey, 2023[19]; Klein, 2020[20]). In the realm of public financial management, biased algorithms can similarly affect the distribution of public funds, social benefits and access to government programs, perpetuating existing inequalities and hindering fair treatment. While more of an automated decision-making system than true AI, Australia’s Robodebt scheme illustrates problems that can arise as a result of algorithmic errors if not caught and addressed by humans (Box 5.11).
Australia’s Robodebt scheme, introduced in 2016, was an automated debt recovery programme designed to identify and recoup welfare overpayments. It replaced a manual process with a data-matching algorithm that compared fortnightly income data reported to Centrelink, the agency responsible for social security payments, with averaged annual income figures from the Australian Taxation Office (ATO). Discrepancies were flagged as overpayments, and debt notices were automatically issued without human verification. This "income averaging" method ignored fluctuations in actual earnings, often generating false debts. The system also reversed the burden of proof, requiring recipients to provide historical pay records to contest debts — a demanding task for many. Over its operation, the scheme issued 470 000 incorrect debt notices totalling EUR 775 million, causing widespread distress and financial hardship.
The scheme’s calculations were declared unlawful in 2019. A Royal Commission was set up in 2022 to enquire into the establishment, design and implementation of the Robodebt scheme; the use of third-party debt collectors under the Robodebt scheme; concerns raised following the implementation of the Robodebt scheme; and the intended or actual outcomes of the Robodebt scheme. The Royal Commission issued a report in 2023, which discussed impacts the scheme had on recipients, including those related to income withholdings and garnishees, emotional and psychological effects, and the loss of faith in the government. Robodebt exemplifies the risks of automating complex social systems without adequate human oversight or rigorous testing. The fallout included significant legal settlements and calls for stricter regulations on the use of algorithms in public policy. While the Royal Commission’s findings were not necessarily representative of the Australian Government’s views, the government agreed, or agreed in principle, to 56 of the commission’s 57 recommendations.
The Robodebt scheme — which used automated data-matching, income averaging and overpayment calculation — can be described as an automated decision-making system. While the scheme did not leverage AI, it helps to illustrate issues in governance, human oversight and algorithmic design.
Implementation challenges the process in a way to minimise
Matching problems to AI solutions
Inflexible or outdated legal and regulatory environments
Outdated legacy information technology systems
Lack of high-quality data and the ability to share it
Skills gaps
Lack of actionable frameworks and guidance on AI usage
One challenge is to match PFM needs and AI technologies. Most finance ministries that have already implemented AI projects in PFM emphasise the importance of mapping processes and activities to pinpoint areas of inefficiency and potential efficiency gains as a prerequisite to deploying AI.3 Once this initial phase is complete, the next step is to assess the suitability of AI technology, or other technologies, for integration to help address them.
Despite AI’s capacity to summarise or draft text using language models, finance ministries have been cautious in rolling out new technologies in fiscal reporting (e.g. the automatic production of fiscal reports). This may be due to concerns over accuracy and whether AI complies with current regulations. There may also be concerns over where responsibility lies when AI is used, and what its use means for people in positions of responsibility, such as external auditors, and people who use the reports, such as lawmakers.
Information technology (IT) systems are crucial for finance ministries to be able to take advantage of AI opportunities. Yet many OECD countries say they are locked into legacy technologies that are significantly fragmented, often outdated and lack the necessary infrastructure and compatibility to integrate advanced AI functionalities. For example, centrally managed FMIS systems are more than 10 years old in most OECD countries (OECD, 2024[21]). These technologies are not specific to OECD countries and are holding back use of AI in PFM across the world (Rivero del Paso et al., 2023[22])
As with any IT system, the quality of output obtained from an AI system depends on the quality of inputs. Finance ministries also indicate that fragmented data, coupled with restrictions on data sharing, frequently impedes the initiation of AI projects (see Chapter 4, section on “Creating a strong data foundation”).9 These challenges underscore the need for improved data management practices and policies that facilitate more effective data accessibility and sharing. Accordingly, various OECD countries plan major upgrades to their FMIS, also recognising the need for stronger data foundations (Figure 5.2).
Note: Refers only to countries currently undertaking major development or replacements of their central FMIS (18 countries). Ratings present the average level of importance assigned to each objective on a scale of 0 to 4 by all respondents. Data for Chile, Colombia, Israel, Mexico, Slovenia and the United States are not available.
Source: (OECD, 2022[23]), Question 24.
Because PFM is a highly technical policy area, implementing AI systems requires substantial training that involves human intervention and supervision (and feedback) from PFM experts. A significant challenge is that while PFM organisations can effectively identify AI-related risks, many lack the specialised staff skills and institutional capacities needed to develop and implement necessary risk management frameworks. Further, it is tempting to overestimate AI’s capabilities. AI systems produce outputs that are based on probabilities, which means they are by nature uncertain. They can produce data outputs that are wrong or texts that sound highly authoritative but are incorrect (“hallucinations”). Therefore, no matter how much training systems may receive, PFM specialists need to be able to exert critical judgment when using the outputs they generate. Such oversight requires specialists that combine technical PFM skills and basic understanding of how AI works.
