## tnmea2026006-s001 ## Source details **Canonical URL:** [tnmea2026006-s001](https://www.imf.org/-/media/files/publications/tnm/2026/english/tnmea2026006-s001.pdf) ## Other formats - [Markdown version](/-/media/files/publications/tnm/2026/english/tnmea2026006-s001.pdf.md) - [Structured JSON version](/-/media/files/publications/tnm/2026/english/tnmea2026006-s001.pdf.json) --- ### Technologies and techniques associated with Narrow AI - Definitions and spectrum: - Artificial Intelligence (AI): development of computer systems to perform tasks that typically require human intelligence. - Narrow AI: AI trained for specific tasks; focus for PFM experts. - Spectrum includes: Reactive Machines; Limited Memory AI; Theory of Mind AI; Self-Aware AI; Narrow/Weak AI; General/Strong AI; Superintelligent AI. - Major Narrow AI technologies and characteristics: - Machine learning (ML): supervised, unsupervised, reinforcement learning; subfields include deep learning and federated learning. - Natural language processing (NLP): text analysis, sentiment analysis, translation, named entity recognition. - Generative AI (GenAI): content creation via transformers, GANs, VAEs. - Agentic AI: planning, reasoning, tool use, multi-step task execution. - Computer vision: object detection, image segmentation, facial recognition, OCR. - Expert systems: rule-based decision-making using knowledge bases. ### Considerations and pre-conditions for AI implementation - Data and model safeguards: - Ensure availability of high-quality, relevant data; address outdated formats, inconsistency, and sharing reluctance. - Data preparation: cleaning, preprocessing, augmentation, annotation; address language and domain specificity for NLP. - Model selection and maintenance: select appropriate models (e.g., BERT, GPT, Prophet, XGBoost), validate thoroughly, retrain to reflect new data and linguistic trends. - Technical infrastructure and tooling: - Hardware/software needs (GPUs; PyTorch, TensorFlow, Scikit-learn); NLP frameworks (Hugging Face Transformers, spaCy, NLTK); APIs and middleware for integration. - Security, ethics, compliance: - Anonymization, regulatory compliance, mitigation of algorithmic bias, transparency and explainability. - Costs, skills, and complexity: - Rolling operational costs for AI SaaS; upskilling staff; interpretability of probabilistic outputs; significant implementation/maintenance costs tied to dataset size and algorithm complexity. - Problem scoping: - Define clear use cases before selecting technologies; avoid hype-driven adoption. ### Applications in Public Financial Management (PFM) — Narrow AI use cases and country examples - Core PFM applications: - Compliance monitoring, fraud detection, macrofiscal nowcasting/forecasts, red flags/anomaly detection, automated reporting, chatbots, information extraction, financial document processing. - Country/project highlights: - Civil society monitoring: - Ukraine: Dozorro evaluates probability of corruption risks in public procurement, adapts methods when human experts detect new circumvention strategies. - Brazil: Operação Serenata de Amor uses data science and unsupervised learning to detect suspicious public spending and engages the public via a web platform. - Nowcasting: - Georgia and Peru: Ministries use high-frequency indicators and ML (including Google Trends) to “nowcast” economic indicators (Mgebrishvili, 2022; Tenorio and Perez, 2023). - Cash flow forecasting: - Guatemala: IMF Fiscal Affairs Department collaboration; ML methods such as Prophet and XGBoost show strong potential for nonlinear, seasonal, high-frequency cash flows. - Budget classification: - Rwanda: ML on a dataset of more than 40 million transactions; moving toward deep learning to tokenize and analyze transaction strings. - Government grant search: - Mexico: NLP (CoreNLP) implemented on the Budget Transparency website since 2018 to map user keywords to grants. - Chatbots: - Guatemala: Ch@tío (2020) supports suppliers; evolved into private messaging in 2022; estimated response time of one second. - Singapore: Ask Jamie (conceptualized 2014) across 70 agency websites; GovTech updating chatbots to LLM engines. - Sentiment analysis: - United States: US Treasury’s Bureau of Fiscal Service uses NLP on e-mails and call center audio to set priorities and optimize operations. - GenAI in PFM: - Pilots include emergency/disaster management (weather forecasting enhancements), productivity tools (GitHub Copilot). - As of 2025, most GenAI uses in PFM were experimental; many countries expanding regulations, using sandboxes and human oversight. ### Generative AI — implementation considerations and risks - Implementation checklist: - Use case identification to avoid hype-driven adoption. - Curate high-quality datasets; run pilots in sandbox environments. - Fine-tune and customize models to domain needs; combine human oversight and validation. - Integrate via infrastructure and APIs with existing PFM systems. - Key risks and constraints: - Hallucinations: plausible but factually incorrect outputs require human evaluation and controlled learning. - Rolling costs: pay-as-you-go models and resource-intensive computations. - Privacy: avoid exposing sensitive data during training or inference. - Regulatory developments: monitor and adapt to evolving rules. ### Agentic AI — capabilities, uses, and governance - Definition and components: - Systems that set/interpret goals, plan, act through tools, and adapt via monitoring/learning loops; components include reasoning/planning, tool/data access, and monitoring/learning. - Examples and country notes: - Ukraine: Diia.AI integrated into national platform to match users to public services and explain documents. - Albania: Diella (previous virtual assistant) appointed Minister of State for AI in 2025 as an aspirational step toward agentic responsibilities in procurement evaluation. - PFM applications: - Service delivery (pre-filled forms, routing), fraud and waste reduction (continuous monitoring, alerting), audit/compliance automation, procurement review and flagging, treasury decision support (short-term issuance/payment sequencing, cost and market impact evaluation). - Governance and implementation safeguards: - Expert-in-the-loop governance; explicit guardrails, permissions, approval thresholds; full logging for auditability; clear accountability and potential need to update PFM laws; cybersecurity and integration with legacy systems; upskilling for “agent literacy.” ### Computer Vision — applications and implementation considerations - Definition and traditional tools: - Interprets visual data; combines with LiDAR, BIM, GPS; OCR for document extraction. - Country examples: - Korea: product image analysis and satellite-based forest fire management. - Philippines: Project DIME (2018) used LiDAR and geotagging to monitor 75 major projects; planned relaunch. - Peru: national guide on Plan BIM (2021) for public investment in infrastructure. - PFM uses: - Document verification, public investment monitoring, image analytics for expenditure efficiency, data extraction from invoices/contracts. - Implementation considerations: - Collect and annotate visual data; choose appropriate algorithms (e.g., CNNs) and platforms (OpenCV); ensure data security/privacy; combine with complementary technologies; test for inadequate models and visual data diversity challenges. ### Expert Systems — components, strengths, limitations, and PFM uses - Core components: - Knowledge base (facts and rules), inference engine (forward/backward chaining), user interfaces and APIs. - Strengths: - Excel at rule-based tasks requiring explicit domain knowledge. - Limitations: - Difficulty capturing tacit expert knowledge; rigidity and lack of common-sense reasoning; maintenance and trust issues. - PFM applications: - Budget baselines, knowledge sharing, risk assessment of contingent liabilities, project prioritization. - Implementation notes: - Include explanation facilities, knowledge acquisition systems, database integration, control structures, error handling, scalability and continuous updates. ### APIs, API First, and interoperability — benefits and implementation guidance - API role and advantages in PFM: - Enable seamless data exchange, transparency, operational efficiency, innovation, interoperability, and governance. - API First approach: - Prioritize API design and contracts (e.g., Swagger/OpenAPI); treat APIs as products. - Pre-conditions and implementation considerations: - Ensure data consistency, real-time synchronization, documentation, security (OAuth, TLS), scalability, standards (JSON, XML, XBRL), and integration with FMIS. - Quick guide (nine essential tasks): - Brainstorming, stakeholder engagement, API contract, style guide, governance, process automation, mock generation, self-service documentation, developer portal. - Country examples: - South Africa: Municipal Money API publishes 17 years of municipal data across over 290 municipalities; data free for commercial and non-commercial use subject to limitations. - Ghana: Ghana e-payment portal (GEEP) to expose open APIs for interoperability; GhIPSS leveraged APIs to achieve interoperability goals (renewed in 2022). - Brazil (São Paulo): APIs connect costing system to administrative records and automate data submission. ### Big Data Analytics, Cloud, Databases, Data Lakes, and Data Standards - Big Data Analytics: - Characteristics: volume, variety, velocity, veracity, variability, value. - PFM applications: budget planning, expenditure efficiency, service delivery monitoring, citizen engagement. - Country examples: Nigeria, Togo (social protection targeting during COVID‑19); Australia (subsidy targeting); Rwanda (national statistical institutes); Ukraine (social media sentiment generative model). - Caution: tools imperfect, risk of manipulation; should serve advisory roles; require strong leadership. - Cloud computing: - Advantages: scalability, interoperability, cost savings (pay-per-use), easier upgrades and recovery. - Service models: SaaS, IaaS, PaaS; deployment models include Public, Private, Hybrid, Sovereign cloud. - Country examples: Brazil (Serpro, Dataprev), Chile (SIGFE, SIAP), India (e-Lekha on IaaS). - Preconditions: cloud readiness, cloud policy, data sovereignty, skills, vendor management; beware vendor lock-in and hidden costs. - Databases and data architectures: - Types: Relational (MySQL, PostgreSQL), NoSQL (MongoDB), In-memory (SAP HANA), Graph (Neo4j), Time-series (InfluxDB), Blockchain DB (BigchainDB). - Data lake vs lakehouse vs warehouse: lakehouse retains flexibility and high performance; comparison preserves terms Cost$$$$$ and People assignments as in source. - Implementation challenges: data model complexity, performance, vendor lock-in, maintenance, governance. - Data standards: - ISO 20022, ISO 17369 (SDMX), ISO 3166, ISO 8601, ISO/IEC 27000; domain-specific standards include OCDS, CoST IDS, IATI, Open Fiscal Data Package, Beneficial Ownership Data Standard, EITI. - Implementation tasks: data mapping, metadata, ETL, schema updates; challenges from legacy systems, downtime, data quality, validation. ### DevSecOps, Digital Identity, and Distributed Ledger Technologies (DLT) - DevSecOps: - Definition: integrates security into the development lifecycle. - Advantages: enhanced security, collaboration, faster delivery, reputational and financial benefits. - Implementation: phased approach, CI/CD integration, Zero Trust, toolchain (Jenkins, GitLab CI, SonarQube, OWASP ZAP, Ansible, Terraform, Prometheus, ELK Stack), cross-functional teams, governance. - Challenges: regulatory compliance, scaling, skills gaps, cultural resistance. - Digital Identity: - Role: key component of Digital Public Infrastructure (DPI); enables authentication and secure transactions. - India — Aadhaar: - 12-digit number; issued to more than 1.3 billion residents, representing 95 percent of the population. - Direct Benefit Transfer delivered through bank accounts (46 percent) or Aadhaar-linked accounts (54 percent). - Ethiopia — Fayda: - Will provide at least 90 million Ethiopians a unique identification number. - Since 2023, more than 8 million people have registered; over 2 million authenticated; over 30 agencies integrated. - Fayda is now the main credential for civil servants. - Considerations: security/privacy, scalability, adoption, integration, legal/regulatory alignment, standards (OpenID Connect, ISO/IEC 24760). - Distributed Ledger Technologies (DLT): - Use cases: budget monitoring, payments and reconciliations, intra-governmental operations, procurement transparency. - Country examples: Philippines and Poland (tokenized bonds, DLT-based bond registries); Kazakhstan (Digital Tenge CBDC pilot). - Decision framework highlights yes/no questions (e.g., can a traditional DB meet needs? do contributors trust each other?) with outcome guidance: mostly "No" → DLT not needed; many "Yes" → potential DLT case; mixed → further research. - Challenges enumerated: - 1 Nascent technology. - 2 Governance risks. - 3 Data privacy concerns. - 4 Compliance challenges. - 5 Transaction costs. ### Fiscal Transparency Portals, GIS, Interoperability, Microservices, Open Data - Fiscal Transparency Portals: - Purpose: public access to budgets, spending, revenues, debt; progressive disclosure by audience. - Examples: - Mexico: portal integrating over 10 platforms; launched in 2011. - Cabo Verde: SOE Manager automates SOE data collection and KPIs using Microsoft Power BI. - Implementation challenges: balance complexity and accessibility; machine-readable formats; automated timely updates; legal frameworks; civil society engagement. - Geographic Information Systems (GIS): - Value: spatial analysis for site selection, monitoring, risk assessment, asset management, FMIS integration. - Country examples: Italy (farmer subsidies GIS verification), India (welfare program monitoring), Latin America (Data on the Streets Rally since 2016). - Considerations: data collection costs, inconsistent standards, skills gaps, technical infrastructure; use open-source tools (OSM, QGIS) as mitigants. - Interoperability: - Levels: Technical, Syntactic, Semantic, Organizational/Legal/Political. - Tools: X-Road, Federated Identity Management (OpenID, OAuth). - Measures: unique identifiers, common standards (ISO 20022), governance for data sharing. - Microservices: - Advantages: modularity, flexibility, resilience, independence. - FMIS transition: break monoliths into services; Honduras and Paraguay use Docker and Kubernetes. - Preconditions: innovative culture, scalable infrastructure, team-level domain expertise. - Challenges: system complexity, network latency, governance, skill requirements. - Open Data: - Principles: non-proprietary, free to use, machine-readable, primary, permanent. - Quick guide steps: understand demand, locate data, process, standardize, ensure sustainability, protect users, promote