Every earlier part of this series described a central bank that works with data: on prices, on liquidity, on bank balance sheets and on payments. Artificial intelligence (AI) changes how that data can be read, and it changes the financial system the RBI has to regulate. Banks use machine learning to score loans and catch fraud. Fraudsters use generative AI to fake voices and documents. The RBI uses AI itself to nowcast the economy, read complaints and find mule accounts. This final part covers what the RBI has actually done so far: the FREE-AI framework of August 2025, MuleHunter.AI, SupTech, the Reserve Bank Innovation Hub and the June 2026 draft on model risk. It then sets out the risks that the BIS and the Financial Stability Board (FSB) warn about, the tokenised "unified ledger" that may sit under future money, and what all this means for jobs. It ends with a map of the whole series.
In the news
- 7 Oct 2026: Governor flags AI-stock risk. Closing his October policy statement, Governor Sanjay Malhotra listed "risks of an unwieldy correction in valuation of AI stocks" alongside the West Asia conflict, tariffs and high bond yields as factors keeping global sentiment "edgy". Why it matters for the exam: AI is now a macro-financial risk in RBI policy statements, not only a technology topic.
- 24 Jun 2026: Draft guidance on model risk management. The RBI released draft "Guidance on Regulatory Principles for Model Risk Management" for comments until 24 July 2026. It covers all models used by banks, NBFCs, cooperative banks, AIFIs, ARCs and credit information companies, including third-party and AI/ML models, and builds on the FREE-AI report. Why it matters for the exam: this is the RBI's first cross-sector draft rule that applies directly to AI models.
- 29 May 2026: Annual Report sets the AI agenda. The RBI's Annual Report 2025-26 said MuleHunter.AI had been implemented in 26 banks by 31 March 2026, with four more under way. It also announced plans for an institution-grade AI platform with "purpose-built large and small language models" for central banking, and steps towards an AI Innovation Sandbox. Why it matters for the exam: these numbers and plans are likely sources for questions on RBI fintech initiatives.
- 6 Mar 2026: Draft rules on customer liability in digital fraud. The RBI issued draft instructions to widen the rules limiting customers' liability for unauthorised electronic banking transactions. The Annual Report adds a planned compensation mechanism for small-value frauds. Why it matters for the exam: AI-driven fraud has made consumer protection in digital payments a live regulatory issue.
- Feb 2026: India AI Impact Summit, New Delhi. The summit closed with the New Delhi Declaration on AI Impact, which the RBI says was endorsed by 92 countries and international organisations. Why it matters for the exam: it puts India's AI policy, including the IndiaAI Mission (₹10,372 crore, March 2024), in the general-awareness syllabus.
- 31 Dec 2025: Financial Stability Report flags AI-led markets. The December 2025 FSR said the global economy had been supported by "strong AI-related investment" but warned of "the risk of a disorderly market correction". Why it matters for the exam: the FSR is the RBI's main financial stability publication (see Part 15).
- 10 Oct 2025: FSB report on monitoring AI. The FSB published Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector, urging authorities to close data gaps and track AI-related third-party dependencies. Why it matters for the exam: it follows up the FSB's November 2024 report on the four main AI vulnerabilities.
- 7–9 Oct 2025: CAT Sandbox and Unified Markets Interface. At the Global Fintech Fest in Mumbai, the RBI announced a CBDC and Asset Tokenisation (CAT) Sandbox and the Unified Markets Interface (UMI), on which certificates of deposit were issued in tokenised form and settled in wholesale CBDC. Why it matters for the exam: tokenisation is the RBI's next step after the e-rupee pilots.
Why AI matters to a central bank
A central bank is affected by AI in three ways.
- As a user. The RBI collects huge amounts of data: bank returns, payment flows, prices, complaints and speeches. AI can read more of it, faster, and find patterns that people miss.
- As a regulator and supervisor. Banks, NBFCs and fintechs use AI for credit scoring, fraud detection, customer service and trading. If those models are biased, opaque or all alike, the RBI has to deal with the consequences.
- As a guardian of stability and trust. AI can make fraud cheaper and more convincing, concentrate the financial system on a few technology providers and make markets move together. It may also change growth, productivity and inflation, which feed into monetary policy.
The RBI's approach so far tries to do two things at once: encourage innovation and contain risk. The FREE-AI committee called this a "dual focus approach".
The RBI's own AI work
Nowcasting and research
Official GDP data arrive with a lag of about two months after a quarter ends, and policy cannot wait. Nowcasting estimates the current quarter from faster indicators. The Annual Report 2025-26 says the RBI's nowcasting and forecasting models "were further upgraded using AI/ML techniques and data on extreme climate events", including a study of how extreme weather affects vegetable prices. The RBI also used night-time light data from satellites, combined with Google search volumes, to nowcast the GSDP growth of major states. It built an AI/ML framework that uses remote-sensing data to assess crop losses after extreme weather.
