India AI Impact Summit Insights
All sessions

AI-Driven Enforcement: Better Governance through Effective Compliance & Services

Contents

Executive Summary

India's Income Tax Department convened a major symposium bringing together government, industry, academia, and regulatory agencies to discuss AI-driven enforcement and compliance. The summit demonstrated that AI is transitioning from experimental initiatives to operational deployments across tax administration, banking regulation, policing, and securities enforcement. The overarching theme emphasized human-centered AI governance aligned with the MANO framework (Moral, Accountable, National, Objective)—ensuring AI enhances compliance, reduces disputes, and strengthens trust-based governance while maintaining ethical safeguards and accountability.

Key Takeaways

  1. AI is operationalized, not aspirational: Across tax, banking, policing, and securities regulation, India is deploying proven AI applications at scale. Nudges have changed taxpayer behavior measurably; Mule Hunter is saving ₹75–100 crores per bank by preventing fraud; investigation co-pilots are reducing case backlogs.

  2. Sovereignty and control matter: Build domain-specific, India-hosted models rather than outsourcing intelligence to foreign platforms. Sovereign SLMs, secure data warehouses, and centralized data integration (not siloed by department) are prerequisites for trustworthy AI governance.

  3. Trust-based compliance beats enforcement: Informing taxpayers of discrepancies and nudging voluntary correction yields faster, higher-quality compliance than punitive audits. This applies across tax, securities, and banking supervision.

  4. Humans drive AI; AI doesn't drive humans: Chairman Agarwal's anecdote (developing code in 5–6 hours with AI) illustrates AI as a productivity tool. Officers, analysts, and investigators must understand and validate AI outputs; blind reliance invites error and corruption.

  5. Multimodal, cross-sectoral data integration is the force multiplier: Siloed data in government departments misses patterns visible only when financial, behavioral, textual, and visual data converge. Project Insight, RBI's aggregation service, and unified crime investigation platforms show exponential gains from breaking silos.

Key Topics Covered

  • Project Insight 2.0: Next-generation AI-enabled taxpayer services, compliance architecture, and return processing
  • Sovereign Language Models (SLMs): Domain-specific, data-controlled LLMs tailored for tax administration
  • AI in Law Enforcement: Computer vision, anomaly detection, graph analytics, and crime prediction applications
  • Financial Intelligence & Risk Analytics: AI-driven detection of mule accounts, fraud rings, and financial anomalies
  • RBI's Mule Hunter Initiative: Real-time detection of fraudulent accounts using ML across 26+ banks
  • MahaCrime OS: Investigation co-pilot for complex cases (cyber crime, organized narcotics, economic fraud, sexual offenses)
  • SEBI's AI Applications: Radar (ad compliance), Sudashan (misinformation detection), InfoMerge (investigation workflow)
  • Saksham Nudge Strategy: Seven-step behavioral intervention framework using data analytics for voluntary compliance
  • AI Governance Frameworks: Seven sutras and six pillars for responsible AI adoption in financial regulation
  • Multimodal AI & Data Integration: Combining financial, visual, textual, and behavioral data for intelligence
  • Civil Liberties & Bias Mitigation: Human-in-the-loop safeguards, fairness validation, and privacy protections
  • International AI Misuse: 17-country task force addressing synthetic identities, deepfakes, and AI-assisted financial crime

Key Points & Insights

  1. Operational AI Transition: AI has moved from pilot phase to scaled production—1.11 crore taxpayers filed updated returns via nudges; ₹8,800+ crore revenue impact; ₹99,000 crore in foreign assets declared; Mule Hunter deployed across 26 banks with 80–90% accuracy vs. 20–30% for legacy rule-based systems.

  2. Sovereign Data Architecture is Critical: Speakers emphasized building sovereign, India-centric AI solutions rather than relying on generic global models. The proposed BharatVers platform and sovereign SLMs for tax administration prioritize data control, security, and compliance with Indian regulations.

  3. Conversational AI Reduces Compliance Friction: Project Insight 2.0's multilingual chatbots, context-driven prefilling, and real-time guidance will shift compliance from enforcement-heavy to enablement-focused. Annual Information Statement (AIS) enhancements aim to provide taxpayers with accurate, corrected information before filing.

  4. Graph Analytics Unmask Collusion: RBI's Mule Hunter and financial institutions' fraud detection use graph neural networks to identify networks of bad actors moving money in circles, creating false income appearances. Traditional supervised learning fails here; unsupervised and ensemble methods succeed.

  5. Human-in-the-Loop is Non-Negotiable: All speakers stressed AI must augment, not replace, human judgment. Intelligence officers, investigators, and analysts review AI-generated leads before enforcement action. Autonomous systems without human oversight risk bias, false positives, and eroded public trust.

