Banks Do Not Struggle with AI Pilots — They Struggle with Scale
An exclusive conversation with Kishan Sundar, CTO, Maveric Systems
India's banking sector has embraced AI experimentation with unusual speed. Credit decisioning, AML transaction monitoring, customer service automation, document extraction — pilots are running everywhere. Yet the gap between a successful proof-of-concept and an enterprise-grade deployment at the scale of a major bank remains wide, expensive, and underestimated.
Kishan Sundar, CTO of Maveric Systems — a specialist in banking technology modernisation — has spent the last two years helping Indian and global banks navigate exactly this gap. His diagnosis is pointed: the bottleneck is rarely the model.
The Pilot-to-Production Chasm
"In our experience, banks can run a successful AI pilot in six weeks. Scaling it to cover 80% of the intended workflow, with the reliability and auditability that regulators require, takes eighteen months. That delta is where most programmes stall."
The reasons are structural, not technical:
| Challenge | Root Cause | What It Actually Requires |
|---|---|---|
| Data quality | Siloed core banking systems with inconsistent schemas | Data lineage infrastructure, not just cleaning scripts |
| Model drift | Production data distribution shifts from training data | MLOps monitoring pipelines, retraining triggers |
| Explainability | Regulators require decision rationale for credit and AML | Model interpretability layer built into deployment |
| Change management | Relationship managers distrust AI-generated recommendations | Operating model redesign, not just a new dashboard |
| Governance | No single owner for AI risk across tech, risk, and compliance | Cross-functional AI governance committee with clear mandates |
What the High-Performers Do Differently
Sundar identifies three practices that separate banks successfully scaling AI from those cycling through pilots:
1. Platform before proliferation. High-performing banks build a reusable MLOps layer — data connectors, feature stores, model registries, monitoring dashboards — before deploying the second AI use case. Banks that skip this step build technical debt at the pace of their pilot velocity.
2. Instrument outcomes, not just outputs. A model that is 92% accurate in testing is meaningless if the business outcome it is supposed to drive — loan processing time, fraud detection rate, customer satisfaction — does not improve in production. Sundar recommends wiring business KPIs directly to model monitoring dashboards from day one.
3. Govern before you scale. The banks that have avoided high-profile AI failures in 2025–2026 share a common characteristic: they established AI risk committees with clear escalation paths before deploying AI into regulated workflows. Retrofitting governance onto a live system is orders of magnitude more expensive than building it in.
The Regulatory Dimension
RBI's guidance on AI in banking is moving from principles to prescription. Sundar expects explainability requirements and model audit trail mandates to become compliance obligations within 18 months — placing banks that have not invested in interpretability infrastructure at a structural disadvantage.
"The question I ask every CISO and CTO in banking is: if RBI asks you to explain why your AI made a specific credit decision on a specific date two years ago, can you answer that question in 48 hours? Most cannot. That is the gap we need to close."
The Roadmap Sundar Recommends
-
Audit your AI inventory — every model in production, its training data vintage, its monitoring status, and its governance owner.
-
Standardise on a feature store — shared feature engineering prevents duplicated effort and enables consistent model behaviour across products.
-
Build the explainability layer now — SHAP values, LIME, or custom attribution methods embedded in the model serving infrastructure, not bolted on later.
-
Run quarterly AI risk reviews at board level, not just technical deep-dives within the data science team.
-
Invest in MLOps talent as a strategic priority — this is the scarcest capability in Indian banking technology right now.