Banking supervision rarely looks dramatic from the outside. Behind the policy directives and risk assessments, however, lies painstaking work: combing through thousands of credit files, evaluating risk management architectures, and verifying compliance line by line so that depositors stay protected and macroeconomic stability holds. For decades, supervisors quietly accepted that some risks would slip through the cracks of this labor-intensive process. The failures of Wirecard, 1MDB, and Silicon Valley Bank, along with the rise of AI-driven identity fraud, have shown how costly those cracks can be. Delayed intervention and flawed risk classification are no longer tolerable in an environment where data arrives in torrents and technology can finally keep pace with it.

Artificial intelligence has therefore moved from the margins of supervisory conversation to its center. The question for Nepal Rastra Bank is no longer whether AI belongs in supervision, but how to adopt it deliberately, safely, and on terms that strengthen rather than substitute supervisory judgment.

Central Banks are on the move
AI in central banking spans machine learning, deep learning, natural language processing, and generative AI, and it now forms a core pillar of supervisory technology, commonly called SupTech. The OECD estimates that AI adoption could lift financial sector productivity by roughly twelve percent over the coming decade. More telling than the projections, though, are the working tools already in production. The Central Bank of Brazil runs Adam, which scans the credit portfolios of supervised entities to flag exposures with inadequately recognized expected losses, alongside Axis, which extracts and processes financial statements from PDF reports automatically. The European Central Bank deploys Athena to search and summarize confidential supervisory documents and Heimdal to assess risk profiles when bank directors are appointed. The Bank of Thailand analyzes the board meeting minutes of commercial banks with AI, Australia runs real-time market anomaly detection, Singapore feeds news and financial statements into a machine learning risk identification platform, and the Netherlands central bank built ChatDNB, a generative AI assistant that lets supervisors navigate thousands of pages of laws and regulations in seconds.

These examples share a common philosophy. None of them replaces the supervisor. Each one compresses the time between data and insight, freeing experienced examiners to concentrate on judgment-intensive questions that no algorithm can answer.

World Bank's stocktake on emerging markets
A 2025 World Bank report prepared for the G20 Finance Ministers and Central Bank Governors, drawing on a survey of twenty-seven financial sector authorities across emerging market and developing economies, offers the most systematic picture yet of where peer regulators stand. Adoption by financial institutions remains nascent and uneven, with fintech firms running ahead of traditional banks. The most common industry applications are customer service chatbots, fraud detection, and AML/CFT and KYC compliance, while credit scoring carries particular weight in markets where large populations lack formal credit histories, a pattern with obvious resonance for Nepal.

Source: World Bank Survey on AI in Supervision, 2025

On the supervisory side, the report finds that basic generative AI tools have already achieved widespread uptake among regulatory staff for drafting, summarization, and data analysis, even though the application of AI to core supervisory tasks such as on-site and off-site inspection, asset quality review, and anomaly detection remains in its infancy. Roughly a quarter of surveyed authorities have a formal internal AI policy, and most of the remainder expect to adopt one within twelve months. The direction of travel is unmistakable: AI adoption for supervision is a board-level strategic priority for three quarters of respondents.

Source: World Bank Survey on AI in Supervision, 2025

The constraints are equally instructive. Data privacy and security, internal skills gaps, model-specific challenges around validation and explainability, and the difficulty of integrating AI with legacy IT infrastructure rank as the leading barriers. Authorities also worry about dependence on a small set of global cloud and AI vendors, with implications for data sovereignty and concentration risk. Crucially, the report anchors everything in two principles: AI must never replace human supervisory judgment, and financial institutions must remain fully accountable for decisions driven by their model outputs.

Source: World Bank Survey on AI in Supervision, 2025

Nepal's stands today
Nepal Rastra Bank has not yet institutionalized AI within its supervisory functions, but the technology is already in the building. Individual supervisors increasingly use generative AI tools on an ad hoc basis to check policy compliance against directives, polish language, summarize reports, and structure data. This organic uptake saves time, yet it carries serious risks around data confidentiality and security when sensitive supervisory information touches public AI platforms. The experience of South African authorities is instructive here: recognizing that it would be difficult to stop the train, they moved quickly to provide sanctioned, internally governed tools with safeguards that keep sensitive data local.

Two policy developments make this an opportune moment. The Government of Nepal has approved the National AI Policy, 2082, creating an enabling national framework. And NRB, fulfilling a commitment made in the Monetary Policy for FY 2082/83, has released its draft Artificial Intelligence Guidelines for public consultation. The draft applies to all licensed banks and financial institutions as well as payment system operators and providers, and it is encouragingly comprehensive. It requires cross-disciplinary AI steering committees and board approval for AI deployment, mandates incident and annual reporting to the supervision wing, introduces risk categorization across vectors such as severity of potential loss and breadth of impact, and insists on model testing, explainability, data minimization, explicit customer consent, and disclosure whenever customers interact with an AI interface. The explainability requirement deserves particular emphasis: if an algorithm approves or rejects a loan, the institution must be able to state the exact basis of that decision. This positions Nepal sensibly within the emerging global consensus that supervisory expectations should be technology-aware without becoming technology-prescriptive.

Pragmatic roadmap for NRB
Regulating the industry's use of AI is only half of the task. The other half is building NRB's own supervisory AI capability, and here the guiding philosophy should be to think big and start small. The first step is to issue internal guidelines and an operational strategy on staff AI usage without delay, with particular attention to sensitive departments such as Bank Supervision, so that the current ad hoc experimentation is channeled into safe and sanctioned practice. The second is to launch small pilot projects in areas where processes are already mature and data is reliable, since the World Bank evidence shows that AI tools succeed when embedded in established workflows and languish when bolted on as optional extras. Document summarization, directive compliance checking, and internal knowledge chatbots trained on circulars and supervision manuals are natural starting points.

The third priority is foundational: data integration, data quality, and secure IT infrastructure are prerequisites, not afterthoughts, and several peer authorities have wisely refused to build AI prototypes until their underlying databases were modernized. Fourth, capacity building must bridge the divide between supervisors and data scientists. Embedding technical specialists within supervisory teams, as the ECB and others have done, transfers knowledge in both directions and produces tools that are operationally relevant rather than merely clever. Partnerships with universities and international bodies can supplement scarce in-house talent, a constraint every public authority in the region shares. Finally, because AI risks are borderless, NRB should deepen its exchange of information and practice with peer central banks, regional groupings, and standard-setting bodies, learning from documented experiences in Brazil, South Africa, and Europe rather than discovering every pitfall independently.

Supervisor's role is prominent
The promise of AI in supervision is real: real-time risk alerts instead of delayed post-facto reporting, holistic risk pictures that connect unstructured data with financial statements, predictive power that compensates for chronic human resource constraints, and better credit access for thin-file customers and small enterprises. The risks are equally real, from hallucinated outputs and opaque models to vendor lock-in and cyber exposure. The way to hold both truths at once is to keep the supervisor firmly in charge. AI should sharpen supervisory judgment, never substitute for it, and every AI-assisted decision must remain explainable and legally defensible. With its draft guidelines now open for consultation and a national AI policy in place, Nepal has a genuine opportunity to leapfrog rather than lag. The institutions that benefit most from this transition will be those that prepare early, govern carefully, and treat AI as a disciplined extension of supervisory craft rather than a shortcut around it.