India Needs an AI Risk Ledger to Measure Financial Risks

When a boiler exploded in Calcutta in 1863, killing several people, the disaster exposed the dangers of operating powerful machinery without adequate inspection. Within a year, Bengal’s authorities passed Act VI of 1864, introducing India’s first boiler inspection law in the city.

Across the world, a different approach was emerging. In Connecticut, engineers established the Hartford Steam Boiler Inspection and Insurance Company in 1866. Its model combined inspection with financial protection: the company would insure boilers, but it would also examine them to understand the risks and help prevent future accidents.

The distinction was important. Regulation established who was responsible for inspection, while insurance created a financial incentive to understand how failures happened, what they cost and how they could be prevented.

More than a century and a half later, Artificial Intelligence is creating a similar challenge for India’s financial system. Regulators are increasingly clear that financial institutions cannot escape responsibility when AI systems make mistakes. Yet the country still lacks a comprehensive, comparable picture of how often those systems fail, how much damage they cause and whether the same underlying technology exposes multiple institutions to the same risk.

India has started answering who is accountable when AI goes wrong. The harder task is determining what those failures are worth and how to protect the financial system against them.

Why Accountability Alone Cannot Control AI Risk

In July 2025, the Securities and Exchange Board of India (SEBI) introduced amendments to its Intermediaries Regulations that made regulated entities solely responsible for their use of artificial intelligence and machine learning tools. The requirement applies whether an organisation develops its systems internally or purchases them from third-party technology providers.

The rules cover responsibility for protecting investor and stakeholder data, ensuring the accuracy of AI-generated outputs and complying with applicable laws. They also make clear that financial institutions cannot avoid regulatory responsibility simply because an external vendor supplied the technology.

This is an important principle as banks, investment platforms and financial service providers increasingly rely on AI for customer support, investment research, trading-related services, compliance and internal operations.

If an AI system produces an inaccurate recommendation or mishandles a customer instruction, the regulated institution remains answerable for the consequences. It may subsequently pursue contractual remedies against the technology provider, but that does not remove its responsibility to customers or regulators.

However, assigning liability does not automatically make the risk measurable. An institution may know that it must compensate a customer after an AI-related failure without knowing how frequently similar incidents occur across the industry or whether the same technology could cause losses at several firms simultaneously.

Five distinct tasks are involved: assigning responsibility, measuring losses, estimating future exposure, controlling failures and transferring some of the financial risk through insurance. India’s regulatory approach has made progress on the first. The remaining tasks require better information.

What a Kill Switch Can and Cannot Do

The risks associated with AI agents are becoming more visible as systems move beyond generating answers and begin taking actions on behalf of users. An AI agent may search for information, access software, communicate with other systems and execute instructions with limited human intervention.

This creates a different category of operational risk. A chatbot that gives an incorrect answer can mislead a customer, but an agent connected to financial systems may also initiate a transaction, change records or act on an instruction that it has misunderstood.

Visa’s Agentic Commerce Benchmark, published on November 5, 2026, offers an indication of the reliability challenges. Its first round tested 22 AI models from 12 research labs across 15 simulated shopping scenarios. The models failed 42.8 per cent of their test runs.

Visa designed the benchmark to assess whether an AI shopping assistant could identify a customer’s requirements and make an appropriate purchase decision, including choosing not to buy when a product did not meet those requirements. The results showed that models still struggled to consistently establish what a customer actually needed before reaching a decision.

Importantly, these were simulated shopping tasks, not real financial transactions. The results should not be treated as a direct estimate of how frequently banking or investment agents will fail. They nevertheless demonstrate why organisations need to test an agent’s decision-making before granting it authority over consequential actions.

On November 3, 2026, cybersecurity company CrowdStrike announced an AI Kill Switch through its Falcon Guardian platform. The system is designed to let security teams terminate rogue AI agents, prevent them from restarting and cut off their enterprise network access.

Such controls can reduce the damage caused by an agent that behaves unexpectedly. But stopping a system is not the same as reversing its previous actions.

A suspicious process can be terminated. A completed payment, settled trade or loan disbursement may be much harder to reverse. Some transactions can be disputed or corrected, but others may require the institution to compensate the affected customer after the event.

A kill switch is therefore an important operational safeguard, not a substitute for financial protection. Institutions still need to understand what happened, identify who was affected, calculate the loss and determine whether the same weakness exists elsewhere.

Why Insurers Are Hesitant to Cover AI Failures

Insurance companies have traditionally helped businesses manage risks that are too large or unpredictable for a single organisation to absorb comfortably. They collect premiums from many customers, estimate the probability of losses and use that information to determine the price and conditions of coverage.