Text is the second big source. The RBI has used large language models (LLMs) to analyse complaints under the Reserve Bank Integrated Ombudsman Scheme and social-media feedback, producing bank-wise sentiment measures. The BIS describes the same idea for markets: LLMs can turn news, social media and earnings-call transcripts into sentiment indices that track financial conditions. Central banks also apply text analysis to their own communication, scoring policy statements and minutes for tone, as the ECB's Athena platform does for supervisory documents (see below). For an exam answer, the point is that AI helps a central bank both read the economy and check how its words are received (see Part 9 on the expectations channel).
Supervision and SupTech
SupTech is the use of technology by supervisors. The RBI's Department of Supervision set up an Advanced Supervisory Analytics Group (ASAG) to apply AI/ML. By 2024-25 it had built several models: micro-data analytics, a governance assessment model, a social media monitoring model, a fraud vulnerability index, a borrowers' vulnerability model and an asset quality prediction model. For 2026-27 the department plans a proof of concept using AI to find the root causes of complaints at select banks.
Other recent steps:
- a thematic study on the adoption of AI in supervised entities (2025-26);
- a planned thematic review of cyber risks from AI tools used by supervised entities (2026-27);
- aligning the incident-reporting module of DAKSH, the RBI's supervisory platform, with the FSB's FIRE (Format of Incident Reporting Exchange);
- a cyber mapping exercise across banks and their technology providers, which explicitly covered AI applications.
Operations inside the RBI
The RBI is also bringing AI into its own work. With its IT subsidiary ReBIT (more than 1,200 professionals), it has been adding a conversational AI chatbot to the complaint management system. An internal working group on using AI and ML in internal audit reported in February 2026. The biggest plan, under the strategy document Utkarsh 2029, is an "institution-grade AI platform" with "purpose-built large and small language models that are trained for the specialised needs of central-banking functions".
MuleHunter.AI
A money mule account is a bank account, often rented or opened in someone else's name, through which fraudsters move stolen money. Cutting off mules makes fraud harder to cash out.
- 6 December 2024: the Governor announced MuleHunter.AI, an AI/ML model developed by the Reserve Bank Innovation Hub (RBIH), Bengaluru. A pilot with two large public sector banks had "yielded encouraging results".
- How it works: it is a supervised machine-learning model. It uses behavioural and transactional features, learned from real fraud patterns, to give each account a confidence score: the probability that it is a mule.
- Who acts: the bank. The Annual Report states that what happens after an account is flagged "rests with the individual banks", based on their own assessment and enhanced due diligence.
- Scale: implemented in 26 banks as on 31 March 2026, with four more under way. Next steps include a Mule Registry and linking MuleHunter with the smart registry of the Digital Payment Intelligence Platform (DPIP). In the second phase of DPIP, an AI/ML model will give banks a real-time risk score for each transaction. Seven banks had joined the DPIP Phase I pilot by March 2026.
Worked example: why a flag is not a verdict
The figures here are illustrative, not MuleHunter's actual performance. A bank has 10,00,000 accounts, of which 0.1% (1,000) are mules. A model catches 90% of mules (its recall) and wrongly flags 1% of genuine accounts.
- Mules correctly flagged = 1,000 × 0.90 = 900
- Genuine accounts wrongly flagged = 9,99,000 × 0.01 = 9,990
- Total flagged = 900 + 9,990 = 10,890
- Share of flags that are real mules (precision) = 900 ÷ 10,890 = 8.3%
Even a good model produces mostly false alarms when the thing it hunts is rare. That is why the RBI leaves the decision to banks after due diligence, and why the FREE-AI report insists on human oversight and grievance redress. Freezing all 10,890 accounts would hurt nearly 10,000 innocent customers.
The Reserve Bank Innovation Hub and the sandbox
RBIH
The Reserve Bank Innovation Hub was set up in 2021-22 as a wholly owned subsidiary of the RBI, with an independent board and its headquarters in Bengaluru. It builds technology that the whole sector can use. Besides MuleHunter.AI, its projects include:
- the Unified Lending Interface (ULI), a digital public infrastructure for credit. By 31 March 2026 it had 117 lenders, 134 data services and more than 9.6 crore API calls across 12 loan journeys;
- the Digital Payment Intelligence Platform, described above;
- the FinTech and EmTech repositories, which record fintechs and the use of emerging technologies such as AI, cloud and DLT by regulated entities. By March 2026, 693 fintechs and 33 regulated entities had been onboarded.
