What Significant Data Fiduciary obligations under DPDP Act 2023 apply to AI model deployment in BFSI?

DPDP Act 2023 Section 16 empowers the Central Government to designate specific Data Fiduciaries as Significant Data Fiduciaries (SDFs) based on volume and sensitivity of data processed, risk to data principals, national security and public order considerations, and the potential impact on sovereignty and integrity. All major Indian banks (SBI, HDFC Bank, ICICI Bank, Axis Bank), large insurance companies (LIC, SBI Life, HDFC Life), and major NBFCs (Bajaj Finance) are expected to receive SDF designation when the Government issues notifications under Section 16(1) — expected during 2026. Draft DPDP Rules 2025 (published January 3, 2025 for consultation) Draft Rule 12 requires SDFs to conduct Data Protection Impact Assessments (DPIAs) before deploying new significant data processing systems. In BFSI, virtually all AI/ML deployments qualify: credit underwriting models process millions of customer financial records; fraud detection models process real-time transaction data on hundreds of millions of accounts; customer segmentation AI processes complete behavioral profiles. The DPIA obligation creates a pre-deployment AI governance requirement that is more specific than any RBI or SEBI circular currently mandates: document the nature and purpose of processing, assess necessity and proportionality, identify risks to data principals, and establish measures to address those risks — all before the AI model goes live.

How does DPDP Act purpose limitation conflict with ML model feature discovery in India BFSI?

DPDP Act Section 8(3) prohibits processing personal data for any purpose other than the specified purpose for which consent was obtained or for which a legitimate use exists under Section 7. ML models trained on BFSI customer data routinely discover emergent correlations — statistical relationships that were not anticipated when the processing purpose was defined. A credit scoring model may learn that transaction velocity at grocery stores in specific PIN codes is predictive of default; a fraud detection model may learn that calls to specific IVR numbers are predictive of SIM swap fraud. These emergent feature discoveries are not planned — they emerge from the model's optimization. The DPDP Act purpose limitation question: if a customer consented to processing their transaction data 'for the purpose of credit assessment,' has the bank violated purpose limitation by training a model that uses transaction data features to detect fraud (a different purpose)? Or does the model's use of cross-purpose correlations constitute a secondary use requiring separate consent? India's DPDP Act is silent on ML model feature discovery. The EU GDPR has developed a 'compatible purpose' doctrine through guidance and case law — DPDP Act enforcement by the Data Protection Board of India will likely need to develop an equivalent doctrine, but until it does, BFSI institutions bear the risk of purpose limitation violations in their AI models. Practical governance implication: ML model development for BFSI AI should include a purpose compatibility assessment — documenting the intended processing purpose, the training data used, and a review of whether the model's learned features are within the purpose scope.

Can a BFSI customer's right to erasure under DPDP Act require ML model retraining?

DPDP Act Section 12(1)(b) grants Data Principals the right to erasure of personal data when it is no longer necessary for the purpose for which it was processed. This creates a direct collision with ML model governance: once a customer's data has been used to train a model, the data is 'in the model' — encoded in the model's weights, the decision boundaries it has learned, the correlations it has discovered. Deleting the raw training data does not remove the customer's influence from the model. The DPDP Act does not explicitly address whether a trained model's parameters constitute personal data processing. However, three scenarios make the tension concrete: (1) A customer whose loan was rejected by an ML model requests erasure of their credit data. The bank deletes the raw data but the model that rejected them (and may have been re-trained on their data) continues operating. Is this erasure compliant? (2) A customer who was subjected to fraud detection AI that generated a false positive — resulting in account freezing — requests erasure. The fraud model's training data is deleted but the customer's experience as a false positive is embedded in the model's calibration. (3) A BFSI institution processes sensitive health data of insurance applicants (before IRDAI's sensitive data restrictions) to train an underwriting model. The customer later exercises erasure. Deleting the raw health data does not unlearn the model's learned correlation between health indicators and mortality. DPDP Act guidance from MeitY or the Data Protection Board of India on model unlearning obligations is not yet published. EU's GDPR guidance on the 'right to be forgotten' in ML contexts — machine unlearning research, differential privacy, model retraining — provides a reference framework, but India's DPDP has its own enforcement trajectory. BFSI institutions should begin assessing which AI models may face unlearning requests and whether current model architecture allows for compliant erasure responses.

