What AI governance obligations apply to custodian banks using AI for SEBI FPI circular trading surveillance?

SEBI (FPI) Regulations 2019 (SEBI/LAD-NRO/GN/2019/45) and SEBI (Prohibition of Fraudulent and Unfair Trade Practices) Regulations 2003 (FUTP Regulations) impose obligations on Designated Depository Participants (DDPs) and custodian banks to monitor FPI trading for circular trading patterns. AI models used by custodians to flag circular trading — coordinated buy-sell patterns between FPI entities with common beneficial ownership, wash trade detection across offshore structures — must meet four governance standards: (1) MRMF validation: custodian banks are RBI-regulated entities; their AI surveillance models are material risk models under MRMF 2026 requiring pre-deployment validation, model card documentation, and ongoing PSI monitoring; (2) FUTP evidentiary standard: SEBI FUTP enforcement uses custodian-identified circular trading as evidence in adjudication proceedings. An AI model flag that is used as the basis for a SEBI investigation must be reproducible — the same inputs, at the same point in time, must produce the same alert. Non-deterministic AI models (those using random sampling or non-reproducible inference) are not suitable for FUTP enforcement purposes; (3) SEBI Integrated Surveillance System (ISS) integration documentation: custodians feeding AI-generated alerts to SEBI's ISS must document the AI model's methodology, false-positive rate, and alert threshold in the ISS interface documentation; (4) Judicially defensible alert log: SEBI adjudication orders since 2022 have cited AI/ML surveillance findings as the basis for FPI enforcement actions. The custodian institution, as the source of the AI-generated surveillance finding, must maintain the full alert log — model version, input trading data used, output score, alert rationale — for potential production in SEBI adjudication proceedings.

How does SEBI's look-through beneficial ownership requirement create AI explainability obligations for DDPs?

SEBI circulars (SEBI/HO/AFD/AFC1/P/CIR/2023/0026, February 10, 2023, and 2024 amendments) impose look-through obligations on FPIs where concentration or structure triggers review: more than 50% of FPI assets from a single beneficial owner (the opaque-structure rule), or FPI structures with multiple layered entities. DDPs use AI models to identify ultimate beneficial ownership across multi-layered offshore structures — mapping Cayman → Mauritius → Singapore → India FPI entity chains to identify the ultimate human controller. These beneficial ownership AI models must be explainable at two levels: (1) SEBI inspection explainability: SEBI has requested that DDPs demonstrate how beneficial ownership AI flags were generated during examination. An AI model that produces a 'high-risk BO flag' based on opaque neural network layers, without a traceable path from the input corporate structure data to the flag, is not defensible in a SEBI examination. DDPs must document the model's decision criteria — what ownership percentages, structural features, or jurisdiction combinations trigger flags — at a level of specificity that allows SEBI examiners to assess whether the flagging criteria are appropriate; (2) Show-cause proceeding explainability: when SEBI issues a show-cause notice to an FPI based on custodian-identified BO concerns, the FPI or its DDP may be required to explain the AI finding. The DDP must be able to produce a human-readable explanation of why the AI flagged the specific BO structure, grounded in the regulatory criteria of SEBI FPI Regulations 2019 Regulation 22 (opaque structures) and SEBI's beneficial ownership circulars.

Why does FPI concentration limit monitoring AI face a unique dual-regulated governance challenge in India?

SEBI FPI Regulations 2019 Regulation 20 imposes a 10% single-investor and 10% single-stock FPI aggregate holding limit. AI models used to monitor and enforce concentration limits in real time must comply with two separate regulatory frameworks simultaneously — a governance requirement unique to India's FEMA-SEBI dual jurisdiction structure: (1) RBI MRMF (for the custodian bank entity): the bank entity's concentration limit AI is a quantitative model used in material risk calculations and must be in the MRMF model inventory, independently validated, and board-reported under MRMF 2026; (2) SEBI system audit (for the DDP entity): the DDP entity's concentration monitoring system is subject to SEBI's system audit framework (SEBI Circular on System Audit for Market Infrastructure Institutions and Intermediaries), which requires third-party system audit of critical trading and settlement systems. A single AI model monitoring FPI concentration limits may require MRMF model validation documentation for the bank regulator and a SEBI system audit finding for the securities regulator — two different validation methodologies, two different audit reports, for the same model. No global AI governance framework — not NIST AI RMF, not EU AI Act, not US Federal Reserve AI guidance — addresses a dual-regulator model governance structure of this kind. India's FEMA-SEBI joint operational guidelines create this structure, and the AI governance documentation requirements follow from it.

What is the SEBI enforcement evidence standard for AI-generated FPI surveillance findings?

