Introduction

Part I of this post discussed how SEBI’s existing rules fail to classify AI-powered trading platforms, since those rules were built for human advisers and basic pre-programmed systems. It showed how platforms exploit this gap by branding themselves as technology products to avoid registration. This part  draws on regulatory approaches from the EU, Singapore, and the United States to propose four design principles for an amendment that would bring these platforms within SEBI’s existing framework.

Comparative Framework and Characteristics of an Adequate Regulatory Definition

The failure of the Indian framework is not unique. Different jurisdictions have responded by choosing different regulatory entry points. Having identified the shortcomings and their causes, this section draws five principles from three jurisdictions: the European Union (EU), Singapore, and the United States of America (USA) to understand the distinct classifications of the AI regulatory system. The comparative analysis does not point to a single model but reveals a set of recurring regulatory techniques.  These considerations can be translated into design characteristics for an Indian amendment framed under SEBI’s existing statutory powers. Before setting out those characteristics, however, it is important to be clear about both the possibilities and the limits of such an amendment.

The suggested amendment deals with (i) closing the classification gap that lets AI investment platforms avoid regulation by calling themselves technology products rather than financial intermediaries, and (ii) giving SEBI a clear legal basis, through a functional definition, to bring these platforms under existing adviser and broker rules and hold them accountable for conduct, disclosure, registration, and enforcement obligations.

At the same time, classification alone cannot resolve the AI-specific concerns that emerge once these platforms are brought within existing regulatory categories. Problems such as systemic risk arising from correlated model behaviour across platforms, governance obligations arising from continuous automated execution, technical standards necessary for meaningful explainability, and the allocation of liability across layered vendor relationships remain inadequately addressed by simply adding definitions.

The first identified principle is that of Technology Neutrality, drawn from the EU AI Act. The EU AI Act was introduced against a backdrop of documented instances (pg.10) of algorithmic bias in consumer-facing financial services. The Act classifies AI platforms used for financial advice and credit scoring as high-risk under Annex III, requiring mandatory pre-deployment conformity assessments, ongoing human oversight, and registration in a public database. Article 3(1) defines an AI system functionally, by reference to its capacity to infer outputs that influence decisions, rather than by reference to any specific technology. It establishes a pre-market governance checkpoint based on a risk-prevention model. Definitions anchored to a specific technology become obsolete when the technology evolves. India’s existing provisions demonstrate this failure. Any definition SEBI introduces must be function-based, specifically whether it generates or materially shapes an investment recommendation or an execution outcome, rather than how it does so.

For SEBI, the definition should focus on function. Any automated system that, for consideration, generates or materially determines an investment recommendation or execution decision, and delivers that output directly to clients without a human actually framing the advice in real time, should fall within the IA Regulations. The key factor should be whether any human genuinely exercised independent judgment over the substance of the advice before it reached the client.

This still leaves room for ordinary advisory practice. If a human adviser merely uses analytical or data tools to assist their work, independently evaluates the output, and gives advice based on their own professional judgment, the system remains a tool. But where the platform itself generates the recommendation and presents it as the basis for investment action, it is effectively acting as the adviser.

The existing Indian framework does not address this problem because it was never designed to. SEBI’s 2012 algorithmic trading circular regulates trading infrastructure and order flow, while the PFUTP Regulations target misrepresentation and market abuse. Neither framework classifies AI-driven advisory systems as investment advisers.

A technology-neutral definition remains applicable across current and future AI architectures without requiring constant reforms. However, functional definitions alone are insufficient if they can be bypassed through nominal human involvement. This necessitates a Materiality standard.

The second identified principle is the Materiality Standard. The Materiality Standard defines the scope of regulation by the substantive role an algorithmic system plays in shaping an outcome, applying wherever such a system materially determines the content of a recommendation or execution decision, regardless of subsequent human review. The decisive criterion is whether human judgment was displaced at the stage where the financial outcome was formed; a human actor’s presence elsewhere in the process does not alter this classification.

