The gross NPA ratio of scheduled commercial banks reached a historic low of 2.15% as of September 2025, the lowest level since 2010 to 2011, according to RBI provisional data. Provisioning has improved. Capital buffers are robust. Profitability has recorded six consecutive years of growth. Yet within this picture of institutional health, a structural vulnerability is accumulating in slow motion: the composition of the credit book has changed, and the risk management architecture has not kept pace.

The Macro Dynamics: India’s “Debt Shift” and the Restructuring Market

The divergence between personal loan growth and industry credit growth over the past decade represents more than a portfolio rebalancing. It reflects a fundamental change in the nature of risk that Indian banks are carrying.

India’s household debt climbed to 41.3% of GDP at end-March 2025, marking a sustained rise from the five-year average of 38.3%, with non-housing retail loans, including personal loans, credit cards, and auto loans, accounting for 55.3% of total household borrowings as of September 2025.

As Professor Dirk Bezemer of the University of Groningen identifies in his research on the “Debt Shift Trap,” the decoupling of bank credit from non-financial firms in favour of household consumption often precedes systemic instability. Unlike industrial loans, personal loans lack asset-price feedback loops. There is no collateral whose rising value partially offsets credit risk. When stress arrives, it arrives fast and uniformly, particularly among lower-income households where the debt-to-income ratio leaves little buffer.

Traditional prediction models were designed for a credit landscape that no longer exists. What banks now require is not a more accurate probability of default. It is a reasoning capability that can identify early warning signals, model liquidity trajectories under multiple stress scenarios, and propose proactive restructuring terms before a borrower crosses into SMA-2 territory. The window between SMA-0 and formal NPA is where value is created or destroyed. Rule-based systems are structurally too slow to act in that window.

The MSME Credit Gap: Identifying Structural Roadblocks

The MSME sector presents the sharpest illustration of what happens when credit infrastructure fails to match the actual risk profile of borrowers. CRISIL Intelligence estimates the total MSME credit gap has widened to approximately ₹117 lakh crore as of FY2025, with only 27% to 28% of total MSME credit demand being met through formal financing channels.

The structural barriers are well-documented and persistent.

Documentation asymmetry sits at the foundation of the problem. Micro-businesses frequently operate on informal accounting. The absence of audited financials or consistent GST history makes traditional credit assessment systems effectively blind to enterprises whose actual cash flows are healthy. Collateral dependency compounds this: a rigid focus on fixed assets locks out MSMEs that do not own commercial property, regardless of operational viability. The result is that interest rates for unsecured MSME credit carry a risk premium that reflects institutional uncertainty rather than actual borrower risk, creating a self-reinforcing exclusion cycle.

The delayed payment problem is equally material. Delayed payments from larger enterprises to MSME suppliers account for over 4.6% of India’s Gross Value Added, forcing small businesses into perpetual working capital firefighting rather than strategic growth. This liquidity pressure is precisely the kind of stress that appears in behavioral data weeks before it appears in formal credit metrics, yet traditional systems only intervene once the formal metrics have already deteriorated.

The transition from SMA-0 through SMA-1 and SMA-2 to NPA is where the banking system’s current architecture fails most visibly. By the time a borrower reaches SMA-2, rule-based systems are typically proposing responses that should have been initiated at SMA-0. The cost of that delay is not just the borrower’s business. It is the bank’s asset quality, provisioning burden, and the compounding write-off that follows.

The Generative Shift: How AI Transforms Restructuring Logic

The distinction between traditional analytics and agentic AI is not primarily about speed. It is about the nature of the output.

A model that outputs a probability of default leaves a human analyst to determine what to do with that probability. An agentic system reasons over that probability alongside cash flow data, sector stress indicators, GST filing patterns, and the specific terms of the borrower’s existing obligations, then proposes a restructuring term sheet that is both financially defensible and compliant with the relevant RBI circular. The shift is from decision support to decision architecture.

For the MSME credit gap specifically, generative AI addresses the documentation asymmetry problem directly. Where formal audited accounts are absent, the model reasons over GST filings, UPI transaction patterns, trade receivable data from TReDS, and sectoral behavioral signals to construct a reasoned credit profile that a traditional scoring system would never produce.

Deep Dive: The AI-Driven Loan Restructuring Eligibility Engine

For mid-to-large Indian banks, the loan restructuring eligibility engine is no longer a reporting tool. It is a component of the bank’s operating infrastructure, functioning as an intelligent triage system that manages risk volume at a scale human review teams structurally cannot replicate.

The engine identifies restructuring candidates specifically within the SMA-1 and SMA-2 windows, before the slide into NPA becomes unavoidable. Through what-if analysis and scenario stress-testing, the engine simulates how specific external shocks, such as raw material price spikes, government payment delays, or sector-level demand contraction, affect a particular borrower’s liquidity position. The output is not a flag. It is a proposed term sheet, grounded in the borrower’s actual data, ready for human review and issuance.