SEBI adjudication orders from 2022 through 2024 have increasingly cited AI/ML-based surveillance findings as the basis for FPI enforcement actions — penalty orders, debarment orders, and disgorgement orders. SEBI's Rules of Procedure for adjudication require that evidence be reproducible and the methodology disclosed. This creates an evidentiary standard for AI-generated FPI surveillance findings that custodians, stock exchanges, and SEBI's own surveillance division must maintain. Four documentation requirements follow from this standard: (1) Model version control: the specific model version and parameter set used to generate a surveillance alert must be preserved — a model that was subsequently retrained or updated cannot retroactively reproduce the original alert without version control; (2) Input data snapshot: the trading data, beneficial ownership records, and market microstructure data that were inputs to the alert must be preserved at the time of alert generation — not reconstructed from current records, which may have been updated; (3) Inference reproducibility: given the preserved model version and input data, the institution must be able to reproduce the original alert score and trigger logic. Non-deterministic model elements (dropout layers active during inference, random sampling in data preprocessing) that prevent reproducibility are incompatible with the SEBI evidentiary standard; (4) Methodology disclosure: if SEBI's adjudicating officer questions the AI methodology, the institution must be able to disclose the model's alert criteria at a sufficient level of technical detail for the officer to assess whether the methodology is appropriate for the alleged market manipulation — not simply assert that 'the AI flagged it'.

AI Governance · SEBI FPI · Surveillance AI · FUTP · MRMF · India BFSI

SEBI FPI AI Governance: Foreign Portfolio Investor Circular Trading Surveillance, Beneficial Ownership AI, Concentration Limit Monitoring, and FUTP Enforcement Evidence Standards India 2026

SEBI FPI Regulations 2019 and FUTP Regulations impose AI governance obligations on custodian banks and Designated Depository Participants that no global AI governance firm covers — circular trading surveillance AI must be FUTP-evidentiary-standard compliant, beneficial ownership look-through AI must be SEBI-inspection-explainable, concentration limit AI faces dual RBI-MRMF and SEBI-system-audit governance, and SEBI adjudication proceedings since 2022 set a reproducibility standard for AI-generated enforcement findings. Zero global competitor coverage.

Zytra Research Team · October 2026 · 13 min read

📖 Part of the India BFSI AI Governance Complete Guide 2026 — RBI, SEBI, IRDAI, NPCI & DPDP in one place.

Foreign Portfolio Investors (FPIs) constitute India's largest pool of foreign equity investment — over USD 700 billion in Indian securities. SEBI's FPI regulatory framework — SEBI (Foreign Portfolio Investors) Regulations, 2019 (SEBI/LAD-NRO/GN/2019/45, September 23, 2019), as amended through 2024 — governs how FPIs register, invest, report, and how they are monitored for compliance.

Three parties in the FPI ecosystem deploy AI for regulatory compliance: Designated Depository Participants (DDPs), who onboard and monitor FPIs; custodian banks, who hold FPI assets and monitor trading patterns; and SEBI's Integrated Surveillance System (ISS), which aggregates AI surveillance outputs from across the market. Each faces AI governance obligations that are distinct from the domestic algo trading AI governance obligations that have received most regulatory attention — and that no global AI governance vendor has addressed.

FPI Circular Trading Surveillance: The FUTP Evidentiary Standard

SEBI FUTP Regulations 2003, Regulation 4, prohibit circular trading — coordinated buy-sell transactions between entities with common beneficial ownership that create artificial trading volumes or prices without genuine change in ownership. Custodian banks monitor FPI trading for circular trading patterns and report to SEBI's ISS.

AI models used for circular trading detection identify coordination patterns across FPI entities: time-correlation in trade execution, price-quantity symmetry between buy and sell legs, beneficial ownership linkages between counterparties. These models generate alerts that feed SEBI enforcement workstreams.

The governance obligation follows from the legal consequence: SEBI adjudication orders cite custodian-identified AI surveillance findings. An AI alert that becomes the basis for a SEBI enforcement action — penalty, debarment, disgorgement — must meet the FUTP evidentiary standard: the finding must be reproducible, the methodology must be disclosable, and the input data must be preserved.

A custodian whose circular trading AI was retrained after the alert was generated, without preserving the original model version, cannot reproduce the original finding. SEBI's adjudicating officer can treat the inability to reproduce as undermining the evidentiary weight of the finding. The custodian institution — not SEBI — bears the governance burden of maintaining the evidence chain for its own AI-generated findings.

Beneficial Ownership Look-Through AI: SEBI Inspection Explainability

SEBI's opaque-structure rule (reinforced through 2023 and 2024 circulars) requires DDPs to perform look-through analysis for FPIs with concentrated or opaque ownership structures. AI models that map multi-layered offshore ownership chains to identify ultimate beneficial owners are among the most technically complex models in the BFSI AI landscape — they must process corporate registry data from multiple jurisdictions, resolve entity name variations across languages and transliteration conventions, and identify ownership thresholds across multiple layers of intermediary entities.