How do DPDP Act cross-border data transfer restrictions affect BFSI AI vendor selection?

DPDP Act Section 16(2)(d) empowers the Central Government to restrict cross-border data transfers for SDFs to specific countries or categories. Draft DPDP Rules 2025 Rule 12(4) indicates that SDFs may face additional cross-border transfer restrictions beyond those applying to all Data Fiduciaries. For BFSI institutions designated as SDFs, this creates a supply chain governance question about AI vendors: a US-based AI governance platform (model monitoring dashboard, explainability tool, bias detection service) that processes Indian customer financial data offshore may fall within the cross-border transfer restrictions when they apply to SDFs. This is compounded by RBI's existing data localization circular (April 2018): payment system data must be stored only in India. A BFSI institution using a US-hosted AI platform for real-time transaction fraud monitoring may already be violating RBI's payment data localization requirement — DPDP Act SDF restrictions add a second, overlapping compliance layer. The practical governance implication: BFSI institutions evaluating AI vendors should assess: (a) where the vendor processes Indian customer data (model training, inference, monitoring); (b) whether processing can be conducted in India or in a government-approved jurisdiction; (c) whether the vendor can contractually commit to DPDP Act data transfer compliance; (d) whether the vendor's AI platform architecture supports India-resident data processing without architectural changes. Global AI governance platforms that do not address Indian data residency requirements have a structural compliance risk for BFSI SDF customers.

AI Governance · DPDP Act · Significant Data Fiduciary · India BFSI · Data Protection

DPDP Act 2023 Significant Data Fiduciary AI Governance for India BFSI: DPIA Obligations, Purpose Limitation vs ML Feature Discovery, Right to Erasure vs Model Unlearning, and Cross-Border AI Vendor Risk 2026

India's Digital Personal Data Protection Act 2023 creates AI governance obligations for BFSI that no RBI, SEBI, or IRDAI circular addresses: mandatory Data Protection Impact Assessments before ML model deployment for Significant Data Fiduciaries (expected: all major banks, insurers, NBFCs), purpose limitation constraints on emergent ML feature discovery, a right to erasure that may require ML model unlearning, cross-border AI vendor restrictions for SDF institutions, and legacy consent adequacy questions for ML training data. India-specific, BFSI-specific, zero global competitor coverage.

Zytra Research Team · October 2026 · 12 min read

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

The Digital Personal Data Protection Act 2023 (Act No. 22 of 2023, notified August 11, 2023) is India's first comprehensive data protection legislation. Its most consequential provisions for BFSI AI governance are not the general obligations that apply to all Data Fiduciaries — which map reasonably well onto existing RBI and SEBI data governance requirements — but the Significant Data Fiduciary provisions of Section 16, which create a new tier of obligations that explicitly intersect with AI model deployment.

India's large BFSI institutions are expected to receive SDF designation. All major public and private sector banks process personal data of hundreds of millions of customers; LIC alone has 250+ million policyholders; Bajaj Finance has 85+ million customer records. The combination of data volume, sensitivity (financial and health data), and systemic importance makes BFSI sector the primary target for SDF designation when Government notifications begin under Section 16(1).

The DPDP Act creates five AI governance obligations that BFSI institutions must address — obligations that arise from data protection law, not from prudential regulation, and that exist independently of whether RBI or SEBI has issued guidance on the same topics.

Obligation 1: DPIA Before ML Model Deployment

Draft DPDP Rules 2025 Rule 12 requires SDFs to conduct Data Protection Impact Assessments before deploying significant new data processing systems. In BFSI, the credit underwriting AI that processes millions of customer records, the fraud detection system processing real-time transaction data, and the customer lifetime value model processing complete behavioral profiles are all significant new data processing systems when deployed.

A DPIA for BFSI AI includes: the nature and categories of personal data processed; the purposes and legal basis for processing; whether processing is necessary and proportionate to the purpose; risks to data principals (incorrect credit decisions, false fraud flags, discriminatory outcomes); and measures to mitigate those risks. This maps directly onto AI model governance practices — model cards, bias assessments, risk documentation — but driven by data protection law rather than prudential regulation.