Artificial intelligence presents a difficult challenge because its failures can be difficult to classify and may affect several businesses at once.

In November 2025, the Financial Times reported that major insurers including AIG, Great American and W.R. Berkley had sought permission from US regulators to introduce exclusions for certain AI-related liabilities in corporate insurance policies. AIG subsequently said it had no plans to use its filing, illustrating that seeking regulatory approval does not necessarily mean an insurer has already implemented an exclusion across its policies.

The concern is not simply that an AI system might make a mistake. Insurers also need to understand whether a failure is isolated or connected to a technology used by thousands of organisations.

Consider a hypothetical situation in which a model supplied by one technology company incorrectly processes financial instructions. If one bank uses that model, the incident may produce a limited number of claims. If hundreds of banks, brokers and payment companies depend on the same model or infrastructure, a single underlying defect could potentially generate losses across the entire network.

This is known as accumulation risk. It is familiar to insurers in areas such as natural disasters and cyber incidents, where one event can affect multiple policyholders. AI introduces a different challenge because the shared exposure may arise from a common model, software update, training-data problem or third-party service.

Insurers cannot confidently price such exposure without evidence about how often these failures occur, their financial consequences and the number of organisations relying on the affected systems.

AI insurance products do exist. Specialist providers, including Armilla, offer policies designed to address liabilities associated with AI model errors, inaccurate outputs and agent actions. Lloyd’s underwriters have also participated in the development of specialist AI liability coverage. Munich Re has separately offered AI performance-related insurance solutions.

However, the availability of specialised cover does not mean that every AI-related loss is insured. Policies can have specific exclusions, conditions, limits and definitions of covered events. The amount of protection available may also be small compared with the potential exposure of a large financial institution.

The central obstacle is the absence of sufficiently comparable loss data. Without a reliable record of AI failures, insurers struggle to distinguish a manageable individual incident from a risk that could affect an entire industry.

Why India’s Financial System Needs a Common AI Incident Ledger

India has several important pieces of the regulatory foundation, but they need to work together. The Reserve Bank of India, SEBI, the Insurance Regulatory and Development Authority of India and the Ministry of Electronics and Information Technology each have responsibilities relevant to AI governance and financial stability.

At present, an AI-related loss may appear in an institution’s operational records as a payment-processing error, a customer-service failure, a compliance problem or a technology incident. Those classifications may be sufficient for ordinary accounting and internal investigation, but they do not necessarily reveal that an AI system caused the problem.

Nor do they consistently identify the model provider, the infrastructure involved or whether other institutions depend on the same technology.

India therefore needs a common AI incident ledger: a structured system through which regulated institutions report significant AI failures and material near misses using comparable categories.

The objective would not be to create another general AI policy document. It would be to build an evidence base that regulators, financial institutions and eventually insurers can use to understand the actual pattern of failures.

What Information Should the AI Risk Ledger Record?

A useful reporting system would need a common set of information across participating institutions. Each report should capture the circumstances of an incident without exposing customer identities, commercially sensitive information or confidential model details unnecessarily.

The core reporting fields could include:

• AI use case: Whether the system was used for lending, investment advice, trading support, payments, fraud detection, customer service or another financial activity.
• Nature of failure: Whether the system produced inaccurate information, misunderstood an instruction, made an unauthorised decision, disclosed data or failed to follow an established control.
• Severity and financial impact: The number of customers affected, estimated losses, confirmed losses, compensation paid and amounts recovered.
• Technology dependency: The model, vendor, cloud provider and relevant infrastructure involved in the incident.
• Human and technical controls: Whether approval procedures, monitoring systems, access restrictions or escalation mechanisms failed.
• Corrective action: What the institution changed to prevent the incident from happening again.

The system should also record near misses. An AI agent that attempts an incorrect transaction but is stopped by a human reviewer may not create a financial loss. However, the event can reveal a weakness that could cause serious damage if the same control is absent in another institution.

Near-miss reporting is particularly valuable because a database containing only completed failures would provide an incomplete picture of how AI systems behave under real operating conditions.

A common taxonomy would allow regulators to compare incidents across sectors while retaining additional categories for banking, securities markets, insurance and payments. Each regulator could establish reporting thresholds suited to the institutions it supervises.

India Already Has the Foundations for AI Incident Reporting

The Reserve Bank of India’s FREE-AI committee report, released in August 2025, provides an important starting point. The committee was established to develop a framework for responsible and ethical use of artificial intelligence in the financial sector.

Its recommendations address several areas, including governance, risk assessment, accountability, data quality, security and regulatory oversight. The report also recommends formal mechanisms for reporting AI-related incidents and adopting a graded supervisory approach.