SEBI examinations of DDPs have begun including requests to demonstrate how beneficial ownership AI flags were generated. The examination is not asking whether the AI works — it is asking whether the DDP can explain what the AI found, why it found it, and whether the finding is consistent with the regulatory criteria in SEBI FPI Regulations 2019 Regulation 22 and the beneficial ownership circulars.

Explainability requirements for beneficial ownership AI:

Regulatory criteria mapping: the AI's flagging criteria must map to SEBI's regulatory triggers — 50% single-BO concentration, specified jurisdiction structures, opaque custodial chains. The model's decision criteria must be documentable in regulatory terms, not only in model architecture terms
Structure visualization: the AI must produce an ownership structure diagram that a SEBI examiner can read — showing the entity chain, ownership percentages, and the specific criteria that triggered the flag
Update frequency: beneficial ownership data changes as FPIs restructure. The AI model must document when the ownership data used for each flag was last refreshed — a flag based on 18-month-old ownership data that has since changed may not be defensible if the FPI can demonstrate it restructured before the SEBI examination

Dual-Regulated AI: MRMF + SEBI System Audit for Concentration Limit Models

FPI concentration limit monitoring AI — models that track the 10% single-investor and 10% single-stock holding limits in real time — sits at the intersection of RBI and SEBI regulation for custodian entities that are both banks (RBI-regulated) and DDPs (SEBI-regulated).

The MRMF obligation covers the bank entity: concentration limit AI is a quantitative model used in material risk calculations (a limit breach creates regulatory liability), requiring MRMF model card, independent validation, and board reporting. The SEBI system audit obligation covers the DDP entity: SEBI's system audit framework requires periodic third-party audit of critical trading and settlement systems, which includes concentration limit monitoring systems.

Two different validation methodologies, two different audit reports, two different regulatory examiners, for the same AI model. The practical governance solution: maintain a single model with dual documentation packages — MRMF model card and validation report for RBI examination, system audit documentation for SEBI examination — with cross-references to avoid inconsistency between the two packages. An inconsistency between what the bank tells RBI the model does and what the DDP tells SEBI the model does is a governance failure that creates exposure with both regulators simultaneously.

FPI Debt Route AI: The Third Regulatory Layer

FPIs investing in Indian debt under the VRR (Voluntary Retention Route) and Fully Accessible Route are subject to joint RBI-SEBI oversight under FEMA Master Direction on Debt Investment by FPIs (2024). AI models monitoring FPI debt investment headroom, generating regulatory alerts for limit breach risk, and tracking utilization against the aggregate FPI debt limit face a third regulatory layer: FEMA compliance obligations.

The practical governance implication: FPI debt limit AI sits under three regulatory frameworks simultaneously — SEBI FPI Regulations for the DDP, RBI MRMF for the bank, and FEMA Master Direction for the debt investment monitoring function. No single regulatory framework's AI governance guidance covers all three. India's institutions deploying AI in this space are navigating an intersection that global AI governance vendors structurally cannot advise on.

SEBI FPI AI Governance Checklist

☐ Circular trading surveillance AI: MRMF model card; FUTP evidentiary standard compliance — deterministic inference, version control, input data snapshot preservation; alert log retention for adjudication window
☐ ISS integration documentation: AI model methodology, false-positive rate, and alert threshold documented in SEBI ISS interface documentation; SEBI examination ready
☐ Beneficial ownership look-through AI: flagging criteria mapped to SEBI regulatory triggers (Reg. 22, BO circulars); structure visualization output; ownership data refresh frequency documented; SEBI inspection explainability package
☐ Dual-regulated concentration AI: MRMF model card (RBI) and system audit documentation (SEBI) maintained as separate packages with cross-references; no inconsistency between packages
☐ FPI debt route AI: FEMA compliance documentation for debt limit monitoring AI; three-framework governance package (SEBI + RBI MRMF + FEMA) maintained
☐ SEBI adjudication evidence readiness: for each AI-generated FPI enforcement referral, preserve model version, input data snapshot, inference reproducibility test, and methodology disclosure package
☐ PSI monitoring: FPI circular trading and BO AI monitored for distribution drift — FPI registration data and trading pattern distribution changes quarterly
☐ Model update change management: any model retrain triggers MRMF change management log and SEBI system audit notification for material model changes
☐ DDP examination readiness: SEBI examination package for FPI AI governance — BO AI explainability, concentration limit audit documentation, surveillance methodology disclosure
☐ Bank examination readiness: RBI MRMF examination package — model cards, validation reports, PSI monitoring dashboards for all FPI surveillance AI models