Obligation 2: Purpose Limitation and Emergent ML Features

DPDP Act Section 8(3) prohibits secondary use of personal data. In ML model development, purpose limitation creates a novel compliance challenge: when a fraud detection model, trained on transaction data collected for fraud prevention, discovers that certain behavioral signals (IVR call patterns, device location history, transaction timing rhythms) are predictive — should those signals be used, or does their use constitute secondary processing beyond the consent scope?

Practical governance protocol for purpose limitation in ML:

Before model training: document the processing purpose and the categories of personal data approved for that purpose
During feature engineering: assess whether proposed features are within the purpose scope or constitute secondary use
Post-training: assess whether the model's learned features include unexpected cross-purpose correlations; document whether those correlations are within or outside purpose scope
For out-of-scope features: either retrain excluding out-of-scope data features, obtain separate consent for the extended purpose, or rely on a Section 7 legitimate use basis if applicable

Obligation 3: Right to Erasure and ML Model Unlearning

Section 12(1)(b) right to erasure is India's equivalent of GDPR's right to be forgotten. For BFSI AI, the question of whether erasure requires ML model retraining or unlearning is unresolved. Machine unlearning — the technical capability to remove a specific data point's influence from a trained model without full retraining — is an active area of ML research but is not yet standard practice in BFSI AI deployments.

Governance steps BFSI institutions should take before DPDP Act enforcement begins:

Assess which AI models were trained on personal data that could be subject to erasure requests
For each such model: can it be retrained excluding an individual's data? What is the retraining cost and timeline?
For models where retraining is impractical: assess whether the model's continued operation constitutes ongoing personal data processing (if not, erasure of raw data may satisfy the obligation)
Develop a legal position on the model unlearning question — informed by MeitY guidance when published, and by DPDP Board of India decisions as enforcement begins
For new model deployments: consider architecture choices (differential privacy, federated learning) that reduce the personal data footprint in trained models

Obligation 4: Cross-Border AI Vendor Compliance

Draft DPDP Rules Rule 12(4) SDF cross-border restrictions, layered on RBI's payment data localization circular, create a vendor selection governance requirement: BFSI institutions evaluating AI governance platforms, model monitoring tools, bias detection services, and explainability dashboards must assess data residency. India-resident data processing is the safest default for BFSI AI vendors.

Obligation 5: Legacy Consent Adequacy for ML Training Data

DPDP Act deemed consent provisions (Section 7) provide safe harbor for "legitimate uses" including contractual necessity and employer obligations. ML training on BFSI customer data collected before the DPDP Act's notification (August 2023) under pre-existing consent frameworks faces a retroactive adequacy question: does the pre-DPDP consent for "processing your data for banking services" cover ML model training for fraud detection, credit scoring, and customer segmentation?

No definitive guidance from MeitY or the DPDP Board. Prudent governance: legal opinion on whether existing consent frameworks satisfy DPDP Act standards for each ML use case; consent refresh plan for high-risk uses where adequacy is questionable; new model deployments using DPDP-compliant consent from data collected after August 2023.

DPDP Act BFSI AI Governance Checklist

☐ SDF designation monitoring: track Government notifications under DPDP Act Section 16; assume SDF designation for major banks, insurers, NBFCs and prepare accordingly
☐ DPIA process: establish pre-deployment DPIA process for new AI/ML systems; DPIA template covers personal data categories, purpose, necessity, risk assessment, mitigation measures
☐ Purpose limitation review: AI/ML model development checklist includes purpose compatibility assessment; out-of-scope feature identification process; legal basis documentation for each data category used in model training
☐ Model unlearning assessment: inventory of AI models trained on personal data; retraining feasibility assessed; legal position on model unlearning obligation documented; machine unlearning options evaluated for new deployments
☐ AI vendor data residency: vendor assessment for AI platforms includes data residency documentation; India-resident processing preferred; cross-border transfer legal basis documented for approved vendors
☐ Legacy consent review: legal opinion on pre-DPDP consent adequacy for each ML training data category; consent refresh plan for high-risk uses; new data collected post-August 2023 under DPDP-compliant consent
☐ Data Protection Board readiness: documentation ready for DPDP Board investigations — processing purposes, consent records, DPIA reports, erasure request handling procedure, model governance documentation