AI in Public Audit: Driving Transparency and Accountability

Key Topics Covered

• AI Strategy Framework for CAG: Four foundational pillars guiding institutional AI adoption
• Information Systems Audit Evolution: Expanding from IT application audits to cyber security audits
• Technology-Based Audit Methods: Big data analytics, OCR/NLP, machine learning, satellite imagery, and drone technology
• Sovereign Large Language Model (LLM) Development: Building country-specific LLMs exclusively for audit purposes
• Data Infrastructure & Management: Secure ETL pipelines, data lakes, and handling of unstructured data (vouchers, PDFs, images)
• Capacity Building Initiatives: Training 7,800+ officers across 140+ offices in data science, AI, and cyber security
• Institutional Safeguards & Security: Data handling protocols and security measures for sensitive government information
• Global Context & Collaboration: CAG positioned as one of few supreme audit institutions with published AI strategy frameworks
• Challenges in Implementation: Data quality, completeness, variety (structured/unstructured), and complexity across diverse government systems

Key Points & Insights

1. Four-Pillar Foundation: CAG's AI strategy rests on (1) using AI in audit operations, (2) auditing government AI systems, (3) capacity building, and (4) R&D infrastructure—with each pillar interdependent and critical to success.

2. Scale of Audit Coverage: CAG operates across 140+ offices with 45,000 employees, having laid 500+ IT audit reports in parliament and audited major systems (GST, IRCTC, GEM application), establishing it as a sophisticated audit institution requiring proportionate AI capabilities.

3. Shift from Sampling to Comprehensive Analysis: Traditional audit methodology selected units based on expenditure parameters. Modern approach analyzes entire datasets to identify red flags systematically, enabling more thorough coverage without proportional resource increase.

4. Multi-Technology Stack: CAG employs OCR, NLP, machine learning algorithms, graph analysis, satellite imagery, and drone technology—each applied to specific audit problems (duplicate beneficiary detection, procurement analysis, infrastructure volumetric assessment).

5. Data Complexity as Core Challenge: The problem space involves massive volumes of unstructured data (vouchers, PDFs, images, speech) across multiple languages and regional variations, compounded by ongoing addition of new data types (GIS, video), making data preparation critical.

6. Sovereign LLM Development: Rather than relying on commercial LLMs, CAG is building audit-specific language models with a roadmap where 50% of development uses CAG-trained officers initially (within 3 years), scaling to 90% internal capability to embed audit domain knowledge.

7. Human-AI Integration Philosophy: The approach emphasizes that "coding is not technical—machines generate code." The critical components are: (1) understanding the audit problem to solve, (2) possessing the data and domain knowledge, and (3) defining meaningful success metrics beyond accuracy alone.

8. Cyber Security Audit Expansion: CAG initiated cyber security audits as pilot programs in major applications in response to increased threat actor activity, expanding assurance scope beyond traditional IT controls to cyber readiness assessment.

9. Reusable Components Strategy: Data processing challenges are addressable by reusing tested components from other government initiatives (GSTN, Centers of Excellence), rather than building entirely new infrastructure—emphasizing pragmatic adaptation over greenfield development.

10. Institutional Collaboration Necessity: CAG explicitly states success depends on cooperation from academia, industry partners, and other government departments—positioning this as a collective governance initiative rather than siloed institutional effort.

Notable Quotes or Statements

• "Coding is not technical. Machines generate code. So the question is: what is the problem that you're trying to solve?" — Emphasizes that domain expertise and problem definition are more critical than technical coding ability in AI implementation.

• "The volume of data that we have is the biggest advantage and the disadvantage." — Captures the dual nature of CAG's position: vast datasets enable comprehensive analysis but create infrastructure and processing complexity.

• "Today I do not cover it and that is where I feel LLMs can really help other than the hype." — Tempers expectations by distinguishing practical utility (exploration of previously uncovered audit areas) from speculative promises.

• "We will not be able to do it alone." — Signals institutional humility and necessity for ecosystem collaboration (academia, industry, other government departments).

• "If you are able to build it, you have the data that we have which is which is in a bounded box." — Suggests competitive advantage in building proprietary LLMs using government's unique, closed dataset.

Speakers & Organizations Mentioned

• Sri K. Sujit — Director, Office of the Comptroller and Auditor General (CAG) of India; primary technical presenter
• Professor Madan — (surname partially unclear in transcript) — Presenter on enabling infrastructure for sovereign LLM initiative
• Dr. Sanjay Kumar — Chief AI and Digital Officer, Wadhwani Foundation; panelist on AI preparedness
• Sri Navin Singh — Principal Director, Commercial and CAG; panelist on audit across organizations and international coordination (UN Board of Auditors)
• Professor Aam Gupta — Associate Professor, IIT Delhi; panelist on technology-societal interface
• Shrinat Chakraati — Senior Vice President, National Institute of Smart Governance; panelist with 30+ years technology advisory experience
• Priyanka Sharma — Partner, Public Finance and Digital Government, KPMG; panel moderator
• Comptroller and Auditor General (CAG) of India — Primary institutional focus; 45,000 employees, 140+ offices
• Wadhwani Foundation — Partner in AI/digital initiatives
• IIT Delhi & IIT Madras — Educational partners for LLM development and talent sourcing
• GSTN (Goods and Services Tax Network) — Referenced as example of government data processing capability
• National Institute of Smart Governance — Governance technology advisory organization

Technical Concepts & Resources

• Big Data Analytics (4 V's): Volume, Velocity, Veracity, Variety framework applied to government audit data
• OCR (Optical Character Recognition) — Used for automating document processing (vouchers, PDFs)
• NLP (Natural Language Processing) — Processing unstructured text across multiple Indian languages
• ETL Pipelines (Extract, Transform, Load) — Secure pipelines for moving data from ministries to cloud infrastructure
• Machine Learning Algorithms & Graph Analysis — Applied to procurement data analysis and red-flag identification
• Satellite Imagery & Geospatial Analysis (GIS) — High-resolution imaging for infrastructure volumetric analysis
• Drone Technology — Physical verification of infrastructure compliance
• AI-Enabled Tool/OCR Toolkit — Beneficiary duplicate detection system requiring non-technical user interface
• Sovereign Large Language Model (LLM) — Custom-built LLM (initiated at IIT Madras/IIT Delhi) for audit-specific applications; roadmap to 90% CAG-officer development within 3 years
• Data Lake Architecture — Centralized storage of structured and unstructured government data from multiple ministries
• Cyber Security Audit Framework — Independent assurance protocols for cyber readiness beyond traditional IT controls
• Information Systems Audit (500+ parliamentary reports) — Established audit methodology for IT applications (GST, IRCTC, GEM)
• Data Quality Challenges: Language diversity, regional variations, multiple unstructured formats (speech, video, GIS to be added)

---

Note on Transcript Quality: The provided transcript contains significant audio transcription artifacts (repeated phrases, unclear passages). The summary prioritizes coherent themes and explicitly stated initiatives while noting where transcript quality limited precision.