The regulatory sandbox
A regulatory sandbox lets firms test new products with real customers under relaxed rules, within limits and under RBI watch. The RBI has run one since 2019. By 31 March 2026, five cohorts had been completed:
| Cohort | Theme | Applications | Shortlisted for testing | Successfully exited |
|---|---|---|---|---|
| 1 | Retail payments | 32 | 6 | 6 |
| 2 | Cross-border payments | 27 | 8 | 4 |
| 3 | MSME lending | 22 | 8 | 5 |
| 4 | Prevention and mitigation of financial frauds | 9 | 6 | 3 |
| 5 | Theme neutral | 22 | 5 | 1 |
| On tap | Closed cohort themes | 22 | 4 | 2 |
| Total | 134 | 37 | 21 |
So about 16% of applicants (21 of 134) have exited successfully. The On-Tap facility now accepts theme-neutral applications. Two newer sandboxes are planned or running: the CAT Sandbox for CBDC and tokenisation (announced October 2025), and an AI Innovation Sandbox (on the 2026-27 agenda, as FREE-AI recommended).
The FREE-AI framework
Origin and committee
On 6 December 2024, the RBI announced a committee to develop a Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) in the financial sector. It was constituted on 26 December 2024 and submitted its report on 13 August 2025. The chair was Dr Pushpak Bhattacharyya of IIT Bombay. The members included Debjani Ghosh (NITI Aayog), Dr Balaraman Ravindran (IIT Madras), Abhishek Singh (MeitY) and Rahul Matthan (Trilegal), with Suvendu Pati of the RBI's FinTech Department as member secretary.
What the surveys found
The committee drew on two RBI surveys:
- Of 612 entities surveyed by the Department of Supervision, 127 (about 21%) were using or building AI. About 27% of 171 NBFCs used AI, and no asset reconstruction company did.
- Of 76 entities in a FinTech Department survey, 67% were exploring at least one generative AI use case, and 85% wanted a regulatory framework.
- Governance was thin. Only about a third of AI users had board-level AI oversight, about a quarter had formal AI incident processes, and only 10% had bias-mitigation protocols.
The seven sutras
The report rests on seven principles ("sutras"). The one-line explanations are the report's own.
- Trust is the Foundation: "Trust is non-negotiable and should remain uncompromised."
- People First: "AI should augment human decision-making but defer to human judgment and citizen interest."
- Innovation over Restraint: "Foster responsible innovation with purpose."
- Fairness and Equity: "AI outcomes should be fair and non-discriminatory."
- Accountability: "Accountability rests with the entities deploying AI."
- Understandable by Design: "Ensure explainability for trust."
- Safety, Resilience and Sustainability: "AI systems should be secure, resilient and energy efficient."
Six pillars, 26 recommendations
The recommendations sit under six pillars, with 13 on each side:
| Group | Pillar | Recommendations |
|---|---|---|
| Innovation enablement | Infrastructure | 1 Financial sector data infrastructure; 2 AI innovation sandbox; 3 Incentives and funding support; 4 Indigenous financial-sector AI models; 5 Integrating AI with DPI |
| Innovation enablement | Policy | 6 Adaptive and enabling policies; 7 AI-based affirmative action; 8 AI liability framework; 9 AI institutional framework |
| Innovation enablement | Capacity | 10 Capacity building in REs; 11 Capacity building for regulators and supervisors; 12 Sharing best practices; 13 Recognising responsible AI innovation |
| Risk mitigation | Governance | 14 Board-approved AI policy; 15 Data lifecycle governance; 16 AI system governance framework; 17 Product approval process |
| Risk mitigation | Protection | 18 Consumer protection; 19 Cybersecurity measures; 20 Red teaming; 21 Business continuity plan for AI; 22 AI incident reporting and sectoral risk intelligence |
| Risk mitigation | Assurance | 23 AI inventory and sector-wide repository; 24 AI audit framework; 25 Disclosures by REs; 26 AI toolkit |
The ideas most likely to be asked about:
- Board-approved AI policy for every regulated entity, with risk classification of AI uses.
- Tell the customer. Consumers should know when they are dealing with AI and should be able to reach a human.
- AI inventory. Each entity keeps a register of its models and use cases, updated at least half-yearly, feeding a sector-wide repository.
- Incident reporting without automatic penalties, so that firms report AI failures early.
- Liability with room to learn. Firms remain liable for customer losses, but supervisors should allow a first or one-off failure to be corrected where the firm followed safeguards such as audits, red teaming and incident reporting. This does not apply to repeated breaches or gross negligence.
- Shared infrastructure: financial-sector data infrastructure as DPI, sector-specific indigenous models as a public good, and an AI sandbox.