use. - Licensing examples: Creative Commons Attribution 4.0 International; PDDL; Open Database License. ### Orchestrators, Containers, Robotic Process Automation (RPA) — adoption metrics and impacts - Orchestrators and containers: - Definitions: containers package applications; orchestrators manage container lifecycles. - Usage statistics presented (categories preserved as listed): 39%; 37%; 10%; 5%; 41%; 47%. - Orchestrator adoption among respondents: - kubernetes: 65% Using in production; 18% Piloting/evaluating. - Other container orchestration tools: 47% Using in production; 41% Piloting/evaluating. - Benefits: efficiency, speed, portability, stability, scalability. - Tools: Kubernetes, Docker Swarm, Apache Mesos, Rancher, Nomad. - Implementation considerations: integration, security, performance, upskilling, monitoring, compliance. - Robotic Process Automation (RPA): - Definition: bots executing programmed rules for repetitive tasks with minimal changes to existing systems. - PFM applications: reconciliation, invoice processing, budget transparency extraction, quality checks. - Country examples and metrics: - Finland (Palkeet, as of 2020): 26 bots automate 70 processes. - United Arab Emirates (as of 2022): more than 8 robots automating 63 processes; outcomes reported: 95% reduction in errors; 65% reduction in average handling time; saving of 39,000 hours in human labor. - United States: multiple agencies use RPA (GSA, Defense Logistics Agency, NASA Shared Services Center). - Implementation considerations: supporting infrastructure, development pipeline, proof-of-concept focus, align with national priorities, address maintenance and integration with FMIS. ### Smart contracts, Tokenization, Zero-Knowledge Proofs (ZKPs) — innovations and cautions - Smart contracts: - Definition: computerized transaction protocols that execute contract terms (Szabo, 1994); often paired with DLT. - PFM applications: automation of business processes, asset management, oversight, procurement execution. - Country examples: US Treasury explorations (2019); GSA acquisition automation; Netherlands (pension and voucher pilots); Sweden and Georgia (land registries). - Implementation challenges: code failures (one audit found critical bugs in 25% of smart contracts), legal enforceability, cybersecurity, stakeholder resistance, oracle integration, wallet provisioning, governing law definition, hybrid text+code approaches recommended. - Tokenization: - Definition: replacing sensitive data with tokens to secure information and represent assets digitally. - PFM applications: budget tracking, asset management, procurement, government payments. - Country examples: - Slovenia: sovereign digital bond of 30 million euros (July 2024) settled on-chain via tokenized cash system. - Philippines: tokenized Treasury bonds sold November 2023 to institutional buyers; plans to expand access to retail via mobile payments. - Luxembourg: regulatory framework for DLT-based dematerialized securities; tokenized fractional forest ownership example. - Considerations: regulatory uncertainty, scalability, data security, technological complexity, implementation costs, interoperability standards. - Zero-Knowledge Proofs (ZKPs): - Definition and types: zk-SNARKs, zk-STARKs, Bulletproofs. - PFM advantages: privacy-preserving verification, regulatory compliance without data exposure, fraud prevention. - Use cases: identity validation, welfare eligibility checks, public health expenditure verification, pension disbursement, confidential procurement bidding, emergency relief distribution. - Implementation considerations: high computational requirements, integration with legacy systems, regulatory ambiguity, resource intensity, cybersecurity, need for pilot phases and robust cryptographic infrastructure. ### Cross-cutting policy and operational recommendations (synthesized from content) - Data governance and standards: - Adopt common data standards (e.g., ISO 20022, SDMX), perform thorough data mapping, implement metadata standards and ETL processes. - Human oversight and accountability: - Embed expert-in-the-loop governance for AI/agentic systems; log actions and reasoning for auditability; maintain human accountability under PFM laws. - Security, privacy, and compliance: - Implement DevSecOps and Zero Trust; employ encryption, role-based access, auditing, and regular security assessments. - Incremental piloting and sandboxes: - Use sandboxes for GenAI, DLT, tokenization, ZKPs, and agentic AI to validate benefits and manage risks before scaling. - API-first and interoperability: - Prioritize API contracts, governance, and developer portals; adopt unique identifiers and semantic standards to enable cross-system data sharing. - Skills, change management, and culture: - Invest in upskilling, build cross-functional teams, foster data cultures, and manage stakeholder engagement to mitigate resistance. - Cost and vendor management: - Assess rolling operational costs (AI SaaS, cloud), avoid vendor lock-in, include exit terms in contracts, and plan for maintenance budgets. - Explainability and bias mitigation: - Require explainability features for AI outputs and implement procedures to detect and mitigate algorithmic bias. *Source: tnmea2026006-s001 - Sections 1–8* ### Section 1 ### Technologies and techniques associated with Narrow AI ### Overview: What AI refers to in this content unit - Artificial Intelligence (AI): development of computer systems to perform tasks that typically require human intelligence; aims to automate tasks or solve problems with greater accuracy than humans. - Narrow AI: AI trained for specific tasks; currently the type implemented in the field. - Spectrum of AI types discussed: - Reactive Machines: follows predefined rules - Limited Memory AI: learns from past experiences - Theory of Mind AI: understands human emotions - Self-Aware AI: understands its own existence - Narrow or Weak AI: trained for specific tasks - General or Strong AI: human-like cognitive abilities - Superintelligent AI: surpasses human intelligence - Note: AI is rapidly evolving; focus is on practical Narrow AI technologies for Public Financial Management (PFM) experts. ### Major Narrow AI technologies and key features - Machine learning (ML): algorithms and models that learn from data to make predictions or decisions. - Common ML types: supervised, unsupervised, reinforcement learning. - Subfields: deep learning, federated learning. - Deep learning: neural networks with many layers; successful in image and speech recognition. - Federated learning: trains models across multiple decentralized devices or servers holding local data samples. - Natural language processing (NLP): enables machines to understand and generate human language; techniques include text analysis, sentiment analysis, language translation, named entity recognition. - Generative AI (GenAI): creates new content (images, text, audio, video, music) by learning patterns from existing data. - Techniques: transformers (language models like those used by ChatGPT), Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs). - Agentic AI: planning, reasoning, and acting autonomously to complete multi-step tasks; coordinates tools, data, and other agents and adapts to feedback. - Computer vision: interprets visual data using object detection, image segmentation, facial recognition. - Expert systems: mimic decision-making processes of human experts using rule-based systems and knowledge bases. ### Considerations for implementation (Pre-conditions and common challenges) - Data-related: - Ensuring the availability of high-quality, relevant data for training and validating AI models; data privacy and confidentiality challenges. - Data availability issues: outdated formats, inconsistency, reluctance or inability to share information across entities. - Data quality and management: cleaning, preprocessing, augmentation techniques. - Language and domain specificity (for NLP): variations, dialects, domain-specific terminology. - Technical infrastructure and tooling: - Setting up necessary infrastructure including hardware (e.g., GPUs) and software (e.g., PyTorch, TensorFlow, Scikit-learn). - Using NLP libraries/frameworks (e.g., Hugging Face Transformers, spaCy, NLTK). - Tools for model training, deployment, monitoring, and middleware/APIs for integration with PFM systems. - Model selection and maintenance: - Choosing suitable AI models and algorithms (e.g., BERT, GPT, Prophet, XGBoost), conducting thorough testing and validation to ensure accuracy and reliability. - Keeping models updated with new data and linguistic trends; retraining needs. - Security, ethics, and compliance: - Implementing security measures to protect data and AI models; regulatory compliance; anonymization protocols. - Ethical concerns: algorithmic bias towards underrepresented groups; transparency considerations. - Costs and skills: - Managing rolling costs associated with AI deployment, particularly with AI software as a service. - Addressing lack of expertise among staff; continuous training and upskilling. - Implementation complexity: - Managing probabilistic outputs, selecting appropriate indicators, interpreting model interactions. - Bearing significant implementation and maintenance costs depending on dataset size and algorithm complexity. - Problem scoping: - Clearly defining the problem to be solved and assessing AI suitability before selecting the appropriate technology. ### Applications in Public Financial Management (PFM) — use cases and country examples - Core PFM applications: - Compliance monitoring: automatically tracking and verifying compliance with regulations and policies. - Fraud detection: detecting unusual behavior in PFM-related systems to prevent corruption and mismanagement. - Macrofiscal forecasts: improving forecasting accuracy by combining historical data with current trends and larger datasets. - Red flags/anomaly detection: identifying and preemptively addressing potential issues related to government funds. - Automated reporting: summarizing budget documents, translating reports. - Chatbots: citizen interfaces to reduce human support in query resolution. - Information extraction: analyzing text-based data from budget documents and reports. - Financial document processing: automating extraction of information from natural language documents. - Country and project examples: - Brazil and Ukraine — Civil society monitoring: - Ukraine: Dozorro evaluates probability of corruption risks in public procurement; system improves predictive ability, retains suspicions that prove accurate, and automatically adjusts methods when human experts detect new circumvention strategies. - Brazil: Operação Serenata de Amor uses data science and unsupervised learning to identify suspicious uses of public spending and engages the public via a web platform. - Georgia and Peru — Nowcasting: - Working papers (Mgebrishvili, 2022; Tenorio and Perez, 2023) show Ministries of Finance using high-frequency indicators to “nowcast” key economic indicators; ML algorithms improve macrofiscal forecasting and enable inclusion of unstructured and high-frequency data (such as Google Trends). - Guatemala — Cash flow forecasting: - Collaboration between the IMF’s Fiscal Affairs Department and Guatemala’s Ministry of Finance assessing AI for government cash flow forecasting; ML-based approaches can offer meaningful gains in forecast precision. - Methods such as Prophet and XGBoost show strong potential when cash flows exhibit nonlinear patterns, seasonality, and high-frequency volatility. - Rwanda — Budget classification: - Ministry of Finance and Economic Planning uses ML to predict accuracy of budget line classification during planning. - Dataset: more than 40 million transactions based on the Chart of Accounts. - Transitioning to deep learning (neural networks that tokenize and analyze relationships between words in transaction strings) to improve prediction accuracy and ensure budget line items are entered and costed correctly. - Mexico — Government grant search: - Since 2018, Ministry of Finance and Public Credit implemented a search platform on the Budget Transparency website using NLP and CoreNLP to map simple user keywords to available government grants. - Guatemala and Singapore — Chatbots: - Guatemala: Ch@tío (implemented in 2020) supports government suppliers in the e-procurement system; evolved into a private messaging tool in 2022; offers 24/7 services and an estimated response time of one second. - Singapore: Ask Jamie (conceptualized in 2014) implemented across 70 agency websites; GovTech updating chatbots to LLM engines to improve performance and personalization. - United States — Sentiment analysis: - US Treasury’s Bureau of Fiscal Service applies NLP to improve customer experiences; sentiment analysis of e-mails revealed neutral messages dominated, with some negative sentiments related to technical issues; NLP also used to analyze call center audio to set priorities and optimize operations. - Government-wide GenAI pilots and regulatory response: - GenAI pilots: emergency and disaster management (improving weather forecasting by injecting noise into training data and generating mitigation plans), productivity tools (GitHub Copilot). - Many countries expanding regulations and policies to address GenAI risks, using sandboxes and combining human oversight where appropriate. - Uses of GenAI in PFM were mostly experimental as of 2025. ### Technology-specific implementation checklists (condensed) - Machine Learning: - Pre-conditions: defined use case, availability of high-quality relevant data. - Implementation actions: data cleaning and preprocessing, model selection and training (e.g., Scikit-learn, TensorFlow), infrastructure (GPUs), security and compliance measures. - Challenges: data privacy, algorithmic bias, implementation costs, complexity in interpreting probabilistic outputs. - Natural Language Processing: - Pre-conditions: language data preparation, domain-specific term handling. - Implementation actions: choose models/techniques (e.g., BERT, GPT, named entity recognition, sentiment analysis), set up computational resources (Hugging Face Transformers, spaCy, NLTK), integrate via APIs/middleware. - Challenges: language variations, data annotation quality, scalability, ethical bias, regulatory compliance. - Generative AI: - Pre-conditions: curated training data, sandbox environments for experimentation. - Implementation actions: pilot prototypes, combine human oversight, monitor regulatory developments. - Challenges: new risks leading to updated regulations and need for secure experimentation environments. *Source: tnmea2026006-s001 - Section 1* ### Section 2 ### tnmea2026006-s001 - Section 2 ### Examples of government experimentation and deployment - Brazil: The National Treasury has carried out nine proof-of-concepts for GenAI. In liability management, it is