RBI DAKSH: What Banks Must Do to Prepare for AI-Powered Supervisory Monitoring

The Reserve Bank of India's supervisory transformation is underway, not forthcoming. DAKSH — the Advanced Supervisory Monitoring System, launched October 2022 — processes compliance workflow data, cyber incident reports, and MIS data from supervised entities in near real-time. The Supervisory Data Quality Index, introduced in 2025, scores the quality of every bank's data submission and creates a direct link between a bank's internal data governance posture and its supervisory risk profile. Banks whose data submissions score poorly under SDQI face escalated supervisory attention. Governor Sanjay Malhotra's January 2026 speech made the direction explicit: near real-time supervision of India's digital financial system is the goal, and traditional periodic on-site inspection is no longer the primary supervisory mode.

This has three direct implications for AI governance: the model inventory Zytra's governance tools maintain must be DAKSH-ready; AI models that feed bank MIS data submissions must produce explainable outputs when DAKSH flags anomalies; and the data quality obligations embedded in DAKSH's SDQI framework create enforceable data lineage requirements for AI pipelines.

What DAKSH Collects from Banks

Four Categories of Supervisory Data

The Supervisory Data Quality Index

How SDQI Scores Your Bank

SDQI, introduced in 2025, scores bank data submissions on five dimensions:

1. Completeness — are all required data fields populated? A model inventory with 85% completion scores lower than one at 100%. Empty fields for validation status, model owner, or last review date are completeness gaps.
2. Accuracy — do submitted values match internal records and external verification? AI models that process data before submission introduce accuracy risk if they apply transformations not documented in submission methodology.
3. Consistency — do values across different returns agree? An AI model generating cash flow projections for one regulatory return must produce values consistent with the underlying data in other returns. Inconsistency flags suggest either model errors or data pipeline fragmentation.
4. Timeliness — are submissions made by deadline? An AI-driven data pipeline that processes large volumes of transaction data is subject to the same submission deadlines as manual processes. Pipeline failures that cause late submissions degrade SDQI.
5. Governance — is there a documented data governance framework? The SDQI framework rewards banks that maintain documented data lineage, data dictionaries, and change management processes for data pipelines — including AI-driven ones.

What Poor SDQI Scores Trigger

A bank with poor SDQI scores enters an escalated supervisory mode that involves more frequent data requests, additional targeted inspections, and in severe cases, supervisory meetings with senior management or board. From an AI governance standpoint: a bank whose AI models are contributing to SDQI degradation (through incomplete model inventory, inconsistent MIS outputs, or undocumented data transformations) creates a direct linkage between AI governance quality and supervisory risk profile.

Explainability for Supervisory Data Requests

When DAKSH's anomaly detection flags a pattern in a bank's submitted data, RBI issues a supervisory data request — asking the bank to explain the underlying driver. If that driver involves an AI model, the bank must produce:

What Banks Need to Produce on Supervisory Request

1. Model card: What does the model do, what data does it process, what is its output, and who is the accountable owner? A model card must be producible within the supervisory response timeline (typically 7–14 working days for initial response).
2. Validation report: When was the model last validated, against what data, by whom, and what were the findings? Validation reports must be current — a model with a 3-year-old validation report for a supervisory data request is a documentation failure, not just a governance gap.
3. Assumption log: What assumptions does the model make, and have those assumptions been tested against current market conditions? For models used in regulatory data processing (capital calculations, provision models, liquidity metrics), assumption drift is a specific DAKSH risk — the model may be operating on stale assumptions that produce systematically biased outputs.
4. Change log: Has the model been changed since last validation? Model changes that were not re-validated create a gap in the supervisory response chain. DAKSH can in principle detect when submitted data patterns shift in ways consistent with model changes.

Data Pipeline Governance for DAKSH Feeds

Four Data Governance Requirements AI Pipelines Must Meet

1. Data lineage documentation: Every AI pipeline that processes data eventually submitted to RBI must have documented lineage — from source data to model input to model output to regulatory return. Undocumented transformations in the pipeline create SDQI accuracy risks.
2. CIMS integration readiness: RBI's Centralised Information Management System (CIMS) is the data infrastructure DAKSH draws from. Banks submitting data through CIMS must ensure their AI pipelines produce outputs in CIMS-compatible formats and that model-generated values are clearly attributed (not merged with manually adjusted values without documentation).
3. Change management for AI pipelines: Model updates, data source changes, and feature engineering changes must trigger a documented review before the changed pipeline feeds DAKSH-visible submissions. A model that changes its output distribution without a documented change management process creates an unexplained shift in supervisory data.
4. Audit trail for AI-generated regulatory data: Any value in a regulatory return generated or materially processed by an AI model must have an audit trail sufficient to reconstruct the computation. This is both an RBI IT Governance Master Direction requirement and a DAKSH data quality requirement.

Utkarsh 3.0 Forward-Look

What RBI's 2025–2027 Strategy Means for AI Governance

Utkarsh 3.0, RBI's medium-term strategy for the period 2025–2027, embeds near-real-time supervisory capability as a strategic objective. Four specific implications:

1. Supervision will move from sampling to comprehensive monitoring — meaning every bank, every return, every anomaly pattern is visible to DAKSH, not just those selected for inspection.
2. Model inventory data will be requested more frequently and with shorter lead times — banks need a continuously maintained model registry, not a point-in-time snapshot produced for inspections.
3. AI-specific supervisory data requests will increase — as MRMF implementation matures, RBI will be monitoring model risk at the portfolio level, not just validating individual models.
4. The explainability bar will rise — a model explanation that satisfied a 2023 inspection may not satisfy a 2026 supervisory data request, as RBI's own supervisory AI capabilities improve and the bar for model documentation rises correspondingly.

Frequently Asked Questions