The FREE-AI report is a committee report, not a regulation. The RBI said in May 2026 that it was assessing the recommendations to frame "more specific guidance", and it plans a principle-based framework for AI in the financial sector in 2026-27.
From framework to rules: the 2026 model risk draft
The first binding step is the draft Guidance on Regulatory Principles for Model Risk Management of 24 June 2026. Its main features:
- Scope: commercial banks (including foreign banks), small finance, payments, local area and regional rural banks, urban and rural cooperative banks, all-India financial institutions, NBFCs, ARCs and credit information companies.
- Coverage: all models, including those bought from third parties and those using AI/ML, across their whole life cycle from development to retirement.
- Lineage: it builds on the RBI's August 2024 draft on model risk in credit and on the FREE-AI report. The Annual Report said these draft directions were "under final stages of completion".
- Comments: due by 24 July 2026. As of 24 October 2026, the final version had not been issued.
News reports on the draft (for example MediaNama, 25 June 2026) highlight "kill-switch" arrangements, so that an AI system can be overridden, suspended or switched off; human-in-the-loop or human-on-the-loop oversight; due diligence, audit rights and exit plans for third-party AI providers; and, for customer-facing generative AI, controls on hallucination, red-teaming and protection against prompt injection.
Do not confuse this with the payments "kill switch" in the RBI's 2026-27 agenda. That is a proposed facility for customers to block all debits from their own account in one go.
The risks: what the BIS, FSB and RBI warn about
BIS Annual Economic Report 2024
Chapter III of the BIS Annual Economic Report 2024 (30 June 2024), "Artificial intelligence and the economy: implications for central banks", is the standard reference. Its main points:
- Uses: LLM-based sentiment for nowcasting; turning unstructured data (job postings, satellite images, payments) into indicators; supervision (Project Ellipse); anti-money-laundering (Project Aurora found machine learning beat rule-based methods, especially with pooled data).
- Risks: herding, because "reliance on the same handful of algorithms" could amplify volatility and fire sales; concentration on a few providers of cutting-edge models; cyber risks such as prompt injection and data poisoning; and model risk, since "hallucinations are a feature rather than a bug in these models".
- Remedy: central banks should form a "community of practice" to share knowledge, data, best practices and tools, and should invest in staff who combine economics and programming.
FSB, November 2024
The FSB's report The Financial Stability Implications of Artificial Intelligence (14 November 2024) named four vulnerabilities:
- Third-party dependencies and service-provider concentration;
- Market correlations (herding);
- Cyber risks;
- Model risk, data quality and governance.
It also flagged generative AI's potential to increase fraud and disinformation. Its October 2025 follow-up noted that a few global technology firms now control more of the AI supply chain.
The risk map
| Risk | What it means | Indian safeguard (actual or proposed) |
|---|---|---|
| Model risk | Wrong, biased or drifting models; black-box decisions | Model risk draft (June 2026); FREE-AI recs 16, 23, 24 |
| Herding | Many firms use similar models and act alike in stress | Monitoring; RBI's cyber and AI mapping |
| Concentration | Dependence on a few cloud and model providers | Third-party due diligence and exit plans; FSB monitoring |
| Cyber | Prompt injection, data poisoning, AI-powered attacks | FREE-AI recs 19–21; planned thematic review of AI cyber risk |
| Deepfake fraud | Cloned voices and faces used to trick customers and staff | MuleHunter, DPIP, customer awareness, liability rules |
| Consumer harm | Unfair denial of credit; customers not knowing it is AI | Disclosure, human option, grievance redress (rec 18) |
Deepfakes deserve a word. Generative AI can clone a voice or face from a few seconds of material, which makes impersonation scams ("this is your bank manager" or "this is the police") far more convincing. The RBI's response has been layered: MuleHunter to block the exit routes, the bank.in domain so customers can recognise genuine bank websites (with fin.in planned for other financial entities), public campaigns against "digital arrest" scams and the 1930 helpline, and the draft rules on customer liability.
What the fraud data show
The RBI's Annual Report gives bank frauds of ₹1 lakh and above by area of operation. The category most linked to digital channels is "card/internet/digital payments".
Earlier Annual Reports give 3,596 such cases (₹155 crore) for 2021-22 and 6,699 cases (₹278 crore) for 2022-23, so reported digital frauds rose about eightfold in two years before falling.
| Year | Card/internet/digital cases | Amount (₹ crore) | All frauds: cases | All frauds: amount (₹ crore) |
|---|---|---|---|---|
| 2023-24 | 28,836 | 1,452 | 35,800 | 11,013 |
| 2024-25 | 13,332 | 517 | 23,722 | 32,803 |
| 2025-26 | 293 | 29 | 10,114 | 48,021 |
Source: RBI Annual Report 2025-26, Table VI.3.