being employed to streamline the monitoring of debt novation of the Salary Variations Compensation Fund, automating the analysis of documents received from a state-owned bank and generating two written products. - Singapore: Open Government Products, in-house team of GovTech Singapore, has deployed an AI-based assistant for public officers called Pair. As of 2025, it has served a total of 149,000 users across 195 agencies. The Building and Construction Authority has been using it to speed the review of infrastructure project reports. - United States: In 2024, several agencies were exploring GenAI’s application to contract writing, having requested industry feedback or developed a prototype. ### AI: Generative AI — Applications in Public Financial Management - Advanced simulations: creating and testing different scenarios to support financial planning, infrastructure development, and risk assessment. - AI assistant: supporting various tasks, such as automating budget proposals, generating summaries of budget documents, creating policy drafts, and more. - Analytics and reporting: providing more compelling insights, natural language explanations, and data visualizations through integration into available tools and platforms. - Data augmentation: generating synthetic data to help address data privacy and availability concerns, enabling compliant data analytics and augmenting training datasets for ML. ### AI: Generative AI — Considerations for Implementation - Use case identification: Clearly defining specific use cases for Generative AI within PFM based on the expected added value, avoiding hype-based adoption. - Data collection and preparation: Gathering and curating relevant high-quality datasets based on use cases (e.g., fiscal data for automated reporting) to train models. - Model training and customization: Selecting and fine-tuning Generative AI models (e.g., GPT) to fit specific requirements, ensuring accurate, relevant, and context-aware outputs. - Ethical and regulatory compliance: Establishing guidelines and compliance frameworks to ensure that generated content adheres to ethical standards and regulations. - Integration and deployment: Developing infrastructure and APIs to integrate Generative AI models with existing PFM systems for specific tasks. - Hallucination risks: Addressing AI hallucinations (where it generates content that is plausible but factually incorrect). Creating a controlled learning environment and requiring human evaluation may help. - Quality and model training issues: Ensuring the data used is reliable and the outputs are validated. Training Generative AI models to understand the domain-specific knowledge, which often requires extensive data. - Rolling costs: Managing high operational costs from large-scale use, frequent consultations, pay-as-you-go pricing models, and resource-intensive computations. - Privacy concerns: Addressing privacy issues during model training and use, such as avoiding sensitive data exposure, ensuring compliance, and safeguarding outputs against revealing confidential information. - Integration challenges: Ensuring that the outputs are compatible with existing workflows and reporting standards, which may require consistent monitoring and re-tuning. ### Agentic AI — What they are - Definition: Agentic AI refers to AI-enabled systems that can set or interpret goals, plan steps, take actions through digital tools, and adapt based on feedback from their environment. - Components in practice: - (i) reasoning and planning, - (ii) access to tools and data (such as APIs, databases, document stores, and workflow systems), - (iii) monitoring and learning loops to improve performance over time. - AI agent: an instantiated worker component that can perceive inputs, decide what to do next, and act via tools within a defined scope (for example, a “budget classification agent” that checks and corrects chart-of-accounts mapping). - Agentic AI: the broader system capability and operating model coordinating one or more agents to complete end-to-end tasks, including task decomposition, delegation, verification, escalation, and auditability. - Common configurations: single-agent systems, multi-agent workflows, and autonomous agents operating within predefined guardrails. ### Agentic AI — Current or future uses in PFM (examples) - Albania and Ukraine: Government AI agents; Ukraine has recently deployed Diia.AI, widely cited as one of the world’s first government AI agents. Diia is integrated into the national platform and digital infrastructure to perform actions: it analyzes a user’s request and situation, matches them to the correct public services, and explains documents to facilitate the application process. - Diella (Albania): Previously a virtual assistant on the e-Albania public service platform, was appointed as the Minister of State for AI in 2025. The appointment aims to curb corruption in public procurement, with plans to gradually transfer responsibilities in tender evaluation from human ministers to AI agents. While it does not yet perform fully as an autonomous agent, Diella represents an aspirational step toward applying agentic AI in PFM processes and governance reforms. ### Agentic AI — Applications in Public Financial Management - Service delivery: guiding citizens to the right programs and services through front-line agents, producing pre-filled forms, explaining requirements, and logging decisions and handoffs. - Fraud and waste reduction: providing continuous monitoring across payment systems, triggering alerts, enriching cases with linked data, and routing prioritized cases to inspectors. - Audit and compliance: executing repeatable compliance checks, assembling evidence packs, and creating exceptions to auditors with contexts and recommendations. - Procurement: reviewing bids and documents against relevant laws and criteria, flagging anomalies, and routing cases for approval or escalation. - Decision support for treasuries: generating feasible short‑term actions (such as adjustments to issuance, repos, term deposits, or payment sequencing) within legal, market, and institutional constraints; calculating expected carry costs, probability of breaching cash buffer limits, potential market impacts, and sensitivity to shocks; checking alignment with policy frameworks and governance rules. ### Agentic AI — Considerations for Implementation - Clear use cases and scoping: Establishing clearly defined goals rather than giving open-ended instructions; design task-specific scopes with explicit boundaries to maximize agent performance. - Agent-ready infrastructure: Providing agents with the ability to act through robust APIs or open standards to interact with existing PFM systems. Ensuring data accessibility, governance, and proper permissions. - Expert-in-the-loop governance: Embedding human oversight where agents execute groundwork while human experts review and approve final actions, such as authorizing payments or finalizing audit reports. - Action guardrails and permissions: Defining explicit execution boundaries, including fine-grained permissions, separation of testing and production environments, approval thresholds for high-impact actions, and the ability to suspend or override agent activity instantly. - Auditability and traceability: Building in logging of agent actions, accessed data, tool calls, and decision paths to support internal controls, audits, investigations, and post-incident reviews in line with accountability requirements. - Opacity and "black box" risks: Ensuring explainability by logging agents’ reasoning and decision-making processes to maintain accountability and support auditability. - Accountability and liability risks: Determining responsibility when autonomous agents make fiscal errors, recognizing that human supervisors remain ultimately accountable under PFM laws, which may need updating. - Integration with legacy systems: Adapting agents to safely interact with outdated or brittle PFM systems without causing crashes, data corruption, or workflow disruption. - Cybersecurity and safety concerns: Controlling the expanded attack surface created by tool-using agents, implementing permissions, isolation, monitoring, audit logs, and fail-safes. - Skill and “agent literacy” gaps: Upskilling practitioners to manage rather than operate agents, monitoring their performance, spotting hallucinations, and correcting behavior. ### AI: Computer Vision — What it is - Definition: Computer vision enables computers to interpret and understand visual data from the world around them. It encompasses tasks from image recognition to video analysis and can be combined with LiDAR, Building Information Modeling (BIM), and GPS. - Traditional tools: optical character recognition (OCR) technology can read and extract data from images, videos, and physical documents to automate data entry and document verification. ### Computer Vision — Current or future uses in PFM (examples) - Korea: Image analytics in procurement and climate risks management. Korea is applying AI across sectors, including product image analysis to recommend product codes and an AI-based forest fire management system using satellite images and AI video analysis. Improvements in data availability, predictive accuracy, monitoring, and detection are expected to help better manage fiscal risks from natural disasters and climate change. - Philippines: Project DIME (Digital Imaging for Monitoring and Evaluation) launched in 2018 by the Department of Budget and Management and the Department of Science and Technology to leverage LiDAR and geotagging to compare fund utilization and physical accomplishment of high-value government projects. In 2018, it monitored 75 major projects comprising the construction of access roads to airports and seaports. The government announced plans to relaunch the project as part of its digital transformation agenda. - Peru: The Ministry of Economy and Finance published a national guide on Plan BIM in 2021 to adopt BIM technology in information management for public investment in buildings and infrastructure, aiming to enhance efficiency and reduce errors during planning, design, and construction phases. ### Computer Vision — Applications in Public Financial Management - Document verification: helping prevent fraud and illegitimate transactions by verifying the authenticity of financial documents through watermarks, security features, and signatures. - Public investment management: enabling digital monitoring of the physical world to enhance construction and facility management. - Image analytics: enabling visual data analyses that can help improve expenditure efficiency (e.g., assessing damage and planning rescue operations in emergency). - Data extraction: recognizing patterns and structures of printed or hand-written invoices, contracts, and other financial documents to read, extract, and enter relevant information. ### Computer Vision — Considerations for Implementation - Use case identification: Clearly defining specific use cases for computer vision within PFM, targeting its application to areas where it is appropriate and impactful. - Data collection and annotation: Collecting and labeling relevant visual data, ranging from scanned documents and receipts to satellite images depending on the task. - Tool and technique selection: Selecting appropriate algorithms (e.g., Convolutional Neural Networks), processes (e.g., OCR), and platforms (e.g., OpenCV) that can help address specific PFM challenges. - Data security and privacy: Implementing robust security protocols to protect sensitive visual data, ensuring that data is encrypted, securely stored, and only accessible by authorized personnel. - Complementary technologies: Developing capabilities to combine computer vision with other technologies, enabling an array of applications and more comprehensive analyses. - Visual data diversity: Managing the wide variety and different formats of visual data ranging from scanned documents to real-time video feeds, identifying and taking different approaches for effective processing. - Inadequate models: Establishing thorough evaluation and testing processes, as well as proper error analysis, to ensure the effectiveness and reliability of computer vision models. - Ethical and privacy concerns: Ensuring that computer vision technologies and collected data are handled carefully, addressing bias, especially in areas like surveillance and facial recognition. - Data challenges: Addressing the poor quality of raw material, inaccurate annotation, bad augmentations, and other issues in visual data management. - Integration challenges: Managing the complexities of combining computer vision with other technologies, which may require advanced data fusion techniques. ### AI: Expert Systems — What they are - Definition: Expert systems are specialized computer programs designed to emulate the decision-making abilities of human experts, built on domain-specific knowledge and a set of rules. - Distinction from ML: Expert systems rely on explicitly programmed knowledge and predefined rules; machine learning uses algorithms and statistical models to analyze data and learn patterns. - Strengths: Expert systems excel in tasks that require explicit domain knowledge and adherence to specific rules; machine learning is more effective for complex, data-intensive problems where patterns may not be readily apparent. ### Expert Systems — Applications in Public Financial Management - Budget baseline: helping define budget baselines by automating the analysis of historical budget data, economic indicators, and government priorities. - Knowledge sharing: capturing wisdom and expertise of senior professionals to make it accessible to less-experienced members. - Risk assessment: evaluating legal agreements, historical claims data, and economic conditions to help manage fiscal risks associated with contingent liabilities. - Project prioritization: evaluating project proposals against predefined criteria, scoring and ranking them to help select priority projects for implementation. ### Expert Systems — Core components and implementation notes - Knowledge base: - Facts: A collection of factual information and data relevant to the domain. - Rules: A set of rules, often represented as "if-then" statements, that encode the domain-specific expertise. - Inference engine: Responsible for reasoning and drawing conclusions based on the knowledge stored; uses mechanisms such as forward chaining and backward chaining. - User interaction and access: Interfaces can be command-line, graphical, web-based, or accessed programmatically via APIs to enable integration with other software systems and workflows for automated decision making and data exchange. *Source: tnmea2026006-s001 - Section 2* ### Section 3 ### tnmea2026006-s001 - Section 3 ### Expert systems: typical components and implementation considerations - Typical components: - Explanation facility to provide transparent reasoning behind the system's decisions, helping users understand why a particular recommendation or decision was made. - Knowledge acquisition system that allows domain experts to modify or expand the knowledge base, keeping the system up-to-date and adaptable. - Databases integrated to enable expert systems to access and retrieve additional domain-relevant data. - Control structure that manages the flow and coordinates interactions between the user, knowledge base, and other components. - Error handling and validation mechanisms to detect inconsistencies in data or user inputs and provide feedback or corrections. - Integration with external systems (sensors or other software applications) to gather real-time data or execute specific actions, depending on the application. - Scalability and performance optimization to handle large datasets and deliver timely responses. - Considerations for implementation: - Building a comprehensive knowledge base that includes rules, policies, and financial regulations relevant to PFM. - Defining and refining rules or decision-making logic for reliable and consistent outputs (Rule optimization). - Training the system and requiring rigorous validation and verification procedures to ensure recommendations are aligned with real-world practices (System evaluation). - Providing training for end users to understand how to interact with the system and interpret and apply its recommendations (User adoption). - Establishing a mechanism for continuous updates with new rules, policies, and financial regulations to keep the system current and relevant (System update). - Key challenges and limitations: - Knowledge capture difficulties: Gathering and accurately encoding domain-specific knowledge; tacit or implicit expert knowledge may be hard to articulate. - System limitations: Lack of common-sense reasoning can lead to rigid or overly literal interpretations of rules and technically correct but impractical recommendations. - Maintenance needs: Ensuring adaptability, managing increasing complexity of decision rules, and optimizing performance as the knowledge base grows. - Knowledge uncertainties: Handling incomplete, contradictory, or highly context-dependent real-world PFM knowledge. - Lack of user trust: Need to enhance transparency and ensure the techniques or reasoning behind the system’s decisions are explainable to users. ### APIs in Public Financial Management (technology card) - Definition and role: - API (Application Programming Interface): “a set of rules and protocols that enables software applications to communicate with each other to exchange data, features and functionality (IBM, 2024).” In PFM, APIs enable seamless data exchange between different PFM systems or provide third-party developers with access to government data. - Advantages in PFM: - Transparency: enabling open data and real-time publication for civil society access and reuse, supporting accountability. - Operational efficiency: automating data transfer, reducing manual intervention and duplicative processes, minimizing errors. - Innovation: enabling third-party developers to build applications using government data. - Integration and interoperability: facilitating data sharing among different systems, reducing data silos. - Integration (API First card): ensuring compatibility and interoperability among different systems. - Flexibility: language- and platform-agnostic interfaces supporting microservices-based architectures. - Efficiency: centralized workspace encouraging collaboration and documentation. - Governance: enforcing consistent standards and incorporating security considerations early. - Pre-conditions and implementation considerations: - Data consistency and real-time synchronization: ensure API provides real-time (or near-real-time) updates while maintaining consistency in financial records. - API maintenance and support: adequate documentation, support, and regular updates. - Vendor lock-in and compatibility: avoid reliance on a single API provider. - Security vulnerabilities and compliance: evaluate encryption, authentication, and authorization mechanisms (e.g., OAuth, TLS); ensure compliance with data protection regulations (e.g., GDPR, PCI DSS). - Cost and licensing assessment: assess licensing fees, transaction fees, or usage limits. - Scalability and performance: determine if the API can handle expected transaction volumes and scale as data and usage grow. - Data compatibility and standards: handle required formats and comply with international standards (e.g., JSON, XML, XBRL). - System interoperability: check integration with existing systems such as FMIS and other PFM software, supporting standard protocols (e.g., REST, SOAP). - Quick guides and recommended practices (API First approach, nine essential tasks summarized): - Brainstorming: identify key services and capabilities, types of APIs needed. - Stakeholder Engagement: identify interested parties and secure participation; establish a shared vision. - API Contract: design a contract establishing standards and best practices; document all APIs and interaction points. - Style Guide: create a cohesive style guide for consistency across teams, standardize status codes, version control, and error handling. - API Governance: implement governance to enforce standards and expected results. - Process Automation: use tools to automate documentation generation, style validation, API simulation, and version control. - Mock Generation: generate mock versions to accelerate UX development and promote team independence. - API Self-Service: ensure documentation is easily accessible to development, operations, security, and test teams. - Developer Portal: create a central place for developers to access mocks, tests, contracts, and documentation. - API First: definition and specific implementation challenges: - API First approach: prioritizes design and implementation of APIs before other software development aspects; treats APIs as products rather than middleware. Begins with defining an API contract (tools like Swagger and OpenAPI). - Implementation challenges: - Integration complexity: enabling communication between legacy technologies and new API-driven platforms. - Change management: securing support and buy-in from finance departments, IT teams, and external partners. - Security and exposure risks: risks of exposing sensitive data when external stakeholders are involved. - Inconsistency across APIs: ensuring data consistency and integrity across various APIs. - Maintaining API performance: monitoring performance, availability, response times, and meeting SLAs in high-volume PFM operations. - API standards and documentation: establishing clear guidelines ensuring consistent API design and documentation. - Interoperability and secure architecture: ensuring APIs are interoperable with FMIS and other applications and implementing authentication, authorization, encryption, and data integrity protocols. - Governance and ownership: defining structures for API lifecycle management, versioning, access controls, and documentation updates. - Scalable and flexible architecture: building loosely coupled, scalable architecture to handle growing transaction volumes and requests. ### Current or future uses and country examples - South Africa: - Municipal Money: a web-based tool developed by the National Treasury of South Africa providing extensive financial information based on Section 71 submissions by municipalities. - Includes a citizen-facing website and an API which publishes data in a machine-readable format. - The API makes 17 years of data — on income and expenditure, cash flows, audit opinions, and more, covering over 290 municipalities — readily available to the public. - Data published through the API is free for commercial and non-commercial use, subject to certain limitations. - Enables citizens and researchers to analyze and compare municipal financial performance, supporting civic education and oversight. - Ghana: - Ghana’s Ministry of Finance (MoF) launched initiatives to accelerate the shift toward digital payments and improve financial inclusion, including upgrading the Ghana e-payment portal (GEEP). - Open APIs will be created to connect the GEEP to third-party platforms and integrators. - Reengineering of business processes accompanies infrastructure upgrades. - The MoF collaborates with institutions such as the Ghana Interbank Payment and Settlement Systems (GhIPSS). - GhIPSS achieved the “financial inclusion triangle” by connecting three independently interoperable systems and leveraged APIs to enhance government-to-person payments. - APIs enable real-time data sharing between ministries and payment platforms, reducing delays and improving tracking. - These reforms are broadly aligned with the government’s commitment to interoperability, renewed in 2022 with the second version of the eGovernment Interoperability Framework that highlights API as an interoperability facilitator. - Brazil (state of São Paulo): - Using APIs to connect the public sector costing system to administrative records. - The Secretariat of Education automated data submission processes through APIs, allowing internal systems to gather and send data by programs and school cadasters to the cost accounting system. - Similar API initiatives are underway in other agencies, expected to streamline data collection and consolidation and enhance efficiency and sustainability of the costing system. ### Big Data Analytics: applications, country experiences, and cautions - Definition and key characteristics: - Big data: large volumes of structured, semi-structured, or unstructured data that grow exponentially in complexity and speed. Key characteristics include volume, variety, velocity (speed), veracity (reliability), variability, and value. - Growth driven by technologies such as Internet of Things (IoT) and widespread mobile internet-enabled devices. - Applications in PFM: - Budget planning: providing more accurate forecasts based on historical trends and economic indicators. - Expenditure efficiency: identifying inefficiencies or irregularities to optimize costs and prevent fraud or wasteful spending. - Service delivery: tracking and assessing program impact to enable more personalized and targeted interventions. - Citizen engagement: enabling sentiment analysis using social media or textual data to understand public opinion. - Current or future uses and country examples: - Social protection during COVID-19: - Nigeria relied on high-resolution poverty maps created using deep learning on satellite imagery combined with census data to improve targeting. - Togo combined machine learning with mobile phone and satellite data and other big data sources to expand its flagship social protection program and prioritize the poorest citizens. - Historical and sector uses: - Australian Department of Finance introduced a new approach to subsidy targeting and costing relying on big data in the mid-2010s. - Vietnam has announced plans to invest in big data technologies as part of broader efforts to build modern, transparent digital finance systems. - United States: Social Security Administration analyzes vast unstructured data to identify suspected fraudulent disability claims; Federal Housing Authority uses big data to forecast indicators and build models to maintain positive cash flow (Helms, 2015). - Rwanda: national institutes of statistics lead projects to build big data and analytical capabilities across sectors including PFM. - Ukraine: recent paper presents a generative model of social media post classification for sentiment analysis on a campaign calling for reallocation in local government budgets during wartime; authors caution that big data tools should serve advisory roles and that further research is needed to prevent manipulation and address technological imperfections (Krynytsia et al., 2024). - Cautions and considerations: - Big data tools may be imperfect and subject to manipulation; they should serve an advisory role. - Effective uses of big data often require government- or country-wide efforts and strong institutional leadership. *Source: tnmea2026006-s001 - Section 3* ### Section 4 ### tnmea2026006-s001 - Section 4 ### Big Data Analytics - Definition: Big data analytics refers to the process of examining and extracting actionable insights from large datasets, often using machine learning, predictive modeling, and other analytical applications. - Enablers: Advancements in artificial intelligence, cloud computing, and data storage solutions have significantly enhanced the ability to leverage big data for improved operational efficiency and decision-making. - Considerations for implementation: - Large volume and variety: Managing the large volume and variety of data from multiple sources, ensuring they are compatible and usable. - Privacy and security concerns: Implementing robust security measures to protect sensitive data and ensure compliance with privacy regulations. - Underperformance: Ensuring the analytics system can effectively scale to handle increasing data volumes without degradation in performance, identifying and mitigating relevant bottlenecks. - Data integration complexity: Overcoming complexities of integrating data from disparate sources with varying formats and structures. - Maintenance challenges: Regularly updating analytics tools and systems to adapt to technological advancements and growing data requirements. - Technical and organizational enablers: - Scalability: Establishing a robust data infrastructure capable of handling large volumes of data, such as cloud storage solutions and data lakes. - Interoperability framework: Implementing a framework to integrate diverse data sources into a unified analytics platform, ensuring interoperability across different systems without extensive ETL. - Data quality management: Ensuring processes are in place for data cleaning, validation, and consistency. - Open-source analytics tools: Deploying advanced, open-source analytics tools and technologies (e.g., Apache Hadoop, Apache Spark) for efficient data processing and analysis. - Technical skills: Hiring and training personnel with expertise in big data technologies, data science, and analytics. ### Cloud Computing - Definition: Cloud computing is the delivery of computing services for an IT platform (including servers, storage, databases, software, analytics, and intelligence) over a network, offering faster innovation, flexible resources, on-demand performance, better system efficiency, and economies of scale. - Advantages in public financial management: - Performance and scalability: providing on-demand availability of IT resources and faster access to latest technology, helping address issues associated with traditional FMIS setups and leading to better system performance and scalability. - Interoperability: enhancing