AI agents start shopping: Mastercard unlocks agentic commerce in India

Even as India’s digital payments ecosystem continues to expand at an unmatched scale, it now finds new territory. Mastercard, at the India AI Impact Summit 2026, has demonstrated the first authenticated agentic commerce transactions in India. The financial tech giant insist they are working with AI companies as well as fintechs and merchants, to accelerate AI-led commerce in India and indeed the Asia Pacific region.

The first payments were done on Mastercard cards issued by Axis Bank and RBL Bank, for tokenised agentic purchases using Cashfree Payments, Juspay, PayU and Razorpay payment aggregators on merchants including Swiggy, Instamart, Vi and Tira. This is aligned with the Mastercard Agent Pay framework, a set of guidelines for tokenisation and security.

Users will have a choice of their own AI agents to work with, and Mastercard hopes they can tackle interoperability with scale and ubiquity. Nitendra Rajput, Senior Vice President and Head of Mastercard AI Garage, tells HT that insights from adoption and payment behaviours feed into “network intelligence” — AI models trained on more than 160 billion transactions processed globally each year.

Add a looming a spectre of AI agents attempting transactions on a user’s behalf. For context, Reserve Bank of India (RBI) data shows credit card spends touched ₹2.12 lakh crore across 552 million transactions in January, just below ₹2.17 lakh crore record of September 2025. As regulators tighten data localisation and security norms, global networks such as Mastercard are doubling down on India’s high-growth market, increasingly leveraging AI.

Edited excerpts. Q. What can go wrong when AI agents are authorised to transact financially? The biggest risks lie in how these systems optimise, scale, and relied upon. One concern is misaligned optimisation, where an agent may prioritise speed or efficiency in ways that unintentionally conflict with a customer’s best interests.

Another risk is cascading errors, because automated workflows amplify mistakes into larger failures if oversight is weak. A third risk is over-reliance, where users may stop questioning or validating AI decisions. It is essential to build systems with strong explainability, continuous monitoring, independent audits, and meaningful human oversight.

Q. As fraud becomes an AI-versus-AI battleground, how does Mastercard ensure its models consistently outmanoeuvre adaptive attacks? Attackers use automation, synthetic identities, and machine learning to scale operations. Our fraud models run multiple AI engines in parallel, what we call a “mixture of experts”, allowing us to evolve models continuously and stay ahead of attackers.

This approach has enabled us to stop billions of dollars in fraud across our network. Global intelligence alone isn’t enough. India is unique with high digital adoption, diverse devices, assisted commerce, and evolving social patterns. Our models combine localised learning, tuned continuously by our teams.

This hybrid approach ensures we stay structurally ahead, not just reactive. Q. As data localisation debates intensify, do they constrain or reshape AI training and cross-border fraud intelligence? Data localisation is fundamentally reshaping AI systems in financial services. For Mastercard, privacy and compliance are non-negotiable, and systems must respect national data boundaries.

This means strong governance, privacy by design, and clear accountability in training, monitoring, deployment, and model evolution. But, fraud does not respect borders. Patterns often emerge globally before they appear locally, and architectures with latest algorithms that work in other geographies are applied with localised data.