  6. Behavioral Nudges Outperform Penalties: Saksham nudge strategy shows voluntary compliance increases 3x when taxpayers receive data-backed guidance showing "here's what we know about you." This soft-touch approach reduces litigation, penalties, and adversarial relationships.

  7. Multimodal Intelligence Multiplies Insight: Combining structured financial data, unstructured text (documents, communications), satellite imagery, facial recognition, and telecom metadata enables law enforcement to detect anomalies invisible to single-modality systems.

  8. Bias & Fairness Require Continuous Testing: Professor Mossum and others warned that ML models trained on historical data can embed discrimination (e.g., against minority communities). Fairness validation, red-teaming, and diverse training data are essential to prevent judicial and administrative injustice.

  9. Real-Time Risk Scoring Shifts from Reactive to Predictive: Instead of catching fraud after it happens, RBI's Digital Payments Intelligence Platform aims to score transactions in real-time, preventing harm before accounts are compromised.

  10. Capacity Building & Training Are Foundational: Income Tax Department's emphasis on staff upskilling, SEBI's democratization of AI (non-IT staff building tools), and Maharashtra Police's training of 233 investigators demonstrate that institutional readiness—not just technology—determines success.

Notable Quotes or Statements

  • Chairman Ravi Agarwal: "You have to drive AI rather than AI driving you. Humans have to drive the AI for that to happen you have to build capacity in your resources in the human resources."

  • Chairman Agarwal (on code development): "Otherwise development of this code would have taken months but then with AI, with spending five to six hours one was able to come up with reasonably robust and matured code... That is the power of AI."

  • Chairman Agarwal (on motive): "While we have in AI a very powerful tool we need to be conscious about how do we put in place and apply AI in our overall governance overall welfare of people happiness of people while being ethical."

  • Professor Mossum: "AI brings up the issues. AI generates a lead but the lead is processed by the human to maintain trust in the system... We cannot make this completely autonomous."

  • Professor Mossum (on bias risk): "We are a very diverse society with so many castes and social strata. If we got those biases in our models, that will be very very devastating."

  • Shari Soenduati (RBI): "Mule accounts in our banking system is a real challenge and given the huge volume of data we would never humanly be able to do it without the use of technology."

  • Shri Shashi Bushan Shukla (CBD): "We should use AI in an informative manner... to enable taxpayers to pay correct taxes at the right time without any penalty... this is what we are trying to achieve."

  • Shri Harsh Podar (Maharashtra Police): "The investigative co-pilot is supposed to serve as a mechanism to tune into all these silos of data and make sense and make an investigative path for the investigating officer."

Speakers & Organizations Mentioned

Government Agencies:

  • Income Tax Department (CBD – Central Board of Direct Taxes)
  • Reserve Bank of India (RBI)
  • SEBI (Securities & Exchange Board of India)
  • Maharashtra Police / Marvel (special purpose vehicle for Maharashtra AI)
  • Ministry of Home Affairs
  • DRDO (Defence Research & Development Organisation)
  • National Institute of Direct Taxes
  • I4C (Indian Financial Crime Consortium)

Technology & Consulting Partners:

  • LTI Mindtree (Shri Romesh Reauru, Shiri Nasanti)
  • Teradata (Martin Wilcox, SVP AI-driven analytics)
  • Microsoft (collaboration on MahaCrime OS)

Speakers:

  • Amandep Dhanova (IRS 2018, moderator)
  • Shri Ravi Agarwal (Chairman, CBD; IRS 1988 batch)
  • Shri Abhishek Kumar (Commissioner, Income Tax)
  • Prof. Mossum (Founding Head, YSAI, IIT Delhi)
  • Shri Hersh Podar (IPS, SP Nagpur; CEO Marvel)
  • Shri Surendeepatty (Chief GM, RBI; Head Fintech)
  • Shri Ganesh Ram Ganesh (Founder, CyberI; cyber security expert)
  • Shri Anish Pande (Executive Director, SEBI)
  • Shri Shashi Bushan Shukla (Principal Commissioner, CBD)
  • Shri K. Mahadevan (Joint Commissioner, Income Tax)
  • Shri Anan Jasur (Principal Chief Commissioner, CBD Delhi)
  • Shrimati Mona Singh (Member, CBD)

International References:

  • Google DeepMind (Demis Hassabis, AGI timeline discussion)
  • PMML, Mojo, ONNX (ML model formats)
  • CRS/FATCA frameworks (international automatic information exchange)

Technical Concepts & Resources

AI/ML Models & Architectures

  • Large Language Models (LLMs): GPT-based models for conversational AI, document processing
  • Small Language Models (SLMs): Domain-specific, parameter-efficient fine-tuning (LoRA – Low-Rank Adaptation) for sovereign, controlled models
  • Graph Neural Networks: Collusion ring detection, network analysis, entity relationship mapping
  • Ensemble Methods: Combining statistical, supervised, and unsupervised ML for anomaly detection
  • Multimodal AI: Integration of text, images, audio, video, and structured financial data
  • Agentic AI / AI Agents: Autonomous workflow triggering and guided investigation paths
  • Generative AI for Code: LLMs generating application code to reduce development time