Worked example: reading the fraud table
- Share of cases, 2023-24: 28,836 ÷ 35,800 = 0.805, so digital frauds were about 80% of all cases (the RBI shows 80.4%).
- Share of amount, 2023-24: 1,452 ÷ 11,013 = 0.132, only about 13% of the money.
- Average size, 2023-24: ₹1,452 crore ÷ 28,836 = ₹0.0504 crore, or about ₹5 lakh per case.
- Average size of an advances fraud, 2025-26: ₹40,774 crore ÷ 8,640 = about ₹4.7 crore per case, almost 100 times larger.
Three cautions apply. First, the table counts only bank-reported frauds of ₹1 lakh and above, so most small UPI and card scams reported to the police through the 1930 helpline are not in it. Second, the year is the year of reporting, not occurrence. The jump in total value in 2025-26 includes 314 older cases worth ₹30,199 crore that were reclassified after a Supreme Court judgment of 27 March 2023. Third, the Annual Report does not explain the sharp fall in digital cases in 2025-26, so do not read it as proof that AI tools ended digital fraud. The exam-safe statement is the RBI's own: by number, frauds were highest in the card/internet/digital category in 2023-24 and 2024-25, while in value they were concentrated in advances in all three years.
How other central banks use AI
- European Central Bank: its Athena platform uses natural-language processing to classify topics, gauge sentiment and recognise entities in supervisory documents. The ECB also uses machine learning to classify more than ten million legal entities by sector and to structure web-scraped price data for inflation analysis (ECB Blog, 28 September 2023).
- US Federal Reserve: runs a staff AI programme with a Chief Artificial Intelligence Officer, an AI compliance plan and a public AI use-case inventory.
- Bank of England: its survey with the FCA (21 November 2024) found that 75% of UK financial firms already use AI, with a further 10% planning to within three years. The top three providers account for 73% of cloud, 44% of model and 33% of data providers named by firms. Its Financial Policy Committee's report of 9 April 2025 identified four areas of systemic risk: AI in core decisions such as credit underwriting, correlated trading in markets, reliance on a few AI service providers, and cyber threats.
- BIS Innovation Hub: runs joint projects for central banks on AI, tokenisation and cyber security. These include Aurora (AML with machine learning), Ellipse (supervisory analytics) and Agorá (tokenised correspondent banking). The RBI joined Project Nexus (linking fast payment systems such as UPI) and, in 2025-26, Project Rialto and Phase 2 of Project Mandala, both on cross-border payments with CBDCs.
Tokenisation and the unified ledger
AI is one half of the technology story. The other is tokenisation: recording money and assets as programmable tokens on a shared platform, so that messaging, reconciliation and settlement happen in one step.
The BIS set out the vision in two annual reports:
- 2023 (Chapter III, "Blueprint for the future monetary system"): a unified ledger "combines tokenised money and assets on a common platform", with central bank money at its core.
- 2025 (Chapter III, "The next-generation monetary and financial system"): money must pass three tests, singleness (settlement at par), elasticity (enough liquidity for payments) and integrity (resistance to fraud and illicit finance). The BIS concluded that stablecoins "fall short of requirements to be the mainstay of the monetary system".
India's moves fit this pattern: the e-rupee pilots since 2022, now with programmable payments for subsidies (see Part 19); the UMI with tokenised certificates of deposit settled in wholesale CBDC; the CAT Sandbox; an MoU on digital assets with the Monetary Authority of Singapore; and planned cross-border CBDC pilots. AI and tokenisation connect. Programmable money needs smart rules, fraud scoring needs real-time data, and both raise the same questions about trust, privacy and concentration that FREE-AI tries to answer.
What it means for jobs and skills
- Exposure is broad. IMF staff estimated in January 2024 that "almost 40 percent of global employment is exposed to AI", about 60% in advanced economies, 40% in emerging markets and 26% in low-income countries. In advanced economies, "roughly half the exposed jobs may benefit from AI integration".
- Banking is highly exposed because much of its work is reading documents, checking rules and answering customers: exactly what language models do well.
- Roles shift rather than vanish. Routine work (KYC checks, data entry, first-line queries) is most at risk. Demand grows for people who can validate models, audit AI, manage data, handle cyber risk and explain decisions to customers and supervisors. The FREE-AI report gives capacity building its own pillar, with recommendations for boards, staff and supervisors.
- For aspirants: an RBI Grade B officer will increasingly supervise models, not just ledgers. Knowing the vocabulary (model risk, explainability, bias, red teaming, AI inventory, SupTech, RegTech) is now part of general awareness and the ESI and Finance papers.
