Introduction and Background:
The Indian BFSI sector is going through a change because of digitalization and this change is being driven by the use of artificial intelligence and machine learning. Smart computer programs are being used more and more for tasks like checking how likely someone is to default on a loan making trades using computers finding fraud helping customers with their questions and following rules about money laundering. These programs make things faster and more efficient.
They also come with some dangers. The dangers include things like decisions made by machines not knowing how some programs work mistakes made by AI that creates content problems with keeping personal information private attacks on computer systems and relying too much on one company for these tools. To deal with these dangers and to create a set of rules the Reserve Bank of India (RBI) has created two new tools. First on August 13, 2025 the RBI shared a report from a group of experts that talked about how to use intelligence in a responsible and fair way.
This report helps build trust and makes sure that AI is used in a way that’s fair and accountable. Second on June 24, 2026 the RBI shared a draft document that talks about how to manage the risks of using models. This document includes all intelligence and machine learning models, not just the ones used for credit scores. It covers the time from when a model is made to when it is used. The goal is to see how companies, in India should use the rules from the report and the new guidelines. Making sure AI is used in a way and managing the risks it brings is now not just something companies are encouraged to do but it is something they must do.
Objectives:
- To conduct an extensive research of the principles, strategic foundations and recommendations stated in the RBI’s Free AI Framework.
- To analyze the requirements in RBI’s Draft Guidance on Model Risk Management issued in June 2026 with regard to governance, risk tiering and validation independent of the model developer.
- To identify the challenges Indian financial institutions might face while shifting from model validation to AI governance.
- To design a strategy for implementation of the technology innovation in accordance with regulatory requirements.
Analysis:
AI Governance in the Indian BFSI Sector entails an elaborate comprehension of the interaction between the ethical principles and the risk management framework, which are quantified. The regulatory framework of RBI utilizes two means to govern AI in BFSI sector: the ethical “Sutras” of Free AI and operational guidelines of Model Risk Management Framework.
The Free AI Framework: Principles and Pillars:
The Free AI Framework is founded on seven core principles known as “Sutras”. They are as follows:
Trust is the Foundation: AI systems should be built to instill and sustain trust in the financial system’s integrity.
People First: Human decision-makers should have ultimate decision-making power in order to ensure that human interests are at stake when using automated systems.
Innovation over Restraint: The regulation encourages responsible innovation and the development of safe cases, not prohibition.
Fairness and Equity: The models should not show any kind of unlawful bias and discriminatory behaviour towards certain individuals or groups of people.
Accountability: Regulated institutions remain fully accountable for the decisions and actions of their artificial intelligence (AI) systems, regardless of whether they are developed by the institution itself or by third parties outside the institution.
Understandable by Design: AI system should have explainable features like documentations to track and understand the automated logic by all users.
Safety, Resilience and Sustainable: The systems shall be resilient and robust against cyber-physical vulnerabilities and infrastructure disruptions.
The following is how the RBI translates the above seven principles to actionable steps in the form of six strategic pillars: Infrastructure, Policy, Capacity, Governance, Protection, and Assurance. The six strategic pillars contain 26 actionable recommendations altogether.
In terms of innovation and infrastructural recommendations, for example, the RBI proposes to set up common compute environments, AI innovation sandboxes, and sector-specific large language/credit models to give equal access to smaller financial entities. When it comes to regulatory and compliance issues, it insists on adopting a proportionate risk-based approach and applying strict criteria to those applications which have significant risks while being flexible towards low-risk utilities. Model Risk Management Guidance (June 2026) While the Free AI Framework deals more with ethical guidelines, the 2026 Draft Guidance on Model Risk Management contains the operational requirements. The practice of MRM was limited to legacy credit underwriting scorecards before 2026. Yet, the new 2026 guidelines extend this concept by defining model risk management as managing risks from any quantitative method/system/algorithms which input information into processing to come up with statistical estimates/classifications/recommendations. It directly regulates advanced machine learning models, deep learning networks, and generative AI platforms.
The MRM guidance outlines an elaborate model lifecycle governance expectation, composed of several important components. They are as follows:
Model Inventory and Version Control: All regulated firms must create and maintain an inventory system that dynamically tracks all active, inactive, under development and retired models. None of the models can become operational without being formally logged in this register.
Risk-Tiering System: All models developed by the financial institutions need to be classified into low, medium and high-risk categories. Contrary to the traditional approach that relied on materiality of money alone, in the 2026 guidelines, it was clarified that the criteria used to determine risk-tiering must include three important vectors, namely, complexity of the technology, entity’s dependence on the model outputs and degree of autonomy given to the model in taking its own decisions in the absence of human intervention.
Model Validation: An important component of the architecture is the independent validation of all credit and decision-making AI models done by an independent team or a group, which is not associated with the model development process. Prior to model deployment, this independent team evaluates the model for conceptual validity, stability, data lineage and software vulnerabilities.
Human Involvement and Controls Over Model Interventions: Human involvement and control over model interventions are mandatory according to the guidance through well-defined architectures such as “human-in-the-loop” (which requires human review and approval of every automated output) or “human-on-the loop” (where automated systems run on their own, but there are realtime override and kill-switch capabilities to prevent any uncontrolled activity).