Data & Platforms

  • Project Insight 2.0: Next-gen data warehouse with AI infusion for taxpayer profiling, risk assessment, e-verification, campaign optimization
  • Insight 1.0: Existing data warehouse ingesting income tax, third-party (banks, stock exchanges), and foreign jurisdiction data
  • Mule Hunter.ai: ML-based detection of fraudulent accounts across 26+ banks; 857+ features engineered; real-time scoring planned
  • BharatVers / BharatAI: LLM agent platform for building multi-agent systems; sovereign, India-built alternative to generic platforms
  • MahaCrime OS: Investigation co-pilot for cyber crime, organized narcotics, economic fraud, sexual offenses (Maharashtra Police)
  • Digital Payments Intelligence Platform (DPIP): Planned real-time transaction scoring and detection framework (RBI)
  • SEBI Tools:
    • Radar: Compliance checking for mutual fund advertisements
    • Sudashan: Multimodal monitoring of misleading financial content and fin-influencers
    • InfoMerge: Workflow intelligence for investigation case management
    • Cyber security compliance audit tool (three-model ensemble for reduced hallucination)

Data Sources

  • Pan (Permanent Account Number) database – 80+ crore individuals
  • ITR (Income Tax Return) filings – 9–12 crore annually
  • Specified Financial Transactions (SFT) – 650+ crore data fields
  • Annual Information Statement (AIS) – aggregated third-party financial data
  • CRS/FATCA – automatic exchange of foreign asset info with 100+ countries (~50 lakh pieces annually)
  • TDS returns, bank data, stock exchange data, telecom CDR (Call Detail Records)
  • Satellite imagery, social media (Facebook, Instagram, PhonePe, Google Pay)
  • Assessment orders, appeal orders, judicial precedents

Key Methodologies

  • Saksham Nudge Strategy: Seven-step framework—Sankalan (data collection), Anushandhan (analysis), Karyavai (action), Sangathan (communication), Aadhyayan (feedback), Aak (handholding), Adikar (enablement)
  • Conversational Chatbots: Context-aware, intent-driven NLP for taxpayer guidance and grievance resolution
  • Prefilling & Auto-Correction: Using third-party data to pre-populate returns and flag discrepancies
  • Litigation Risk Assessment: Tagging issues in assessment/appeal orders via LLM; predicting case viability
  • Anomaly Detection: Unsupervised/semi-supervised learning to identify unusual patterns in financial behavior
  • Red-Teaming & Fairness Testing: Systematic bias detection and mitigation before deployment
  • Centralized vs. Federated Learning: Aggregating insights from multiple banks (Mule Hunter) without centralizing sensitive data

International Frameworks & Initiatives

  • MANO Framework (Government of India, adopted Nov 5, 2024): Moral and ethical AI, Accountable governance, National sovereignty, Objective and legitimate systems
  • RBI's Seven Sutras for AI Governance:
    1. Trust in the system
    2. People first
    3. Innovation over restraint
    4. Transparency
    5. Responsibility & accountability
    6. Data & algorithm quality
    7. Resilience & continuity
  • AI Sandbox (RBI proposal): Secure testing environment with cross-sectoral, anonymized data for model developers
  • 17-Country OECD/G20 Task Force (led by India): Addressing AI-assisted financial crime (synthetic identities, deepfakes, fabricated documents)

Civil Liberties & Governance Safeguards

  • Bias Mitigation: Ensuring fairness across caste, gender, socioeconomic strata (especially critical in India's diverse society)
  • Human-in-the-Loop: Mandatory human review before enforcement action; AI generates leads, humans validate
  • Explainability & Transparency: Interpretable models; documented decision-making
  • Privacy & Data Minimization: Sovereign hosting, data residency, limited scope of surveillance
  • Accountability Frameworks: Clear liability for AI-induced errors; redressal mechanisms for affected citizens
  • Civil Liberties Oversight: Ensuring surveillance doesn't erode constitutional protections

Contextual Notes

  • Event Date & Venue: Income Tax Department symposium (exact date not specified but references "AI Impact Summit" with theme "Sarvjanay Suraj Sukhai" – welfare and happiness for all)
  • Policy Backdrop: Implementation of the new Income Tax Act 2025 (effective April 1, 2026), positioning the tax system as technology-driven with rule-based automation to reduce interpretational ambiguity
  • Institutional Maturity: Income Tax Department is positioned as a pioneer in adopting AI responsibly—emphasizing citizen welfare, taxpayer services, and trust-based voluntary compliance over coercion
  • International Context: India's MANO framework reflects alignment with global AI governance trends while prioritizing national sovereignty and cultural values