The series in one map: Parts 1 to 20
This is the last part of the series. Each part answers one question about the RBI.
Foundations. Part 1 traced the RBI from the Hilton Young Commission to the e-rupee, and Part 2 set out its functions and governance. Every later part zooms in on one of those functions.
Monetary policy. Part 3 explained the framework. Part 4 covered repo, SDF and MSF, and Part 5 covered CRR and SLR. Part 6 described the MPC, Part 7 inflation targeting, Part 8 the price indices, and Part 9 how policy reaches the economy. AI now feeds this chain through nowcasting and text analysis.
Liquidity and the balance sheet. Part 10 on OMOs, Part 11 on the LAF and MSF, Part 12 on currency, Part 13 on the balance sheet and Part 14 on the surplus transfer show how the RBI creates and manages money.
Stability and the external sector. Part 15 on financial stability, Part 16 on the exchange rate, Part 17 on forex reserves and Part 18 on regulating banks and NBFCs. AI risk is now part of this theme: model risk, cyber risk and herding are stability questions.
The future. Part 19 on the digital rupee and this part on AI show where central banking is heading: programmable money, tokenised assets and machine-assisted supervision, still anchored by the same goals of price stability, financial stability and public trust that the RBI Act set out in 1934.
Common exam traps
FREE-AI is not a law. It is a committee report (13 August 2025) with 26 recommendations. The binding step so far is the draft model risk guidance of 24 June 2026, which was not final as of October 2026.
Seven sutras, six pillars, 26 recommendations. Do not mix the numbers. The 26 recommendations split 13 + 13 between innovation enablement (infrastructure, policy, capacity) and risk mitigation (governance, protection, assurance).
Who built MuleHunter? The Reserve Bank Innovation Hub (Bengaluru), not NPCI and not a private bank. It was announced by the Governor on 6 December 2024. The bank, not the RBI, decides what to do with a flagged account.
RBIH vs ReBIT. RBIH is the innovation subsidiary (MuleHunter, ULI, DPIP). ReBIT is the RBI's IT and cyber-security subsidiary.
Two kill switches. The model-risk draft is reported to require kill switches for AI systems. The payments agenda separately proposes a customer-side kill switch to block all debits.
Unified ledger vs CBDC. A CBDC is central bank money in digital form. A unified ledger is a proposed platform where tokenised central bank money, tokenised deposits and tokenised assets sit together.
Fraud numbers. RBI fraud tables cover frauds of ₹1 lakh and above, by year of reporting. Digital frauds dominate by count, and advances frauds dominate by value.
Key facts at a glance
| FREE-AI | Framework for Responsible and Ethical Enablement of Artificial Intelligence |
| FREE-AI committee | Announced 6 Dec 2024; constituted 26 Dec 2024; report 13 Aug 2025 |
| Chair | Dr Pushpak Bhattacharyya (IIT Bombay) |
| Structure | 7 sutras, 6 pillars, 26 recommendations (13 innovation + 13 risk) |
| Innovation pillars | Infrastructure, Policy, Capacity |
| Risk pillars | Governance, Protection, Assurance |
| Model risk draft | "Guidance on Regulatory Principles for Model Risk Management", 24 Jun 2026; comments to 24 Jul 2026 |
| MuleHunter.AI | RBIH's supervised ML model for mule accounts; announced 6 Dec 2024; 26 banks by 31 Mar 2026 |
| RBIH | Wholly owned RBI subsidiary, Bengaluru, set up 2021-22 |
| Regulatory sandbox | Since 2019; 5 cohorts; 134 applications, 21 exits (31 Mar 2026) |
| AI platform plan | Institution-grade AI platform with LLMs and SLMs (Utkarsh 2029) |
| BIS AER 2024 Ch. III | "Artificial intelligence and the economy: implications for central banks" |
| FSB, 14 Nov 2024 | Four AI vulnerabilities: third-party concentration, market correlations, cyber, model risk |
| BIS AER 2025 Ch. III | Singleness, elasticity, integrity; stablecoins fall short |
| Digital frauds (₹1 lakh+) | 28,836 cases (2023-24), 13,332 (2024-25), 293 (2025-26) |
Exam corner: what to remember
- FREE-AI = 7 sutras + 6 pillars + 26 recommendations, report dated 13 August 2025, chaired by Pushpak Bhattacharyya.
- The first sutra is "Trust is the Foundation". Remember also People First, Innovation over Restraint and Understandable by Design.
- Key recommendations: board-approved AI policy, AI inventory, disclosure to customers, incident reporting, AI innovation sandbox, indigenous financial-sector models.
- MuleHunter.AI: built by RBIH, announced 6 December 2024, in 26 banks by end-March 2026.
