CAG to Deploy AI-Powered Auditing System for Efficiency and Transparency

CAG’s AI-Powered Audit System: Efficiency vs Accountability?

On September 18, 2025, India’s Comptroller and Auditor General (CAG) announced its plan to deploy an AI-powered Large Language Model (LLM) for auditing government accounts—a move projected to transform efficiency and transparency in public auditing. The LLM promises automated analysis of massive datasets and anomaly detection across ₹45 lakh crore worth of annual government transactions. But the reliance on AI to evaluate public finance management raises fundamental questions about accuracy, security, and institutional accountability.

The Ambition: Large Language Models in Audit

The policy instrument here is both technological and structural. The AI framework will enable automated evidence collection, risk prioritization, and performance reporting across ministries, Public Sector Undertakings (PSUs), and Panchayati Raj Institutions (PRIs). Backed by digital portals like CAG-Connect, which already integrates nearly 10 lakh auditee entities, the system builds on prior initiatives such as West Bengal’s paperless PRI audits and Telangana’s remote compliance efforts. The goal: shift from exhaustive manual reviews to AI-supported hybrid audits so auditors can focus fewer field visits on high-risk cases.

But alongside the efficiencies, one must consider what is truly being automated. Deep learning algorithms central to LLMs interpret unstructured data—records, reports, transaction histories—and predict patterns. For instance, ₹3.5 lakh crore disbursed annually under GST revenues could see faster anomaly detection through probabilistic analysis. Yet, the technology also moves audits away from subjective judgment embedded in human review.

The Case for AI-Aided Auditing

Proponents argue that integrating AI could eliminate inefficiencies long embedded in India's audit approach. Currently, the traditional system requires months to collect field-level evidence before audits can highlight irregularities. By scanning decades of inspection precedents, LLM-based algorithms promise greater consistency in standards, cutting across subjective variations between auditors. Early testing in Haryana’s Public Works hybrid audit demonstrated faster feedback loops.

Significantly, the sheer scalar challenges of government oversight require such interventions. India has 2.7 crore PRI workers managing decentralized finances—a volume near impossible for manual monitoring. West Bengal’s virtual audit shows that AI integration reduces operational bottlenecks, allowing better real-time feedback and expanding audit reach without proportional staff increases. High-value governance programs—such as ₹15,000 crore allocated under housing schemes—could especially benefit.

Furthermore, AI could bolster enforcement in areas plagued by opacity. Take anomaly detection in GST receipts—a domain where pattern recognition across large transactional datasets is critical to uncover tax evasion. Pattern-based targeting fills gaps left by human-based sampling audits, particularly in sectors with fragmented data trails. Instead of auditing 1.2 crore GST registrants piecemeal, the AI system could identify clusters with higher financial irregularity risks.

The Case Against: The Perils of Over-Automation

While efficiency gains are undeniable, skeptics raise pressing concerns—both institutional and technical. First, data maturity varies drastically between states, ministries, and PRIs. A significant portion of governance records remains paper-bound or poorly digitized. AI systems relying on high-quality inputs risk uneven outcomes across regions like Tamil Nadu with robust e-governance versus states lagging behind.

The accuracy of the Large Language Model itself is another Achilles heel. Unlike deterministic logic tied to traditional auditing methodologies, AI algorithms work probabilistically. How would auditors justify errors generated by opaque black-box models to parliamentary committees? Faulty anomaly detection could wrongly flag compliant entities—a problem amplified by overly aggressive reliance on automation.

Data privacy and cybersecurity loom large. Handling ₹45 lakh crore in transactions across sensitive databases (income disclosures, social benefits, subsidy receivers) presents unprecedented privacy exposure. But institutions that safeguard such information—India's CERT-In or the Department of Revenue—are yet to fully integrate AI security protocols capable of supporting inter-department AI audits. The gap between ambition and infrastructure is real.

Perhaps the biggest critique lies in how far human oversight can balance AI-driven recommendations. The CAG (Duties, Powers, and Conditions of Service) Act, 1971, central to audit autonomy, lacks explicit provisions for blending automation with traditional methods. Institutional safeguards to mandate manual crosschecks on AI findings remain absent, risking overreliance on machines at the expense of human evaluative rigor.

An International Comparison: South Korea’s Automated Oversight

South Korea offers valuable lessons. The Board of Audit and Inspection (BAI) faced similar auditing challenges amidst mounting digital financial transactions. In 2021, it deployed an AI-based audit tool with anomaly detection in public works projects worth $87 billion annually. While earlier pilots flagged more financial irregularities than manual reports, these findings came with caveats—the AI frequently misinterpreted outliers as corruption risks. South Korea mitigated this by requiring mandatory human vetting of all AI-generated anomalies before final audit conclusions.

India’s challenge is similar but distinct. While South Korea combined automation with radical institutional reform, India’s fragmented access policies (state-centre differences) and uneven digitization suggest a tougher road ahead.

Where Things Stand

India’s leap into AI-led auditing is bold, necessary, and fraught with risks. The undeniable benefits—speed, coverage, and improved detection—make the case for reform urgent. However, the true efficacy of this transition will depend on how flexibly institutions design safeguards. Without embedding human accountability in CAG processes, overreliance on AI could backfire spectacularly.

The real risk? That flashy systems overshadow systemic reform. Standardized data formats, continuous model audits, and enforceable explainability measures must precede full-scale adoption.

Exam Practice

• Prelims MCQ 1: Which constitutional articles govern the framework of CAG's duties?
  (a) Articles 138 to 141
  (b) Articles 148 to 151
  (c) Articles 112 to 116
  (d) Articles 151 to 155
  Correct Answer: (b)
• Prelims MCQ 2: What is a key advantage of an AI-powered audit system?
  (a) Double the number of staff visits
  (b) Increased cybersecurity risks
  (c) Enhanced anomaly detection
  (d) Exemption from human review
  Correct Answer: (c)

Mains Question: To what extent does the proposed AI-powered audit system by the CAG address systemic inefficiencies in public financial oversight? Assess its risks, opportunities, and structural limitations with examples from India and abroad.

Practice Questions for UPSC

Prelims Practice Questions

Frequently Asked Questions