Convergence Issues in Indian BFSI:
BFSI Banks and NBFCs in India face unique technical and practical problems in converging Free AI standards and the MRM guidelines. The first issue that arises from this convergence is the problem of the “black-box” nature of the deep learning algorithms and ensembles used. Traditionally, validation was done using coefficients of linearity, but the new AI techniques work through multi-dimensional, non-linear matrices and thus are hard to make explainable. When it comes to the convergent framework, if the logic behind the model cannot be explained and related to human logic, then the principle of “understandable by design” would not be satisfied and would not meet the MRM documentation requirements. The second operational issue is the quality of data, concept drift, and data privacy. The quality of training input data is very crucial when using AI models. Within the Indian credit scenario, using unstructured data like utility bills, social footprints, or local smartphone data might lead to the formation of proxy variables leading to discrimination against marginalised demographics. Models experience “concept drift”, meaning macroeconomic
changes or black swan market conditions cause the real-world data patterns to differ from the patterns from which it was trained, reducing its accuracy and unmonitored financial risk. Finally, the third-party dependency
vector leads to compliance risks. Almost all the Fintech firms, mid sized NBFCs, and cooperative banks do not build their AI models from scratch but rely on third-party tech vendors for specialised models and API services. According to the RBI’s MRM guidelines, relying on AI models provided by third-party vendors does not serve as an excuse for operational failures or non-compliance. When the vendor declines to reveal proprietary training data and weights, citing intellectual property concerns, the regulatory entity becomes responsible for the software’s biases, hallucinations, and anomalies.
Findings and Recommendations:
Following a detailed review of regulatory documentation and sector implementations, the following findings have emerged:
AI risk has evolved far beyond the mere credit loss issue to cover areas of ethical fairness, technical transparency, and algorithmic security, making previous IT governance procedures inadequate.
Ultimate accountability for the governance of AI has clearly been shifted to the top; in this respect, the RBI regards the issue as non delegable at the board level.
Vendor models represent a compliance blind spot because of the lack of access to data lineages and intellectual property structures.
In order to effectively manage the above-mentioned changes, financial institutions operating in India must implement the following strategic roadmaps:
Adopt Board-Approved AI Governance Policy and Committee:
Each regulated institution must urgently devise and implement enterprise-level, board approved AI policy. The process of
responsibility assignment will take place through creation of a dedicated multi-party Responsible AI Committee which will involve representatives from risk management, compliance, data engineering, product development and legal departments. The above-mentioned committee will be responsible for establishing the risk appetite for the institution and will validate all deployments classified as high-risk by the bank.
Develop Dual-Vector Risk Tiers Matrix:
Financial institutions will need to build their internal catalogues for classification of all uses cases for Ai implementation. Each use case must be assigned the proper risk classification (low, medium, high) based on intersection of materiality and operational dependency. Materiality includes financial and customer impact. Operational dependency involves model complexity, degree of automation and lack of human oversight. High-risk deployments like retail loans approval engines and algorithmic collections need to go through validation.
Deploy Advanced XAI and Fairness Toolkits:
Tools that satisfy the transparency requirement include integrating mathematical explainability measures into the development pipeline. Deploying algorithms such as SHAP (Shapley Additive Explanations) or LIME (Local Interpretable Model-agnostic Explanations) would allow developers to decompose complicated non-linear models into human-interpretable feature importance breakdowns. Moreover, fairness testing for disparate impact and demographic parity in the model must be implemented as part of the automatic validation pipeline to avoid algorithmic discrimination due to proxy variables.
Mandate Realtime Red-Team Testing, Drift Monitoring, and Logging AI Incidents:
The evaluation of model risk is not a one-time check. The RE must establish realtime monitoring systems to detect input data drift and change in output distribution. Alerting systems should trigger fall back procedures or manual overrides if there is a fall in model
performance under certain statistical benchmarks. Additionally, regular red-team simulations must be conducted as part of the validation to test the AI model’s robustness against malicious prompt injection and cyberattacks. It can also help to monitor model performance during tail risks in the economy. There should be a logbook for tracking AI incidents and continuously improving institution’s defenses.
Overhaul Third-Party AI Systems Service Level Agreements and Vendor Diligence:
There needs to be a framework for procurement of third-party AI systems. There should be contractual provisions obligating vendors to provide comprehensive model documentation and data lineage, and audit access to models. In case a vendor uses an auditable proprietary AI system, the RE should implement wrapper methods, benchmarking testing, and human-in-the-loop validation procedures to manage these risks.
Conclusion:
Regulatory mandates of the Reserve Bank of India including the August 2025 Free AI report and the June 2026 Draft Guidance on Model Risk Management indicate the dawn of a new era where artificial intelligence shall be regulated in the banking and finance industry. In addition to mandating banks and NBFCs to have policies that regulate the use of AI, this set of guidelines has moved the governance of artificial intelligence from the scope of optional IT innovation to the scope of regulatory compliance and risk management at the board level. The execution of such regulations calls for a delicate balance where although the regulated entity should have strong mechanisms in place like independent validation teams, dynamic risk-tiering and explainability tools, they should not end up hampering the revolutionary benefits of artificial intelligence towards financial inclusion and efficiency of processing across India.
References:
- The Press Release of RBI (2024): Constitution of the Committee on Framework for Responsible and Ethical Enablement (Free) of Artificial Intelligence (AI) in the Financial Sector. Central Office, Mumbai.
- RBI (2025): Report of the Committee on Framework for Responsible and Ethical Enablement of Artificial Intelligence (Free AI Framework). Department of Fintech, Mumbai.
- RBI (2026): Draft Guidance on Regulatory Principles for Model Risk Management. Department of Banking Supervision / Regulation, Mumbai.
- Risk Management Association of India (RMAI). (2026). Special Insights on AI Adoption and Model Risk Vulnerabilities in Indian Commercial Banking. Portal Documentation.