- Draft model risk guidance: 24 June 2026. It covers all models, including AI/ML and third-party models, at all regulated entities.
- ASAG in the Department of Supervision builds AI/ML models for supervision. DAKSH is the supervisory platform.
- RBIH: Bengaluru, wholly owned subsidiary (ULI, DPIP, MuleHunter). ReBIT: IT subsidiary.
- BIS AER 2024: herding, concentration, cyber and model risk; "community of practice". FSB 2024: four vulnerabilities.
- Unified ledger (BIS 2023), and the three tests of singleness, elasticity and integrity (BIS 2025).
- RBI tokenisation steps: UMI, tokenised CDs, CAT Sandbox (October 2025).
For exam-specific syllabus and pattern details, see our pages for RBI Grade B, SBI PO and IBPS PO.
Test yourself
The RBI's FREE-AI committee was chaired by
Show answer
Answer: (b) Dr Pushpak Bhattacharyya of IIT Bombay chaired the committee; Dr Ravindran was a member.
How many sutras and recommendations does the FREE-AI framework contain?
Show answer
Answer: (c) Seven sutras guide 26 recommendations under six pillars.
Which of these is NOT one of the three risk-mitigation pillars of FREE-AI?
Show answer
Answer: (d) Infrastructure is an innovation-enablement pillar, with Policy and Capacity.
MuleHunter.AI, used to detect mule bank accounts, was developed by
Show answer
Answer: (c) RBIH, Bengaluru, built it. The Governor announced it on 6 December 2024.
In June 2026, the RBI issued a draft that applies to all models used by regulated entities, including AI/ML models. What is it called?
Show answer
Answer: (b) The draft was released on 24 June 2026, with comments due by 24 July 2026.
According to the RBI Annual Report 2025-26, in how many banks had MuleHunter.AI been implemented by 31 March 2026?
Show answer
Answer: (c) It was in 26 banks, with four more under way. 117 is the number of lenders on ULI.
A model flags mule accounts in a bank with 5,00,000 accounts, of which 500 are mules. It catches 80% of mules and wrongly flags 2% of genuine accounts. Roughly what share of flagged accounts are actually mules?
Show answer
Answer: (c) It catches 400 mules and wrongly flags 2% × 4,99,500 = 9,990 genuine accounts; 400 ÷ 10,390 ≈ 3.8%.
In 2023-24, banks reported 28,836 card/internet/digital frauds involving ₹1,452 crore. The average amount per case was closest to
Show answer
Answer: (b) ₹1,452 crore ÷ 28,836 ≈ ₹0.05 crore, or about ₹5 lakh.
Which BIS publication contains the chapter "Artificial intelligence and the economy: implications for central banks"?
Show answer
Answer: (b) Chapter III of the 2024 report, published on 30 June 2024. The 2023 and 2025 chapters dealt with the future monetary system and tokenisation.
The FSB's November 2024 report on AI listed four vulnerabilities. Which of these was one of them?
Show answer
Answer: (b) The others were market correlations, cyber risks, and model risk, data quality and governance.
Further reading
- Report of the FREE-AI Committee (RBI, August 2025): the full report, with the surveys, the seven sutras and all 26 recommendations explained. Read the executive summary and Chapter 4.
- RBI Annual Report 2025-26, Chapter VI: Regulation, Supervision and Financial Stability: the latest official word on MuleHunter, RBIH, the sandbox, the AI agenda and the fraud tables.
- RBI press release on the draft model risk management guidance, 24 June 2026: the first step from AI principles to AI rules, with a link to the draft.
- BIS Annual Economic Report 2024, Chapter III: Artificial intelligence and the economy: the clearest overview of how central banks can use AI and what can go wrong.
- FSB, The Financial Stability Implications of Artificial Intelligence (November 2024): the global regulators' list of AI vulnerabilities. Short and exam-friendly.
- BIS Annual Economic Report 2025, Chapter III: The next-generation monetary and financial system: tokenisation, the unified ledger and the three tests for money.
- Bank of England, Financial Stability in Focus: AI in the financial system (April 2025): how a major central bank's financial policy committee thinks about AI as a systemic risk.
- IMF Staff Discussion Note, Gen-AI: Artificial Intelligence and the Future of Work (January 2024): the source of the widely quoted estimates of jobs exposed to AI.