interoperability capacities among various systems and modules by offering standard products and supporting interface protocols such as APIs and File Transfer Protocols (FTPs). - Cost savings: offering pay-per-use services, reducing upfront infrastructure costs, helping avoid overbuilding for peak load, potentially leading to lower per-unit cost as computing power gets cheaper in the future. - Easy upgrade and recovery: making CSPs responsible for maintenance, upgrade, and patching, ensuring timely backup and recovery in case of disasters as defined in the contracts. - Types and uses in PFM (examples from practice): - Brazil: Serpro and Dataprev provide cloud services used for several modules of the Integrated System of Federal Government Financial Administration. - Chile: SIGFE and SIAP are administered by the Budget Department and hosted on cloud platforms; the Treasury operates a platform with many modules on cloud networks. - India: e-Lekha (fiscal reporting module) is hosted on IaaS; debt management and internal audit modules are hosted on PaaS. - Cloud service models: - SaaS: provides the entire application stack via the cloud on a subscription basis. - IaaS: provides access to computing infrastructure (e.g., virtual server space, bandwidth, network connections). - PaaS: in addition to IaaS features, provides a platform to design and develop programs and services. - Other: CaaS, DRaaS, FaaS, XaaS. - Deployment models: Public, Private, Hybrid, Sovereign cloud, on-premise cloud, government community cloud. - Pre-conditions and implementation considerations: - Cloud readiness: Assessing suitability of the application/module for cloud (microservices vs monolithic). - Cloud policy: Establishing comprehensive, up-to-date policies to handle sensitive data and set standards/protocols for CSPs. - Data sovereignty framework: Developing clear policies for data residency, ownership, access, and management. - Skills and mindset: Building technical expertise and shifting procurement approaches. - Vendor management: Preparing comprehensive contracts including exit management terms. - Challenges and risks: - Reluctance to change: Resistance due to culture, job security concerns, or lack of trust in the data center jurisdiction. - Data security: Need for clear regulatory framework and robust security measures. - Legacy systems: Integrating existing on-premise systems with cloud services can be complex or costly. - Unforeseen costs and vendor lock-in: Hidden costs (e.g., data processing and transfer) and CSP charges for moving data out of the cloud. ### Databases (for Public Financial Management) - Definition: Databases are structured collections of data managed to facilitate efficient storage, retrieval, and manipulation of information. - Advantages in PFM: - Scalability: designing the right database enables scaling as public financial systems grow. - Efficiency: streamlining collecting, processing, and reporting data, saving time and cost. - Interoperability: facilitating data exchange and integration across systems. - Transparency: making fiscal data more accessible, supporting open data initiatives. - Database types, examples, and common uses in FMIS: - Relational databases — MySQL, PostgreSQL, Oracle Database, SQL Server: For structured data, ensuring data integrity and complex queries. - NoSQL databases — MongoDB, Redis, Cassandra, Neo4j: For unstructured data and integrating diverse sources. - In-memory databases — SAP HANA, Redis, Memcached: For real-time analytics and transactions. - Object-oriented databases — db4o, ObjectDB: For complex data relationships (rarely used). - Graph databases — Neo4j, Amazon Neptune: For relationship analysis in financial transactions and project linkages. - Time-series databases — InfluxDB, TimescaleDB: For monitoring transactions and analyzing time-indexed metrics. - Hierarchical databases — IBM IMS: Typically legacy use cases. - Network databases — Integrated Data Store (IDS), TurboIMAGE: Rare, primarily legacy. - Blockchain databases — BigchainDB, CovenantSQL: Nascent use cases for contract management and public debt settlement. - Considerations for implementation: - Data model complexity: Choosing databases that handle required relationships and data types. - Vendor lock-in risks: Avoiding dependency on proprietary technologies that limit flexibility. - Trade-offs: Balancing consistency versus availability in distributed/real-time systems. - Performance and scalability: Ensuring capability to handle large volumes without degradation. - Maintenance challenges: Performance tuning, updates, backups, and monitoring needs vary by type. - Data requirements and resource management: Understanding structure, volume, licensing, hardware, and operational costs. - Interoperability and security requirements: Integration with existing infrastructure and necessary encryption, access control, and auditability. ### Data Lakes and Data Lakehouses - Roles: Technologies for recording, storing, and managing information for data analytics and business intelligence in PFM. - Typical flow: Ingest → ETL → Data Lake (raw data) → Data Warehouse (processed data for querying and reporting). - Data lakehouse: Hybrid approach combining data lakes (flexibility for different data types) and data warehouses (schema enforcement). - Data lake characteristics: - Scalable architecture. - Tools for data governance. - Centralized catalogs of data and metadata inventory. - Single interface for management and access. - Selective erasure, compression, and deduplication processes. - Adherence to ‘bringing the process to the data’. - Comparison table (preserved terms and symbols): - Data Lake: Type of Data — Any (structured, unstructured, semi-structured); Format — Open format; Cost — Cost$$$$$; Scalability — Scale at large amount of data; Reliability — Varies based on implementation; People — Data scientists; Ease of use — Difficult; Performance — Low. - Data Lakehouse: Type of Data — Any (structured, unstructured, semi-structured); Format — Open format; Cost — Cost$$$$$; Scalability — Scale at large amount of data; Reliability — High; People — Data scientists and machine learning engineers; Ease of use — Easy; Performance — High. - Data Warehouse: Type of Data — Structured; Format — Proprietary; Cost — Cost$$$$$; Scalability — Expensive to scale; Reliability — High; People — Data analysts; Ease of use — Easy; Performance — High. - Implementation considerations and challenges for data lakes: - Large amount of data: Need robust data strategy; use fast access media for recent data and slower media for historical data. - Varying data structure: Manage different structures and formats requiring flexible design. - Data integration challenges: Modeling and cleaning data from diverse sources to ensure quality and integrity. - Low accessibility and underperformance: Enhance response times, may require costly special software. - Clear objectives and data governance: Define use cases, data catalogs, traceability, storage, backup, and restoration strategies. - Security and privacy: Implement robust access controls and data segmentation. - Organizational culture and technical capacity: Foster a collaborative data-driven culture and acquire skilled resources (e.g., data scientists) with training. ### Data Standards - Definition: Data standards are agreed-upon conventions and guidelines for collecting, formatting, sharing, and using data, including data schemas and data specifications. - Purpose: Resolve ambiguity, ensure data quality, consistency, compatibility, and interoperability; define structure and meaning of data to help systems and people interpret it. - Key components: - Data schemas: Define rules and constraints governing data management within a database. - Data specifications: Provide detailed descriptions of usage, flow, semantics, security requirements, and other considerations beyond technical structure. *Source: tnmea2026006-s001 - Section 4* ### Section 5 ### tnmea2026006-s001 - Section 5 ### International standards for financial and statistical data exchange - The International Organization for Standardization (ISO) is an independent organization composed of representatives from national standards organizations. - The ISO has developed and published more than 25,000 standards covering almost all aspects of technology, management, and manufacturing. - ISO 20022: - Provides a common language and a standardized approach (methodology, process, and repository) for financial communications. - Objective: improve electronic data exchange between financial institutions and businesses. - ISO 17369: - Provides an integrated approach to facilitating the Statistical Data and Metadata eXchange (SDMX). - SDMX aims to modernize and facilitate the exchange of statistical data and metadata by setting a standard for describing, structuring, and transmitting data across different systems as well as organizations. - SDMX is an international initiative sponsored by the IMF, World Bank, Bank of International Settlements (BIS), OECD, and the United Nations Statistics Division (UNSD). - Other ISO standards noted: ISO 3166 (country codes), ISO 8601 (date and time format), ISO/IEC 27000 (information security management). ### Advantages of standards in public financial management (PFM) - Data quality: - Ensuring data consistency and uniformity, improving its reliability and accuracy, reducing errors, enhancing data-driven decision making. - Efficiency: - Streamlining the process of collecting, processing, and reporting data, saving time and cost related to transforming or standardizing data. - Interoperability: - Facilitating seamless exchange of data, broadening the scope of data applications, helping integrate different systems, enhancing coordination. - Transparency: - Making fiscal data more accessible and understandable to various stakeholders, supporting open data initiatives, enhancing trust. ### Open data and domain-specific standards (examples) - Open Contracting Data Standard (OCDS): For disclosing and documenting information in public contracting. - CoST Infrastructure Data Standard (CoST IDS): For publishing and validating infrastructure data. - International Aid Transparency Initiative (IATI) Standard: For publishing development and humanitarian data. - Open Fiscal Data Package: For publishing government budget and spending data. - Beneficial Ownership Data Standard: For collecting, sharing, and using high-quality data on beneficial ownership. - Extractive Industries Transparency Initiative (EITI) Standard: For disclosing information on natural resources. - Note: OCDS and CoST IDS were combined to create the Open Contracting for Infrastructure Data Standard (OC4IDS). ### Data standards implementation: pre-conditions, challenges, and considerations - Common implementation tasks: - Data mapping: Conducting a thorough data mapping exercise to align current database fields with international data standards. - Metadata implementation: Adding standardized metadata tags and descriptors to ensure proper data categorization and easier retrieval. - Data transformation (ETL): Implementing ETL (Extract, Transform, Load) processes to clean, standardize, and migrate existing data to the updated schema. - Schema updates: Updating database schemas to accommodate standardized data fields, types, and relationships. - Field name adjustment: Renaming existing database fields to match international standards. - Implementation challenges and pre-conditions: - Integration challenges: Merging data from various sources and formats into a unified, standardized database structure. - Legacy systems: Ensuring new standards are compatible with existing systems, which may have rigid structures. - System downtime: Managing potential downtime and disruptions during database schema updates and migration. - Data quality issues: Ensuring data accuracy and consistency during transformation, particularly with legacy data. - Validation and testing needs: Thoroughly validating and testing the updated database to ensure it meets the new standards and functions correctly across all applications. ### DevSecOps: definition, advantages, and implementation guidance - Definition: - DevSecOps — short for development, security, and operations — is a development approach that integrates security throughout the development lifecycle. It is an extension of DevOps practice. - Advantages in PFM: - Enhanced security: integrating security early to identify and mitigate vulnerabilities, reducing risk of breaches. - Collaboration: fostering collaboration among development, security, and operations teams. - Reputational and financial benefits: enabling early detection of vulnerabilities and reducing post-deployment fix costs. - Faster delivery: automating security checks in the development pipeline to accelerate software releases. - Quick guides: how to implement DevSecOps - Adopt a phased approach: - Begin with assessing the current software development lifecycle and identifying gaps in tools and practices. - Integrate automated security tools into the CI/CD (Continuous Integration/Continuous Deployment) pipeline. - Use an incremental approach (proof-of-concept projects or leveraging existing automation tools). - Integrate Zero Trust Security: - Operates on the premise that no entity is trusted by default; mandates verification from all users and devices. - Incorporate principles and tools such as mutual transport layer security (mTLS) authentication across the software delivery lifecycle. - Tools and activities by software lifecycle phase: - Plan: Team collaboration system, issue tracking system, risk management. - Code: Debugger, build automation, collaborative code review, infrastructure as code. - Build: Dependency management, artifact repository, static application security test. - Test: Automated test tool suite, dynamic application security test, database security scan. - Release: Release packaging tools, configuration integration testing, release validation. - Deploy: Data masking and encryption, configuration automation, post-deployment scan. - Operate: Data and system backup, load balancing, feedback. - Monitor: Compliance monitoring, logging, alerting, InfoSec continuous monitoring. ### DevSecOps: pre-conditions, challenges, and implementation considerations - Challenges and pre-conditions: - Regulatory challenges: Implementing governance frameworks to manage and monitor DevSecOps practices, navigating complex regulatory requirements, ensuring continuous compliance. - Performance issues: Ensuring practices and tools can scale while balancing security with performance and efficiency. - Skills gaps and training needs: Addressing lack of expertise and providing continuous training. - Cultural resistance: Overcoming resistance to change and building a culture of collaboration and shared responsibility. - Integration complexity: Integrating security tools and practices into existing