A graded approach could encourage institutions to disclose genuine mistakes, conduct audits and carry out red-team exercises without automatically treating every first-time failure as evidence of negligence. This would not remove an institution’s responsibility to compensate customers or excuse deliberate misconduct. Instead, it would recognise the value of identifying weaknesses before they become larger problems.

India’s AI Governance Guidelines, released by MeitY on November 5, 2025, also recommended developing a national AI incident database. The guidelines proposed encouraging public and private organisations to report incidents through confidentiality-protecting arrangements, without creating a threat of penalties simply for making a report.

The purpose would be to identify emerging patterns, understand the impact of AI-related harm and improve policy responses. However, recommendations for a national database should not be confused with an already operational, comprehensive reporting system.

The challenge now is implementation. The FREE-AI recommendations and the proposed national incident database need to translate into reporting standards, institutional responsibilities and a practical system for sharing information.

Why the RBI’s April 2027 Operational Risk Rules Matter

A further development makes the timing important. The RBI has set April 1, 2027, as the implementation date for revised operational-risk capital requirements for applicable commercial banks under its Basel III framework.

The revised approach uses a Business Indicator linked to a bank’s financial activity, with historical loss information incorporated into the operational-risk capital calculation under the prescribed requirements.

These requirements are designed to ensure that banks hold capital against operational risks. They do not, by themselves, create an AI-specific incident reporting system.

That distinction matters. If an AI system causes a payment-processing failure, the resulting financial loss may be recorded as an operational incident. The bank may account for the loss correctly and include it in the relevant reporting process.

But unless the incident record identifies AI as a contributing cause, the information needed for future risk analysis may disappear into a broader category. Regulators may see the financial damage without knowing whether it came from a conventional processing error, an AI model, a third-party agent or a failure in human oversight.

The same problem applies to shared technology dependencies. A bank’s records may show that it suffered an AI-related loss, but they may not reveal that several other institutions rely on the same model provider.

India has an opportunity to improve this information before the revised operational-risk framework takes effect. A common reporting taxonomy could help distinguish AI-related losses from other operational incidents and make the resulting information more useful for supervisory analysis.

Why Shared AI Models Could Create Systemic Financial Risk

AI risks are not limited to individual customers or institutions. A failure in a widely used technology provider could potentially affect several organisations at the same time.

For example, multiple lenders might use the same external model to analyse loan applications. If a defect causes the model to misinterpret a particular type of financial information, those lenders could make similar incorrect decisions.

Likewise, payment companies and banks may depend on common cloud infrastructure, third-party software or AI services. A disruption or defect in one shared component could create problems across multiple services.

This is why recording the name of the technology provider is not enough. Institutions need to understand the dependencies behind the service, including the model, hosting infrastructure and other critical components that may be shared with competitors.

Such information would help regulators identify concentration risk: the possibility that a large part of the financial system depends on a relatively small number of technology providers.

It would also help insurers understand whether several policies could be exposed to the same underlying failure. A collection of individually manageable risks can become much more serious when they are connected to one common source.

India’s Digital Payments Scale Makes the Issue More Urgent

India’s position in the global digital payments market makes AI reliability particularly important. According to ACI Worldwide’s 2024 report, India accounted for approximately 49 per cent of global real-time payment transactions in 2023, with nearly 130 billion transactions recorded during the year.

The scale of this activity creates opportunities for AI applications in fraud detection, customer support, transaction monitoring, risk assessment and financial services. It also means that failures in widely used systems could affect large numbers of people and businesses.

India’s technology services industry adds another dimension. Indian IT companies develop, integrate and operate AI systems for financial institutions in domestic and international markets. As these arrangements expand, contracts involving AI performance, liability and indemnification will become increasingly important.

An Indian technology provider may face contractual claims from an overseas bank if an AI system causes losses. Conversely, an Indian financial institution may need to pursue a foreign model provider after a failure while remaining directly accountable to its own customers and regulator.

Without comparable incident data, both sides may struggle to estimate the probability and potential cost of such failures. This can make insurance negotiations more difficult and leave important risks outside existing coverage.

How Regulators Could Build the System Without Creating Another Authority

India does not necessarily need a completely new regulator to establish a common AI incident reporting framework for financial institutions. Existing supervisory bodies already have the authority and institutional experience to oversee their respective sectors.

The Financial Stability and Development Council’s sub-committee offers one possible platform for coordination. It brings together key financial regulators and government representatives to discuss issues affecting the stability and functioning of the financial system.

The RBI, SEBI and IRDAI could work with MeitY through an agreed process to establish a common core taxonomy. Sector-specific regulators could then define the additional information and reporting thresholds required for the institutions they supervise.