The technical architecture that makes this possible rests on three components. A VectorDB and RAG layer grounds the AI in real-time enterprise data and the current library of RBI regulatory circulars, ensuring that proposed restructuring terms are not just commercially logical but legally compliant with hard regulatory requirements. An agentic reasoning layer decomposes complex requests into multi-step tasks, autonomously conducting due diligence, pulling ERP data, and verifying policy compliance without manual hand-offs between systems or teams. TReDS integration treats the Trade Receivables Discounting System as Digital Public Infrastructure, securitising receivables as a high-quality, self-liquidating data input that unlocks private capital for MSME borrowers whose balance sheets would otherwise disqualify them from restructuring support.

The commercial case is grounded in verified data. EY’s research on GenAI in Indian financial services projects productivity improvements of 34% to 38% across financial services by 2030, with banking operations specifically potentially reaching 46%, driven by automation of credit assessment, customer service, and regulatory reporting workflows. KPMG’s global banking survey found 29% of banks are targeting operational cost reductions exceeding 20% by 2030, with AI integration identified as the primary enabling technology alongside data analytics.

The banks that close that gap will be those that moved beyond discrete AI tools to an integrated agentic architecture operating across the restructuring lifecycle.

Roadblocks to Adoption: Governance, Sovereignty, and the 2026 Outlook

Deploying autonomous systems in a regulated financial environment requires governance architecture that is as rigorous as the AI architecture itself.

The model hallucination risk is not theoretical in a high-stakes restructuring context. An agent that confidently proposes a moratorium period or interest rate concession that violates the specific terms of an active RBI circular does not merely create a compliance problem. It creates a legal liability. RAG-based grounding against a continuously updated regulatory corpus is the architectural solution, not an optional enhancement.

The human-in-the-loop imperative is equally non-negotiable. To prevent algorithmic redlining and systematic bias in MSME lending, agents must be designed with self-reflection capabilities: the ability to identify when confidence in a recommendation drops below a programmed threshold and to seek human guidance before a term sheet is finalized. This is not a limitation of the technology. It is the governance architecture that makes the technology deployable in a regulated environment.

Digital sovereignty is the emerging frontier. Indian regulators are closely monitoring global developments including the EU Tech Sovereignty Package. For Indian banks operating or planning to operate with AI systems, this points to an investment imperative: localized data fabrics, model transparency frameworks, and documented AI governance that anticipates rather than reacts to regulatory requirements.

2026 Strategy: Three Critical Takeaways for Banks

For bank technology leaders evaluating how to position their institutions for the agentic era, three architectural decisions are foundational.

1. The first is modularization for agents. Monolithic legacy core systems are structurally incompatible with agentic workflows. The transition to headless application services that expose capabilities via clean APIs is a prerequisite, not a future-state aspiration, for any bank that intends to deploy autonomous restructuring intelligence at scale.
2. The second is a semantic layer investment. An Enterprise Knowledge Graph that ensures agents and human teams share a unified, query-able understanding of business entities, borrower relationships, and policy constraints is the connective tissue of the agentic enterprise. Without it, agents operating across multiple systems produce outputs that are locally coherent but globally inconsistent.
3. The third is agentic governance through policy-as-code. As autonomous systems begin to operate at machine speed across restructuring workflows, the speed advantage is only sustainable if the governance framework keeps pace. Policy-as-code, where risk limits, compliance rules, and ethical guardrails are encoded as executable constraints on agent behavior, is the only architecture that scales with the autonomous workforce rather than being bypassed by it.

The Indian banking sector’s asset quality trajectory is the result of a decade of structural reform. Protecting that trajectory in a credit environment defined by pro-cyclical household debt and an under-served MSME sector requires a fundamentally different tool than the one that got the sector here. The institutions that build reasoning into their restructuring infrastructure now will be the ones that maintain that asset quality through the next credit cycle.

Crizzen works with financial institutions to design agentic restructuring and credit intelligence systems that move beyond rule-based workflows by integrating fragmented borrower data, enabling early distress detection, and orchestrating compliant, end-to-end resolution across the lending lifecycle.

This article is part of the Crizzen Enterprise AI Playbook exploring how AI is reshaping operational models across industries.

#banking #restructuring #loans #financialinclusion #Crizzen #fintech #casestudy

Reach out at info@crizzen.com.

Sources: RBI Financial Stability Report, December 2025; Press Information Bureau, GNPA Historic Low, February 2026; EY India, GenAI Productivity in Financial Services, March 2025; KPMG International, Banking Transformation Survey, September 2025; CRISIL Intelligence, MSME Credit Report, FY2025; RBI Sectoral Deployment of Bank Credit Data, November 2025; Business Standard, India Household Debt GDP Analysis, December 2025; IMARC, India Personal Loan Market Report, 2025