India's robo-advisory market reached over 70 million registered investors in 2025. The largest mutual fund platforms — Groww, Zerodha, Paytm Money, ET Money — use AI recommendation engines that generate fund selection, portfolio rebalancing, and SIP suggestions for millions of customers. Banks' wealth management arms use AI for client segmentation, product recommendation, and portfolio health scoring. Newer entrants use large language models to deliver conversational investment advice.
Every entity in this space is operating under SEBI's Investment Advisers Regulations, 2013 (SEBI/LAD-NRO/GN/2013/14) — whether or not they have framed their AI product as "advisory." The Regulations are technology-neutral: the obligation is triggered by providing investment advice for consideration, regardless of medium. SEBI's 2016 Consultation Paper on Robo-Advisors confirmed the application. SEBI's 2020 amendment tightened the suitability and conflict-of-interest obligations in ways that have direct AI governance implications.
This post maps the full AI governance obligation from SEBI IA Regulations to the specific documentation, audit, and model governance requirements that robo-advisory AI systems must meet.
The Technology-Neutral Application: What Counts as Investment Advice
The IA Regulations define "investment advice" as advice relating to investing in, purchasing, selling, or otherwise dealing in securities or investment products — including financial planning. The definition has no channel restriction. A mobile app that uses an AI model to generate a "recommended portfolio" based on a customer's risk questionnaire responses is providing investment advice. An LLM chatbot that responds to "what should I do with my ELSS allocation" is providing investment advice.
SEBI's 2024–2025 enforcement wave against AI-driven finfluencers crystallized the boundary: an AI system that generates investment-specific recommendations directed at identifiable investor segments — including through social media recommendations, personalized notifications, or portfolio suggestion screens — is providing advice that requires IA registration or must be provided through a registered IA.
Platforms that have positioned their AI as "goal-setting" or "financial planning" tools rather than "advisory" tools should review whether the AI's actual outputs — fund selection recommendations, allocation percentages, rebalancing triggers — constitute investment advice under the IA Regulations' definition.
The AI Suitability Obligation: Per-Client Documentation at Scale
The IA Regulations require that every investment recommendation be suitable for the specific client. Suitability is assessed across four dimensions: risk tolerance (assessed via SEBI-standard risk profiling questionnaire), investment horizon, financial objectives, and existing portfolio. The IA must document the suitability assessment for each recommendation and retain it for five years.
For AI robo-advisory at scale — serving millions of clients — the per-client suitability documentation obligation creates a system design requirement: the AI recommendation engine must generate and store per-client, per-recommendation suitability rationale that is auditable by SEBI.
This rationale cannot be generic ("recommendation is within client's moderate risk profile"). It must be specific enough to demonstrate compliance: which features of the client's risk profile and portfolio drove the recommendation, why the recommended product is within the client's risk category, and how the recommendation accounts for the client's existing portfolio concentration.
Three AI model design requirements follow from this obligation:
Explainable recommendation logic: A black-box model that cannot generate post-hoc explanations for individual recommendations fails the suitability documentation requirement. The AI must produce, for each recommendation, an explanation that connects client features to recommendation output in terms auditable by SEBI.
Per-recommendation suitability flag: The model must include a suitability check — before the recommendation is surfaced to the client, the model or an overlaid rule set confirms the recommendation is within the client's assessed risk category. Recommendations outside the client's risk category should be blocked or flagged for human review.
Model change suitability review trigger: When the AI recommendation model is updated, a review should assess whether the change affects suitability outcomes for any client segment. A model update that changes the recommendation output for a client segment whose risk profile hasn't changed is a suitability event requiring documentation.
Conflict-of-Interest in Commission-Weighted Training Data
SEBI's 2020 IA Regulations amendment prohibited IAs from receiving trail commissions from product manufacturers — the intent was to eliminate commission-driven bias in recommendations. For AI recommendation engines, this prohibition applies to the model's current incentive structure. But it does not address a distinct governance problem: models trained on historical data from the commission era.