Ensuring transparency and explainability in AI-driven decisions requires robust frameworks and standards that govern and oversee AI processes, including the traceability and accessibility of training datasets (e.g. data provenance, data records) and, when possible, the source code for AI algorithms. However, such frameworks and standards have until recently lacked in many OECD countries. Traditional oversight bodies, such as supreme audit institutions (SAIs) and independent fiscal institutions, are beginning to adapt their methodologies and develop new skills to effectively oversee AI-driven fiscal processes (Box 4.8). Additionally, identifying and implementing safeguards against potential misuse or over-reliance of AI in fiscal governance is crucial as AI systems become more prominent in decision-making processes. PFM specialists need to address these issues to shape a new fiscal governance framework that exploits AI's potential while maintaining transparency, accountability and integrity.
Untapped potential and way forward
Finance ministries are currently adopting a cautious approach to AI, prioritising task automation and predictive applications over more prescriptive AI. While predictive AI focuses on forecasting outcomes, prescriptive AI goes further by suggesting courses of action to achieve desired goals or mitigate risks. However, a systematic application of prescriptive AI could profoundly impact the roles and responsibilities within PFM systems, potentially altering the activities of fiscal stakeholders and oversight bodies.
Any shift from human judgment to system-based outcomes in fiscal decisions necessitates a risk-based re-evaluation of accountability, and the assignment of roles and responsibilities in an increasingly automated environment. This also requires considering how AI could alter the letter and spirit of PFM institutions and processes, and how it could reshape the functions of fiscal stakeholders, including external oversight bodies such as SAIs and independent fiscal institutions. In this context, questions regarding the future of transparency and accountability need to be addressed.
What mechanisms will be needed to ensure transparency in automated PFM decisions? As forecasting and budget planning and monitoring could increasingly be conducted by AI, there should be robust frameworks to ensure transparency and explainability regarding governments’ use of AI systems and their underlying data. This involves creating and implementing standards that govern AI use and oversight.
How will the roles of traditional oversight bodies evolve as AI is adopted? SAIs and independent fiscal institutions will need to adapt their methodologies to effectively oversee and audit AI-powered fiscal processes. This might include developing data science and AI training for staff members or updating audit processes to incorporate AI-specific considerations, such as data integrity and completeness.
What new safeguards will be necessary to protect against misuse of AI in fiscal governance? As AI systems play a more prominent role in fiscal decision-making, identifying possible types of intended or unintended misuses and safeguarding against them becomes paramount.
By addressing these questions, PFM specialists can help shape a new fiscal governance framework that accommodates the full innovative potential of AI, while safeguarding the key principles of PFM.
Finance ministries’ AI efforts need a coordinated approach and should integrate lessons from other governmental projects. This could greatly minimise risks of failure and implementation challenges, while enhancing the quality of outcomes. For instance, community of practice on AI projects can help to prevent common and avoidable mistakes and mitigate risks from inherently complex AI projects. These risks stem, for example, from the involvement of multiple stakeholders with diverse interests or lack of understanding of technologies and systems.
While many financial management agencies identify a potential for large-scale productivity gains from AI, public studies on feasibility and costs are rare and assessments on results and impacts remain anecdotal. Evidence on costs and impacts are either not collected or not publicly available due to the provisional nature of projects still in pilot or early development phases.
Ideally, the results and impact of AI use in PFM should be assessed by using evaluation frameworks to track costs of projects and key performance indicators from completed projects, including cost savings, effectiveness and efficiency gains, error reduction and compliance enhancement. This would involve, among other things, monitoring full costs of projects, comparing metrics from before and after AI implementation, conducting stakeholder surveys for satisfaction, and analysing data to see how AI outcomes match with fiscal policy goals.
Due to the lack of impact data, external oversight bodies recently called for greater transparency in AI projects across government; they said that greater scrutiny and evidence collection is required alongside substantial investment (see Chapter 4, sections on “Investing purposefully” and “Empowering oversight and advisory bodies to guide responsible AI”).
Establishing robust project selection and evaluation frameworks within finance ministries is critical. These should be aligned with government-wide frameworks. They should track not only cost and performance indicators but also the qualitative impacts of AI, such as error reduction and compliance enhancement. This is needed, as multiple projects will likely compete in the future for limited investment resources.
To scale up their ambitions progressively and safely, finance ministries could adopt a sequenced approach to the introduction of AI technologies. This is well illustrated by the case of The Finnish Government Shared Services Centre for Finance and HR (Palkeet), as discussed above, which started with “small” uses of RPA and is now moving towards hyper-automation with ML. According to Palkeet, starting small with their advanced digitalisation journey also helped increase the acceptability for civil servants of more sophisticated AI technologies (Palkeet, 2024[14]).
← 1. Macrofiscal forecasting is the process of predicting key economic and government finance indicators, including GDP growth, inflation, tax revenues, public spending, and debt levels. These forecasts help governments plan budgets, set fiscal policy, and manage public finances over multi-year periods.
← 3. Based on OECD discussions and working party meetings with PFM officials.