CI/CD pipelines without disrupting workflows. - Considerations for implementation: - Toolchain and technologies: Deploying CI/CD tools (e.g., Jenkins, GitLab CI), security tools (e.g., SonarQube, OWASP ZAP), automation tools (e.g., Ansible, Terraform), and monitoring tools (e.g., Prometheus, ELK Stack). - Cross-functional team: Building a team including DevOps engineers, security engineers, developers, and quality assurance engineers. - Collaborative culture: Building shared responsibility among development, operations, and security teams. - Security practices: Integrating security early, conducting continuous security training, performing regular security audits and compliance checks. - Governance framework: Ensuring compliance with data protection, development practices and cybersecurity regulations. ### DevSecOps applications in PFM - Authorization: enabling multifactor authentication using mobile phone or biometrics, and single sign-on setups. - Data security: role-based access, encryption/decryption for secure access and transmission. - Cash transfer programs: identifying targeted beneficiaries, eliminating duplicated or ineligible accounts, facilitating electronic payments. - Payroll and pension management: eliminating ghost workers, tracking attendance, simplifying enrollment and verification, automating disbursements. ### Digital Identity: definition, role in DPI, and PFM relevance - Definition: - Digital Identity (Digital ID) is an electronic method of establishing authenticity through an authentication protocol acceptable across IT systems, linked to the ID issuing system or an agent system. - Typically involves unique identifiers such as biometrics and passwords; enables tamperproof transaction processing and secure data exchange. - Role in Digital Public Infrastructure (DPI): - Digital ID is a key component of DPI, enabling countries to provide economic opportunities and deliver social services. - Provides a trusted and secure means of identification for access to public services and engagement with government and private entities. ### Current or future uses in PFM: India (Aadhaar) and Ethiopia (Fayda) - India: Aadhaar - Aadhaar is a 12-digit number issued by the Unique Identification Authority of India for digital authentication. - Since its launch in 2010, Aadhaar has been issued to more than 1.3 billion residents of India, representing 95 percent of the population. - Direct Benefit Transfer mechanism: subsidies delivered through either bank accounts (46 percent) or Aadhaar-linked accounts (54 percent). - Aadhaar-enabled attendance system used to eliminate ghost employees and reverse the burden of proof of life for senior citizens, ensuring accurate disbursement of salaries and pensions. - Ethiopia: Fayda - Fayda will provide at least 90 million Ethiopians, as well as refugees and migrants, with a unique identification number based on biometric and demographic information. - Residents holding a Fayda ID can use it to authenticate themselves and obtain a unique tax number. - Since 2023, more than 8 million people have registered for Fayda. - Over 2 million have been authenticated, and over 30 agencies have been integrated. - Fayda is now the main credential for civil servants; efforts are made to use it in health care and banking. - Data protection elements such as the principle of data minimization are embedded in the legal framework for Fayda. ### Digital ID: pre-conditions, challenges, and implementation considerations - Challenges and pre-conditions: - Security and privacy risks: Ensuring robust protection against data breaches, unauthorized access, and privacy violations. - Scalability challenges: Ensuring the system can scale efficiently without compromising performance or security. - Lack of user adoption: Building trust among PFM practitioners and encouraging widespread adoption. - Integration challenges: Integrating the digital ID system with existing PFM systems and databases. - Legal and regulatory issues: Navigating regulatory landscapes and ensuring compliance with national and international laws and standards. - Robust identity verification: Ensuring authenticity via biometric data or secure documents. - Considerations for implementation: - Standards compliance: Ensuring the system complies with international standards (e.g., OpenID Connect, ISO/IEC 24760) and can interoperate with other systems. - Data privacy protection: Implementing comprehensive privacy policies, user consent, data minimization, and zero-knowledge proofs. - Secure infrastructure: Establishing encryption, secure data storage, and reliable authentication protocols. - System and process alignment: Ensuring seamless integration with existing PFM systems and aligning processes to utilize digital IDs effectively. ### Distributed Ledger Technologies (DLT): definition and PFM applications - Definition: - Distributed ledger technology (DLT) refers to protocols and supporting infrastructure that allow computers in different locations to propose, validate, and synchronize transactions across a network. - Uses independent nodes that record and independently update transactions in their respective electronic ledgers; the ledger is shared among the nodes. - DLT facilitates peer-to-peer transactions without a central coordinating entity. - Blockchain is the most well-known DLT; other forms include Directed Acyclic Graphs. - Types of DLT vary based on change-making permissions; ecosystem includes hardware, software, business logic, and protocols for nodes. - Applications in PFM: - Budget monitoring and execution: optimizing budgeting and providing real-time access to reliable data when line ministries operate on systems outside FMIS. - Payments and bank reconciliations: improving efficiency and transparency in payment settlement and clearing houses, digital currencies, cross-border payments, and settlement. - Intra-governmental operations: facilitating real-time settlement of transfers, data exchange, electronic records, and digital ID management across departments. - Procurement: making bidding more transparent, promoting open data on government spending, and creating a level playing field for bidders and vendors. ### Current or future uses in PFM: country examples - Philippines and Poland: - DLT used to streamline bond issuance and enhance transparency in record-keeping. - Philippines has sold tokenized treasury bonds maintained in a blockchain-based registry. - Poland launched a DLT-based bonds reporting system. - Kazakhstan: - In CBDC pilots, DLT is used to issue digital tokens representing central bank reserves in domestic currency. - Kazakhstan’s pilot, the Digital Tenge, has been used for public procurement and delivering school meals. *Source: tnmea2026006-s001 - Section 5* ### Section 6 ### tnmea2026006-s001 - Section 6 ### Assessing whether DLT or blockchain is the right solution for a PFM challenge - Decision flow highlights key yes/no questions: - "Can a traditional database technology meet your needs?" → If "No", continue; if "Yes", "DLT not needed". - "Do contributors know and trust each other?" → Determines public vs permissioned DLT use case. - "Does more than one entity need to contribute or update the data?" and "Would they trust a third party?" - "Can or should the data be public?" and "Do you need copies of the database in multiple computers (due to the risk of attack or censoring)?" - "Do you need to control functionality and decide who can make changes to the software?" - Other specific indicators: "Trying to remove intermediaries or brokers", "Working with digital assets", "Able and wish to create a permanent, authoritative record of the asset". - Outcome guidance: - If most answers are "No": DLT not recommended. - If many answers are "Yes": May have DLT use case (public or permissioned depending on trust). - If mixed answers or infrastructure/governance concerns remain: "Conduct further research on gain-feasibility". ### Distributed Ledger Technologies — Considerations for Implementation - Pre-Conditions and primary implementation considerations: - Fit-for-purpose: Ensuring that the DLT application is needed and appropriate; centralized databases often suffice. - Technological readiness: Adequate infrastructure and standards, network connectivity, computing and data storage capacity, interoperability with existing systems and other DLT platforms. - Security and privacy: Adopt robust measures, structures, and standards to ensure integrity, security, and privacy of the network, ledgers, and their data. - Regulatory framework: Establishing or updating legal and regulatory frameworks based on clear understanding of DLT requirements. - Community adoption: Promoting collaboration among government agencies, technology providers, and other stakeholders. - Challenges and risks (enumerated): - 1 Nascent technology: DLT still not widely regarded as fully mature for large-scale applications. - 2 Governance risks: Mitigating lack of centralized oversight and ensuring trust, transparency, and accountability in community self-governing models. - 3 Data privacy concerns: Challenges to anonymity posed by broadcasting data among all DLT participants. - 4 Compliance challenges: Legal and regulatory compliance difficulties due to underdeveloped frameworks and open network participation. - 5 Transaction costs: Potentially high transaction costs from sharing information, validation, and storage across all nodes. ### Fiscal Transparency Portals — What they are and advantages - Definition: - Digital platforms created by governments to provide public access to detailed fiscal information on budgets, spending, revenues, and debt; may consist of multiple platforms dedicated to specific PFM areas. - Progressive disclosure: Present information in layers of increasing complexity to cater to different audiences using tailored analysis, visualizations, and open data formats. - Advantages in PFM: - Data-driven decision making: Enable open, machine-readable data for in-depth analysis and cross-referencing. - Improved communication: Clear, structured information through interactive visualizations and multiple channels. - Cost reduction in providing information: Automate publication and dissemination, reducing manual dissemination costs. ### Fiscal Transparency Portals — Examples and operational details - Mexico’s fiscal transparency portal: - Integrates over 10 platforms focused on different fiscal data aspects (budget allocations, public investments, performance indicators). - Designed user-centered; allows exploring approved and spent budgets, geolocating public investments, accessing performance metrics through interactive visualizations and open data formats. - "The portal was launched in 2011" and has iterated to remain accessible and relevant; it also provides access to courses on PFM and data. - Cabo Verde’s State-Owned Enterprises (SOE) Manager: - Developed by Unidade de Acompanhamento do Setor Empresarial do Estado (UASE) within Cabo Verde’s Ministry of Finance. - Automates data collection, processing, and analysis; integrates with SOEs' enterprise resource planning software; offers dashboards for key performance indicators. - Part of broader public sector digitalization; utilizes advanced technologies like Microsoft Power BI for data analytics. - Objectives: centralize monitoring, reduce administrative tasks, standardize performance methodologies, facilitate data sharing between SOEs and the Ministry of Finance. ### Fiscal Transparency Portals — Considerations for Implementation - Key implementation challenges: - Balancing information complexity and accessibility: Ensure detail for experts while remaining comprehensible for non-specialists. - Providing data in usable formats: Release data in machine-readable, standardized formats. - Active user engagement: Promote meaningful two-way interaction between government and users. - Maintaining timely and automated updates: Automate updates from various internal systems to avoid delays. - Integrating diverse platforms into a coherent experience: Avoid fragmented user experiences when multiple platforms exist. - Robust financial management information systems: Governments need internal systems that generate, store, and update fiscal data efficiently. - Pre-conditions and enabling factors: - Collaboration with civil society, academia, journalists, and others during development. - Clear legal and policy framework enabling publication and setting transparency standards. - Commitment to open data formats and standards. - User-centered design approach driven by understanding diverse user needs. ### Fiscal Transparency Portals — Current or future uses in PFM (selected country examples) - Italy: Farmer subsidies - GIS analysis via the Italian Informative Agricultural System verifies farmers' declarations to curb overdeclaration of land for subsidies; ensures guarantees and simplification for producers and agencies. - India: Welfare programs - GIS used to plug leakages and wrongful reporting of completion of construction in poor housing schemes; track assets created through wage employment guarantee program. - GIS added to Internal Audit system to support recording of audit evidence and to the asset register module on FMIS. - Latin America: Social audits of public works (Chile, Colombia, Mexico) - "Data on the Streets Rally" (since 2016) uses geolocation of infrastructure projects and project data from public investments management systems to let citizens verify construction progress and post findings on social media; best teams are selected as winners and reports are followed up by responsible authorities. ### Geographic Information Systems (GIS) — What they are and applications in PFM - Definition and value proposition: - GIS are spatial systems that manage, analyze, and map geographic and spatial data layers; most PFM data have geographic attributes. - Value lies in analysis of data layers and revealing patterns, relationships, and geographic context beyond map-making. - Applications in PFM: - Site selection: Identify optimal locations for public infrastructure (roads, bridges, schools, hospitals) based on population density, accessibility, environmental impact. - Monitoring, evaluation, and accountability: Assess progress and outcomes of public investment, evaluate budget program policy outcomes, and facilitate public participation. - Risk assessment: Assess vulnerability of public infrastructure to natural disasters and prioritize investments in resilience and mitigation. - Asset management: Track condition and performance of government assets over time for better maintenance planning and budgeting. - FMIS development or update: Simplify integration of components from various vendors for new or improved systems. - Subnational government benchmarking: Facilitate data sharing across subnational FMIS to assess and compare performance. - Enhanced decision making and budget accountability: Promote coordination among entities and enable independent external audits via standardized data exchange formats. ### GIS — Considerations for Implementation - Challenges: - Data collection challenges: Lack of credible land and property surveys; need competent, accountable surveyors. - Implementation costs: Large direct (software acquisition) and indirect (training) costs. - Inconsistent data standards: Variations in classification, collection, and base map scale affect accuracy. - Lack of data culture and knowledge gaps: Need to establish a data-driven culture; shortage of skilled GIS personnel and knowledge transfer. - Data availability and quality: Governments must ensure sufficient spatial and attribute data quality. - Technical infrastructure: Invest in software, hardware, network capabilities, training, and data security. - Implementation pre-conditions and enablers: - Clear objectives: Distinguish core from less crucial data to reduce time and cost of collection. - Technical skills and capability transfer: Provide on-the-job training, hire GIS specialists, phase roll-outs. - Ongoing technical assistance from technology providers as demonstrated in Belo Horizonte, Brazil. - Use of open-source solutions (e.g., OSM, QGIS) or mobile-based mapping applications as cost mitigants. ### Interoperability — Definition, levels, tools, and implementation considerations - Definition: - Interoperability enables products and systems to "work together" — both technical system-to-system interactions and organizational interactions through meaningful data exchange. - Usually implemented through APIs (Application Programming Interfaces). - Levels of interoperability: - Technical: Basic data exchange capabilities, linking computer systems and services, middleware, presentation, accessibility, and security. - Syntactic: Specifies the format and syntax of data (examples: Electronic Data Interchange, eXtensible Markup Language, JavaScript Object Notation). - Semantic: Provides standard meaning across systems (examples: Chart of Accounts, Fiscal Data Registry (national), ISO 3166, ISO/IEC 11179). - Organizational, legal, political: Involves organizations, legislation, policies, and processes to ensure collaboration and integration. - Example tools and solutions: - X-Road: centrally managed distributed data exchange layer enabling unified and secure data sharing; open-source; originally developed in Estonia and deployed in over 20 countries. - Federated Identity Management (FIM): Share user authentication and authorization data using protocols like OpenID and OAuth to enhance citizen access across agencies and borders. - Implementation challenges: - Integration complexity across diverse architectures and technologies. - Security and privacy concerns when sharing sensitive financial data. - Lack of unique identifiers to link and integrate data. - Data inconsistency across formats and structures. - Maintenance challenges: ensuring low latency, high performance, and regular updates to APIs and middleware. - Implementation measures: - Introduce unique identifiers (e.g., UUID/GUID) to link data across platforms. - Adopt common data standards (e.g., ISO 20022) or generate bridge tables. - Establish data governance frameworks for data sharing, quality, and security. - Develop robust APIs and middleware and perform extensive compatibility testing. ### Microservices — Advantages and definition - Advantages in PFM: - Modularity: Easier to code, maintain, replace, and scale; supports agile development and component reusability. - Flexibility: Allows adoption of new technologies and processes; teams can experiment without overhauling systems. - Resilience: Isolates failures, preventing cascading failures; enhances security through data separation. - Independence: Avoids long-term commitment to a single technology stack; enables best-fit technology selection for each service. - Definition: - Microservices are an architectural and organizational approach where software is composed of small independent services that communicate over well-defined interfaces. *Source: tnmea2026006-s001 - Section 6 (tnmea2026006-s001 - Section 6)* ### Section 7 ### tnmea2026006-s001 - Section 7 ### Microservices: definition and design principles - Microservices break an application into independent components that run each application process as a service. - Each service: - performs a single function; - communicates with other services using lightweight APIs; - is owned by a small, self-contained technical team. - Design goals: - Low coupling (degree of interdependence between modules). - High cohesion (degree to which elements of a module are functionally related). - Benefits: easier to scale, faster to develop, accelerates time to market for new features. - Risks/characteristics: highly distributed systems; modularization supports independent updates, deployments, and scaling of services. ### Microservices for Financial Management Information Systems (FMIS) - Traditional FMIS: monolithic architecture centered on the general ledger; high coupling and sequential processing; susceptible to cascading failures and large-scale loss of service. - Transition approach: - Break FMIS modules (e.g., accounting, budget execution, treasury, cash management) into smaller independent services. - Honduras and Paraguay built FMIS based on microservices, planning to use Docker and Kubernetes to deploy and manage containerized microservices. - Key implementation considerations: - Ensure loose coupling and high cohesion by packaging functions that change together as a single service. - Decentralize design decisions but maintain project-wide standards for cross-cutting functionality (example: logging). - Develop expertise in business domain and microservices design within each team. ### Comparison: Monolithic vs Microservices (as presented) - Structure: Monolithic — Single codebase; Microservices — Individual services. - Domain expertise: Monolithic — May not be necessary; Microservices — Risky without it. - Deployment: Monolithic — Easier with fewer modules; Microservices — Faster but more modules. - Network latency: Monolithic — Minimal or almost zero; Microservices — Can be high. - Modifications: Monolithic — Complex, can break rules; Microservices — Simple and flexible. - Observability: Monolithic — Straightforward; Microservices — Complicated. - Problem detection: Monolithic — Difficult to isolate; Microservices — Easier to isolate. - Appropriate when: - Monolithic: Small, simple application, low complexity required, limited resources. - Microservices: Need high scalability, error isolation important, expect frequent changes. ### Microservices: pre-conditions, challenges, and implementation considerations - Pre-conditions / Organizational readiness: - Promote an innovative culture and sharing of reusable services. - Support with robust and scalable technological infrastructure. - Challenges: - System-level complexity: more moving parts than equivalent monolithic application. - Development challenges: writing and testing many small services; existing tools may not fit new approach; refactoring across service boundaries is challenging. - Network congestion/latency: avoid overly chatty services and long chains of service dependencies. - Unclear governance: decentralized approaches can lead to maintenance difficulties and proliferation of languages/frameworks. - Compatibility issues: poor design, DevOps culture, or code versioning can create breaking updates. - Considerations for implementation: - Careful design to ensure loose coupling and high cohesion, based on comprehensive understanding of business processes. - Decentralization: allow teams flexibility while avoiding rigid communication protocols or shared database schemas. - (Decentralized) Governance: enforce some project-wide standards for cross-cutting concerns. - Skilled workforce: develop team-level expertise in business domain and microservices development. ### Advantages of microservices/Open approaches in public financial management (PFM) - Transparency and accountability: allow private sectors and civil society organizations to analyze financial data and hold government accountable. - Enhanced decision-making: more comprehensive and up-to-date financial data for policymakers. - Increased public participation: information that helps citizens engage in budget and policy-making processes. - Efficiency: better monitoring and evaluation of government projects and expenditures; helps identify and eliminate inefficiencies. ### Open Data: definition, principles, and implementation steps - Definition: digital data that can be accessed, used, and shared by anyone for any purpose, subject at most to attribution and share-alike requirements. - Open data characteristics (must be): non-proprietary, free to use, integral, timely, non-discriminatory, primary, permanent (or preserved), and machine-readable. - Tim Berners-Lee’s five-star deployment program (criteria preserved as stated): - available on the web (in any format) with an open license; - available as structured data; - in a non-proprietary format; - using open standards; - linked to other data to provide context. - EU designation: government expenditure and debt identified as high-value datasets. - Quick guide steps for opening budget data (as presented): - Step 1. Understanding demand — Identify target audience and use cases for open data. - Step 2. Locating data — Identify the source and format of the data to publish. - Step 3. Processing data — Debug and clean the data to ensure its quality. - Step 4. Standardizing data — Verify and use available data standards. - Step 5. Ensuring sustainability — Establish mechanisms and capabilities for updating. - Step 6. Protecting users — Provide publicly available terms and conditions of use. - Step 7. Promoting use — Encourage the use of data and wide participation. - Use case example summary: - What: data on the approved and executed budget according to different expenditure categories. - Who: budget experts, academia, investors, auditors, and more. - Why: create knowledge, enable citizen participation. - How: publish structured data to allow interested actors to carry out their own analysis. - Licensing examples presented: Creative Commons Attribution 4.0 International; Public Domain Dedication and License (PDDL); Open Database License. - Factors to consider for publication: - Update periodicity (ideally every time new information is available, but will depend on processing time). - System integrations, technological infrastructure, time and human resources. - Common errors to avoid: use of special characters, misspellings and homonyms, new categories based on arbitrary criteria, inaccurate aggregations (including subtotals), data misrepresentation, empty rows and columns, lack of a data dictionary. - Implementation actions and mitigations: - Data publication framework: establish criteria for dataset selection, cleaning, and standardization. - Data format conversion: convert data to non-proprietary formats (e.g., CSV, JSON). - Metadata and documentation: create comprehensive metadata to improve discoverability and usability. - Automated reuse: develop APIs for programmatic access to datasets. - Interoperability: ensure linkability with other datasets (e.g., using Uniform Resource Identifiers). - Continuous monitoring needs and preserving low latency for updates. ### Orchestrators and containers: role and usage in PFM - Definition: orchestrators (container managers) automate configuration, coordination, and management of computer and software systems; serve as container platforms; enable management of microservices-based architectures; provide portability across environments. - Containers: small, fast, isolated units packaging applications into immutable images for swift and efficient deployment. - Orchestrators’ role: globally deploy, manage, and configurate containerized applications from a single place. - Usage statistics (as presented): - Containers usage breakdown: - 39% - 37% - 10% - 5% - 41% - 47% - Orchestrators (among the 91% of respondents): - kubernetes: 65% Using in production; 18% Piloting/evaluating. - Other container orchestration tools: 47% Using in production; 41% Piloting/evaluating. - Note: the source labels indicate categories like "Used for most/all applications", "Used for a few applications", "Piloting or evaluating", "Planning to use in a year", "No plans", "Not sure" (percent values presented above). - Advantages in PFM: - Efficiency: optimal packaging of applications, infrastructure reusability, reducing operating costs. - Speed: homogeneous development and test environments, real-time deployments, reduced time-to-market. - Portability: independence from underlying environment or hardware; easier movement across cloud providers; helps avoid vendor lock-in. - Stability and scalability: automated and continuous deployment; adapts to varying loads and requirements. - Available open-source orchestration tools summarized: - Kubernetes: popular platform for deploying, scaling, and managing containerized applications; offers storage orchestration, load balancing, self-healing; large community; steep learning curve. - Docker Swarm: native clustering and orchestration in Docker Engine; simpler and easier to use; relatively limited features. - Apache Mesos: abstracts compute resources for elastic distributed systems; supports container and non-container workloads; designed for large-scale deployment; complex to manage. - Other options: Rancher (user-friendly, supports multiple orchestrators); Nomad (known for simplicity). - Pre-conditions, challenges, and implementation recommendations: - Integration complexity: overcome technical challenges integrating orchestration tools with existing systems. - Security and compliance risks: manage container vulnerabilities and network security; ensure continuous compliance with regulations and standards. - Performance issues: ensure orchestrated environment can scale efficiently; address orchestration overhead bottlenecks. - Upskilling and maintenance needs: train IT staff; regularly update orchestration tools and configurations. - Infrastructure assessment: evaluate existing IT infrastructure for compatibility and identify components needing orchestration. - Skilled personnel: assemble teams with expertise in orchestration platforms (e.g., Kubernetes, Apache Mesos). - Security and compliance protocols and interoperability (APIs/middleware) are essential. - Automated monitoring: deploy automated monitoring tools for performance and health, and implement automated management for scaling, load balancing, and failure recovery. ### Robotic Process Automation (RPA): definition, applications, and PFM examples - Definition: software robots (bots) execute programmed rules to perform routine, repetitive, time-consuming tasks (interacting with applications/systems, making