If an AI recommendation engine was trained on five years of historical advisory records from a period when the IA operated under a commission model, the training data reflects commission-biased recommendations. The model has learned to recommend products that historically attracted commissions — mutual fund direct plans were systematically underweighted relative to regular plans, for example, because regular plans generated trail commissions for distributors and IAs often recommended them before the commission prohibition.
A model trained on this data carries the commission-era bias in its weights even after the IA transitions to a fee-only model under the 2020 amendment. The conflict-of-interest is embedded in the model rather than the adviser's current incentive structure. Governing this requires:
- Training data audit: what fraction of training recommendations were for commission-generating products versus direct plans or commission-free products?
- Feature importance analysis: does the model's learned feature weights reflect product margin or distribution fee as a learned signal?
- Neutral product universe backtesting: compare the model's recommendations against a recommendation set generated with all products held equal on commission — systematic differences indicate commission-era bias
- Disclosure: SEBI IA Regulations require disclosure of all conflicts of interest; training data lineage that includes commission-era data is a disclosable conflict
SEBI Algorithm Inspection Authority
SEBI's 2016 Consultation Paper established that SEBI's examination authority under the IA Regulations extends to the algorithm. For AI models, this means a SEBI inspection of a robo-advisory IA can request:
- Model documentation sufficient to verify that recommendations are suitability-compliant
- Version history of model changes with change rationale and suitability impact assessment
- Backtesting documentation showing recommendation performance against suitability criteria
- Audit logs of individual client recommendations with suitability rationale
- Conflict-of-interest audit documentation
A robo-advisory IA that cannot produce this documentation during a SEBI examination faces enforcement consequences equivalent to failing to produce human adviser documentation — potential suspension of IA registration, client remediation orders, and monetary penalties.
Consumer Protection Act 2019: Mis-Selling AI Liability
SEBI enforcement is not the only liability channel for AI mis-selling in India. The Consumer Protection Act 2019 defines "unfair trade practice" broadly enough to capture AI-generated unsuitable investment recommendations. A recommendation generated by a robo-advisory AI that was unsuitable for the client's assessed risk profile, resulted in financial loss, and cannot be explained or defended on suitability grounds creates Consumer Protection Act civil liability for the IA.
The Consumer Protection Act liability is cumulative with, not alternative to, SEBI regulatory action. A client who receives an unsuitable AI recommendation, suffers loss, and discovers that the AI's suitability rationale was inadequate or unavailable has both a SEBI complaint pathway and a Consumer Protection Act civil remedy.
For AI governance, this creates a documentation retention obligation beyond SEBI's five-year requirement: Consumer Protection Act claims can be filed within two years of the cause of action, which for investment losses may not be discovered until years after the recommendation was made. Practically, suitability documentation should be retained for the investment's full holding period plus five years.
Robo-Advisory AI Governance Checklist
- ☐ IA registration confirmed: entity providing AI investment recommendations is registered with SEBI as IA, or recommendations are provided through a registered IA arrangement
- ☐ Risk profiling AI: questionnaire-to-risk-profile mapping is documented; model changes to risk categorization logic are change-managed and client re-profiling triggered where needed
- ☐ Per-client suitability rationale: AI recommendation engine generates and stores per-recommendation suitability rationale; 5-year retention minimum
- ☐ Suitability check at recommendation: model or rule-set confirms recommendation is within client's risk category before surfacing; out-of-profile recommendations blocked or flagged
- ☐ Explainability mechanism: recommendation explanations are auditable, specific, and produced at the client level — not generic
- ☐ Commission-era training data audit: if model trained on historical commission-era recommendations, conflict-of-interest audit completed and disclosed
- ☐ Feature importance analysis: product margin/commission not a learned feature driving recommendations; documented
- ☐ SEBI inspection readiness: model documentation, version history, backtesting, conflict audit, recommendation logs — packaged and accessible for examination request
- ☐ Model change management: model changes assessed for suitability impact; client re-profiling triggered where model change affects recommendations for client segments
- ☐ Consumer Protection Act documentation: suitability rationale retained for holding period plus five years; process for responding to Consumer Protection Act complaints with AI recommendation documentation