Access to the ledger should be carefully divided into three levels:

• Regulatory access: Supervisors would receive detailed incident reports to investigate failures, examine institutional controls and identify shared technology dependencies.
• Industry access: Participating institutions could receive anonymised summaries and aggregate information to understand emerging risks and improve their own safeguards.
• Insurance and research access: Insurers, actuaries and researchers could receive carefully prepared statistics that remove customer data and protect confidential commercial information.

Strong privacy safeguards would be essential. In cases involving rare incidents or highly concentrated technology dependencies, even anonymised information could reveal the identity of a particular institution. Access controls and disclosure thresholds would therefore need to account for the possibility of re-identification.

Reporting incentives would be equally important. If institutions believe that disclosing every mistake will automatically trigger punishment, they may have little reason to report near misses or minor incidents. A carefully designed supervisory approach should encourage transparency while preserving accountability for negligence, misconduct and failures to protect customers.

Why an AI Risk Ledger Could Eventually Support Insurance

A common incident database would not immediately make AI risks easy to insure. Insurance pricing requires sufficient historical information, credible estimates of future losses and an understanding of how different exposures are connected.

A few years of reporting may reveal useful patterns, but it would not automatically produce a reliable actuarial model. AI technology changes quickly, and a model’s behaviour can vary depending on its application, data, updates and operating environment.

Nevertheless, a structured ledger could help insurers answer questions that are difficult to assess today. How frequently do AI agents make consequential errors? Which applications produce the largest financial losses? Are failures concentrated among particular vendors? How often do human approval systems prevent a loss? How many institutions depend on the same technology?

These answers could support better insurance products, more accurate pricing and clearer policy conditions. They could also help institutions negotiate vendor contracts that specify responsibility, reporting obligations, audit rights and compensation arrangements.

GIFT City, India’s international financial services centre, could eventually play a role in bringing additional insurance and reinsurance capacity into the market. Its international insurance ecosystem may provide opportunities for specialised risk-transfer products, although the availability of capital alone cannot solve the underlying problem of insufficient loss data.

Insurers would still need to assess the quality of the evidence, the concentration of exposures and their own capacity to absorb a large number of correlated claims. A ledger would be a foundation for that work, not a guarantee that affordable AI insurance would immediately become available.

The Lesson From the Boiler: Inspection Must Produce Knowledge

The history of boiler safety offers a useful comparison. The creation of Hartford Steam Boiler in 1866 connected inspection with insurance and industrial safety. Over time, accumulated experience helped develop more consistent engineering practices and technical standards.

The American Society of Mechanical Engineers published its first comprehensive Boiler and Pressure Vessel Code in 1915, following work on standardising boiler design, construction, inspection and testing. The National Board of Boiler and Pressure Vessel Inspectors held its first meeting of chief inspectors in 1919, helping bring greater consistency to inspection practices.

The lesson is not that AI and steam boilers are identical risks. Their technologies and failure mechanisms are very different. The connection lies in the institutional process: incidents need to be recorded, examined and converted into knowledge that improves safety and makes risk easier to assess.

India’s financial regulators have begun defining responsibility for AI use. The next step is to ensure that failures are recorded in a way that allows institutions to learn from one another rather than treating every incident as an isolated event.

The RBI, SEBI and IRDAI have an opportunity to agree on a common AI incident taxonomy before April 2027. Such a system could help the new operational-risk framework capture not just the financial cost of past failures but also the technology and dependencies behind them.

A small business wrongly refused a loan, an investor whose instruction is mishandled or a customer affected by an automated payment error needs the institution responsible to correct the harm. But the wider financial system also needs to know whether the same defect is affecting other customers and institutions.

Regulation can establish who answers when an AI system fails. A reliable incident ledger can help establish how much those failures cost, how widely they spread and what can be done to prevent them. Only then will insurers have a stronger foundation for pricing the risk and offering meaningful protection.

Calcutta’s boiler law established the need for inspection. Hartford connected inspection with insurance and the accumulation of engineering knowledge. India has begun defining accountability for artificial intelligence. Its next challenge is to build the evidence system that makes that accountability measurable.

About the Author

Dr. Aditya Vikram Kashyap

Aditya Vikram Kashyap is an AI governance researcher and practitioner whose work focuses on institutional accountability, agentic AI, model risk and the governance of AI in regulated institutions. He is currently Vice President at Morgan Stanley in New York and an award-winning technology leader specialising in enterprise-scale AI, digital transformation and ethical innovation. The views expressed are his own and do not represent any organisation or affiliation, past or present.