calculations, extracting/copying/inserting data, filling forms, completing analyses and reports). - Integration note: RPA typically uses established control mechanisms and communicates with existing systems; integration often requires minimal changes to existing IT landscapes. - PFM applications highlighted: - Reconciliation: ensure compliance with accounting rules and coherence among systems when reconciliation processes differ in templates or FMIS connections. - Invoice processing: reduce errors in applying payments to right accounts and invoices; bots log into FMIS, open invoice processing screen, copy/paste invoice information after ensuring valid purchase order. - Budget transparency: extract information from multiple data sources (e.g., FMIS of state-owned enterprises) to post on government websites in user-friendly format. - Quality checks: identify discrepancies between budget programming and execution or between accounting system and reporting tools; flag possible mistakes, including in contract compliance. - Country examples and metrics: - United States: - RPAs used across various agencies. - General Services Administration: RPAs to prepare budget justifications. - Defense Logistics Agency: automated data handling for interagency transfers; RPA use helps audit processes by logging bot steps with time stamps. - NASA’s Shared Services Center: bots automate distribution of funds and automate purchase requisition processes. - Finland (Palkeet, as of 2020): - Employed 26 bots to automate 70 processes spanning finance, accounting, payroll, invoicing, and procurement. - Use cases: inspect expense receipts, verify expenditure, reconcile accounting, retrieve and check e-invoices, maintain and update supplier register. - United Arab Emirates (as of 2022): - Ministry of Finance had more than 8 robots automating 63 processes and subprocesses across accounts payable, payroll and pensions, and financial accounting. - Outcomes reported: 95% reduction in errors; 65% reduction in average handling time; saving of 39,000 hours in human labor. - Further expansion of RPA expected to improve efficiency. - Implementation pre-conditions and challenges: - Supporting infrastructure: appropriate IT architecture (varies by RPA technology), hardware capacity, security requirements; virtual environments generally recommended. - Development pipeline: create a pipeline for bot development; focus on proof-of-concept demonstrations for interest and funding; align with national priorities and target labor-intensive, repetitive tasks. *Source: tnmea2026006-s001 - Section 7* ### Section 8 ### Section 8 ### Prioritization and Implementation Considerations for RPA and Automation - Adequate testing - Conduct specific time-bound tests to ensure code functionality before widespread deployment, with periodic testing to minimize the risk of errors during bot execution. - Stakeholder engagement - Identify and take into account the needs of various users and stakeholders (e.g., managers, auditors, technical staff) to facilitate acceptance and utilization. - Technical capacity - Build self-sufficiency of staff to deliver, modify, and scale the automation program through training existing staff, having IT team with RPA expertise, and building partnerships. - Complexity - Manage increased costs and difficulties for implementing complex RPAs involving other technologies such as artificial intelligence. - Usability challenges - Address the lack of intuitive, graphical interfaces; such interfaces, while required for user-friendly implementation, can increase maintenance costs. - Changes in workforce - Anticipate a shift in knowledge and skillset of ministry personnel, with more time spent on specialized financial tasks; replacement of the workforce is not likely in the short term. - Maintenance issues - Create a specialized task force or a strong IT team to provide ongoing support in code updates, security checks, and functionality changes. - Lack of integration - Address the lack of FMIS interconnectivity for complex RPAs, which can pose bigger challenges when building bots that interact with each other. ### Smart Contracts — Definition, Applications, and Country Examples - What they are - A smart contract is “a computerized transaction protocol that executes the terms of a contract (Szabo, 1994).” It refers to a computer program that automatically enforces a contract's terms, encoded in lines of code. Smart contracts can be code-only or ancillary to text-based contracts and are commonly associated with distributed ledger technology (DLT) such as blockchain. Smart contracts may not always be legally binding; enforceability depends on jurisdiction, clarity of encoded terms, and complementarity with legal frameworks. - Applications in Public Financial Management - Automation: automating simple business processes (e.g., generating email alerts for completed transactions) to more complex ones (e.g., streamlining payroll and grant disbursement). - Asset management: tracking, verifying, and monitoring assets; enabling automated transfers and transactions; scheduling maintenance based on asset conditions. - Oversight: facilitating accurate data collection, executing rules for compliance and audit, ensuring immutability of transactions when paired with most DLT. - Procurement: executing contracts, calculating and transferring payments, and automating other tasks based on information stored in procurement systems. - Current or future uses (country examples) - United States: In 2019, the US Treasury explored using blockchain to automate certain manually performed controls — including reconciliations, drawdown approvals, and drawdown limits — using smart contracts. The General Services Administration has been looking to modernize government acquisition by automating some processes through smart contracts; offers submitted by companies are automatically reviewed and evaluated, reducing processing time while increasing transparency and accuracy. - Netherlands: Governments at different levels are implementing blockchain and smart contracts to manage pension and redistribution programs. At the national level, the pension provider is exploring smart contracts to determine rules for pension balances and data access. At the subnational level, a smart voucher system in the Municipality of Groningen programs eligibility criteria and spending conditions (e.g., beneficiary profiles and voucher usage limits) in smart contracts, helping eliminate paper-based processes and reduce manual labor. - Sweden and Georgia: The Swedish Government is working with private partners to build a blockchain-based land registry enabling independent smart contract audits and decentralized identity management; the use of smart contracts disintermediates notaries and reduces transaction time between signing and registering a contract. Georgia’s land registry roadmap includes smart contract functionality to execute, among others, escrow services. Other countries such as Ghana are also exploring similar models. - Considerations for implementation — challenges and pre-conditions - Code failures: Address the possibility of mistakes and errors; one audit revealed critical bugs in 25 percent of smart contracts. - Legal status challenges: Keep up with debates regarding the validity or acceptance of smart contracts as legally binding instruments. - Cybersecurity challenges: Address vulnerabilities to evolving cyber threats. - Stakeholder resistance: Address reluctance of contracting professionals to adopt new technology. - Integration limitations: Gain direct access to “real-world” event data; oracles can help but may dilute decentralization benefits and introduce points of failure. - DLT framework: Establish robust legal, data governance, and privacy frameworks for DLT operations to avoid disputes and confusion. - Risk mitigation: Protect against potential code failures through code review and insurance mechanisms. - Wallet provision: Require linked wallets with sufficient funds; a text-based clause specifying a minimum amount could strengthen legal arguments in payment disputes. - Governing law: Specify applicable laws and jurisdiction to prevent ambiguity and venue shopping. - Hybrid approach: Use a combination of text and code; temporarily use code-only smart contracts for simple transactions until greater clarity around validity and enforceability exists. ### Tokenization — Definition, Applications, and Country Examples - What it is - Tokenization replaces sensitive text or data with unique, secure, non-sensitive placeholders called tokens, representing original information in a secure and manageable way. Often associated with blockchain, tokenization can be implemented across digital systems for finance, health care, online transactions, and identity management. - Applications in Public Financial Management - Budget tracking: real-time tracking of public fund distribution with tokenized allocations to enhance transparency and ensure proper use of public funds. - Asset management: tokenizing public assets for precise tracking and transfer to improve asset visibility and reduce risk of loss or mismanagement. - Procurement: ensuring secure, traceable processes through tokenized contracts to reduce fraud and enhance government purchase efficiency. - Government payments: accelerating payment processing, reducing errors, enhancing security, and making disbursements faster and more reliable. - Current or future uses (country examples) - Slovenia: In July 2024, Slovenia became the first European Union country to issue a sovereign digital bond valued at 30 million euros as part of the European Central Bank’s wholesale central bank money (CeBM) settlement experimentation program. The bond was settled on-chain through the Bank of France’s tokenized cash system and facilitated by BNP Paribas through a private tokenization platform leveraging blockchain and smart contracts. - Philippines: In November 2023, the Philippines sold its first-ever tokenized Treasury bonds to institutional buyers; the government plans to use a widespread mobile payments service to offer tokenized bonds to small retail investors who previously lacked access to government securities. - Luxembourg: Has established a regulatory framework allowing DLTs to issue and circulate dematerialized securities while preventing entities from issuing tokens without a central account keeper or settlement organization. The legal framework recognizes DLT-based token transfers as equivalent to traditional securities account transfers. Example projects include a digital platform fractionalizing forest ownership where each token is priced at one euro and represents a stake in forests bought and managed by the company. - Considerations for implementation — challenges and pre-conditions - Regulatory uncertainty: Navigate complex, unclear regulatory environments and align tokenization practices with evolving legal frameworks and international standards. - Scaling tokenized systems: Design systems for scalability while maintaining performance, reliability, and security as adoption increases. - Data security risks: Protect digital tokens and infrastructure from cyber threats with robust measures and continuous monitoring. - Technological complexity: Manage technical complexity through specialized expertise and rigorous project management. - Implementation costs: Allocate sufficient budget and resources for scalable technology development, infrastructure upgrades, and staff training. - Regulatory framework: Establish clear regulations for the use, management, and oversight of tokenized assets. - Data standardization and digital infrastructure: Develop robust infrastructure and consistent data standards to support large-scale tokenization. - Technology readiness assessment: Conduct thorough assessments to ensure infrastructure and systems are prepared to support tokenization. - Cybersecurity measures: Implement comprehensive cybersecurity protocols including encryption, multi-factor authentication, and regular security audits. - Interoperability standards: Adhere to standards that facilitate seamless integration of tokenized systems across government departments and financial systems. ### Zero-Knowledge Proofs (ZKPs) — Definition, Advantages, and Use Cases - What they are - Zero-Knowledge Proofs (ZKPs) are cryptographic methods allowing a prover to prove knowledge of a value to a verifier without revealing the information itself. Main types include: (1) zk-SNARKs (Succinct Non-Interactive Argument of Knowledge) — highly efficient but require initial trusted setup; (2) zk-STARKs (Scalable Transparent Arguments of Knowledge) — offer post-quantum security and transparency but have larger proof sizes; and (3) Bulletproofs — suited for simpler applications like range proofs. - Advantages in Public Financial Management - Enhanced privacy: protect sensitive financial data while verifying integrity and confidentiality. - Regulatory compliance: verify compliance without exposing confidential details. - Improved security: reduce risk of data breaches by not revealing underlying information. - Fraud prevention: verify transactions and claims without disclosing sensitive information. - Key PFM processes where ZKPs add value - Identity validation for payments. - Social welfare distribution: validate eligibility without exposing personal data. - Public health expenditure verification. - Pension disbursement management. - Secure voting systems and budget execution: validate allocation decisions without revealing voter identities. - Confidential bidding in public procurement. - Data sharing between agencies and cash management. - Emergency relief fund distribution: validate recipients without compromising privacy. - Considerations for implementation — challenges and pre-conditions - Computational requirements: Manage significant computational power needed to implement and sustain ZKP algorithms. - Integration difficulties: Ensure seamless integration with legacy public finance systems. - Regulatory ambiguity: Deal with unclear or evolving regulations concerning advanced cryptographic methods; develop explicit regulations and legal guidelines. - Resource intensive: Secure necessary budget and resources for development, deployment, and maintenance. - Cybersecurity threats: Address sophisticated cyber threats targeting ZKP systems and processes; implement comprehensive cybersecurity measures. - Regulatory clarity: Develop explicit regulations and legal guidelines for ZKP use in public finance. - Initial test phases: Run controlled pilot programs to validate ZKP systems before broad deployment. - Advanced cryptographic infrastructure and data harmonization: Establish robust cryptographic systems and consistent data formats and standards across departments. *Source: tnmea2026006-s001 - Section 8* --- _Source: https://www.imf.org/-/media/files/publications/tnm/2026/english/tnmea2026006-s001.pdf_