Sources
- RBI, Governor's Statement, 7 October 2026: "risks of an unwieldy correction in valuation of AI stocks" (accessed 24 October 2026)
- RBI press release 2026-2027/528, 24 June 2026: draft Guidance on Regulatory Principles for Model Risk Management, scope, coverage of AI/ML and third-party models, link to the August 2024 draft and FREE-AI, comments by 24 July 2026 (accessed 24 October 2026)
- MediaNama, 25 June 2026: reported provisions of the model risk draft on kill switches, human oversight, third-party AI and generative AI controls (accessed 24 October 2026)
- RBI Annual Report 2025-26, Chapter I: FREE-AI assessment, India AI Impact Summit and New Delhi Declaration (92 endorsements), IndiaAI Mission outlay, UMI and tokenised CDs, Projects Rialto and Mandala, MAS MoU, MuleHunter as a supervised ML model and Mule Registry plan (accessed 24 October 2026)
- RBI Annual Report 2025-26, Chapter VI: model risk directions in final stages, MuleHunter in 26 banks, Box VI.2 on FREE-AI, regulatory sandbox table, CAT Sandbox and UMI, RBIH projects (ULI, DPIP, repositories), fraud Tables VI.2 and VI.3 and notes, AI thematic study, DAKSH and FIRE, cyber mapping, Box VI.3 on cyber-enabled fraud, SupTech agenda, draft customer liability instructions of 6 March 2026 (accessed 24 October 2026)
- RBI Annual Report 2025-26, Chapter IX: DPIP phase 2, payments kill switch, institution-grade AI platform (Utkarsh 2029), fin.in domain, ReBIT staff and handbooks (accessed 24 October 2026)
- RBI Annual Report 2025-26, Chapter X: AI/ML nowcasting with climate data, LLM analysis of complaints and social media, night-time lights to nowcast GSDP, crop loss framework (accessed 24 October 2026)
- RBI Annual Report 2025-26, Chapter XI: working group on AI and ML in internal audit, report February 2026 (accessed 24 October 2026)
- RBI Annual Report 2024-25, Chapter VI: Advanced Supervisory Analytics Group and its models, FinTech and EmTech repositories, MuleHunter pilot, AI chatbot in complaint management, card/internet fraud data for 2022-23 (accessed 24 October 2026)
- RBI Annual Report 2023-24, Chapter VI: card/internet fraud data for 2021-22 (3,596 cases, ₹155 crore) (accessed 24 October 2026)
- RBI Annual Report 2021-22, Chapter VI: RBIH set up as a wholly owned subsidiary with headquarters in Bengaluru (accessed 24 October 2026)
- RBI, Statement on Developmental and Regulatory Policies, 6 December 2024: FREE-AI committee proposed; MuleHunter.AI piloted by RBIH with two large public sector banks (accessed 24 October 2026)
- RBI, Governor's Statement, 6 December 2024: announcement of FREE-AI committee and MuleHunter.AI (accessed 24 October 2026)
- RBI press release on the FREE-AI report, 13 August 2025: committee constituted 26 December 2024; seven sutras, six pillars, 26 recommendations (accessed 24 October 2026)
- Report of the FREE-AI Committee: chair and members, survey findings, sutra definitions, list of recommendations, tolerant approach to first-time failures (accessed 24 October 2026)
- RBI Master Direction on NBFC Account Aggregators, 2 September 2016: date of the account aggregator framework (accessed 24 October 2026)
- RBI press release on the Financial Stability Report, December 2025: AI-related investment and risk of a disorderly market correction (accessed 24 October 2026)
- BIS Annual Economic Report 2024, Chapter III: AI uses and risks for central banks, Projects Aurora, Ellipse and Agorá, community of practice (accessed 24 October 2026)
- BIS Annual Economic Report 2023, Chapter III: unified ledger definition (accessed 24 October 2026)
- BIS Annual Economic Report 2025, Chapter III: singleness, elasticity and integrity tests; stablecoins verdict (accessed 24 October 2026)
- BIS media release, 1 July 2024: RBI to join Project Nexus (accessed 24 October 2026)
- FSB, 14 November 2024: four AI vulnerabilities and recommendations (accessed 24 October 2026)
- FSB, 10 October 2025: monitoring AI adoption and third-party concentration (accessed 24 October 2026)
- Bank of England and FCA, AI in UK financial services, 21 November 2024: 75% of firms using AI; provider concentration figures (accessed 24 October 2026)
- Bank of England, Financial Stability in Focus, 9 April 2025: four areas of AI-related systemic risk (accessed 24 October 2026)
- ECB Blog, "Careful embrace: AI and the ECB", 28 September 2023: Athena, entity classification, web-scraped prices (accessed 24 October 2026)
- Federal Reserve Board, Artificial Intelligence at the Federal Reserve: AI programme, Chief AI Officer, compliance plan, use-case inventory (accessed 24 October 2026)
- IMF Blog, 14 January 2024: AI exposure estimates of 40%, 60%, 40% and 26% (accessed 24 October 2026)