A definition confined to fully automated systems has an obvious loophole, since it allows actors to insert a nominal human reviewer who provides no meaningful input, letting an output be formally characterised as human advice while remaining functionally automated. This is the human-in-the-loop problem. The European Securities and Markets Authority (ESMA), in its 2018 Guidelines on MiFID II Suitability Requirements, confirmed that algorithm-generated advice constitutes investment advice regardless of whether it is reviewed by a human before delivery. Singapore reinforced this by anchoring classification to the absence of “contemporaneous human intervention” in generating the recommendation, rather than in its delivery, a position echoed in its FEAT Principles and Guidelines on the Provision of Digital Advisory Services, which define digital advisory services in identical terms and require platforms to continuously monitor model performance, fairness metrics, and decision audit trails throughout the model lifecycle. Even Singapore, however, has acknowledged that voluntary principles alone are insufficient, and its move from FEAT Principles to binding AI risk management guidelines confirms that the regulatory approach is evolving toward harder obligations as the technology matures.

Third, any regulatory attempt shall include a counter to the Black Box Problem. An AI system does not always provide a rationale for the executed decisions. They are often generated through probabilistic inference across a training dataset, and the individual responsible for deploying the system may have no more visibility into why a specific output was produced than the investor herself. Singapore’s FEAT Principles addressed this by requiring financial institutions to explain algorithmic decisions to affected clients in meaningful terms, treating explainability not as a best practice but as a precondition for deployment. India’s existing framework makes some attempt in this direction. SEBI’s 2012 algorithmic trading circular requires brokers to maintain logs of all trading activity and tag each algorithmic order with a unique identifier to establish an audit trail. The 2025 circular on retail algorithmic trading goes further, requiring brokers to preserve detailed API activity logs for five years (pg. 4). The June 2025 Consultation Paper additionally proposes model explainability requirements, process documentation, and independent audits of AI systems as part of a board-approved governance framework.

These measures, however, address a different problem. An audit trail records what a system did, the sequence of orders placed, modified, or cancelled. It does not and cannot explain why the system reached the decision it did, which is the question that explainability actually demands. A log that confirms an AI platform sold a particular security at a particular time tells a regulator nothing about whether that decision was appropriate for the client, what inputs drove it, or whether the model had drifted from its intended behaviour. The gap between recording outputs and explaining reasoning is precisely where India’s framework currently falls short.

The MAS has since acknowledged that newer and more complex AI technologies make explainability increasingly difficult to achieve. This is why the displacement of contemporaneous human judgment must operate as a definitional trigger. Where a human actor is unable to explain what the system decided and the reasons underlying that decision, it indicates that no meaningful review has occurred. In such circumstances, the system has effectively made the decision, irrespective of any formal human involvement.

Finally, any definition remains incomplete if confined to the moment of advice, given the continuous operation of AI platforms. This introduces the Temporal Dimension. Unlike traditional advice, an AI system can execute hundreds of portfolio decisions over months or years once authorised, regardless of unanticipated market conditions. An adequate definition should therefore account for this temporal dimension by covering not only the moment of initial recommendation but also subsequent automated execution decisions that flow from the model’s ongoing operation. The US SEC identified this as one of challenges in its 2017 Robo-adviser guidance. A definition that captures only the initial recommendation, but not the ongoing execution decisions of the same system, would leave the most consequential outputs of an AI platform entirely unregulated.

Conclusion

Despite doctrinal differences, all three approaches bring AI platforms within regulatory scope through functional classification, accountability triggers, or fiduciary extension.

Drawing on the characteristics identified above, an Algorithmic Investment Service Provider may be defined as ‘any automated system that, for consideration, generates or materially determines an investment recommendation or execution decision delivered directly to a client.’

A system should not fall outside this definition merely because a human reviews its output before delivery. Where that human has not independently evaluated the substance of the recommendation (where the automated output reaches the client in essentially the form the system generated it) the system remains the effective adviser for regulatory purposes. Equally, where the deploying entity cannot recover or explain the logic underlying the output at the point of delivery, no meaningful human review can be said to have occurred, and the system is deemed to have made the decision irrespective of any formal human involvement.

Classification alone is insufficient. The 111 broker settlement illustrates that intermediaries can derive commercial benefit from AI platforms and disclaim liability through vendor terms. The December 2024 SEBI Board Paper recognises that deploying entities must remain the primary accountable party for AI outputs, a position the Supreme Court affirmed in SEBI v Rakhi Trading Pvt Ltd (2018) 13 SCC 753, holding that regulatory obligations cannot be displaced by private contractual arrangements. SEBI’s existing powers are sufficient. The only question that remains is of regulatory capacity.

The authors are Year IV students at the NALSAR University of Law.