NPCI’s AtOM uses AI agents to automate UPI change testing, validation and readiness certification across partner institutions.
MUMBAI: The National Payments Corporation of India launched two agentic AI platforms, AiNxt and AtOM, at the Bharat AI Zone during the Global Fintech Fest on September 9. The move targets a genuine operational bottleneck: coordinating changes across UPI’s sprawling network of banks and payment apps, which currently processes over 24.5 billion transactions monthly worth roughly ₹30 lakh crore, accounting for 89% of India’s digital retail payments.
The two platforms solve different problems. AiNxt lets banks and developers build their own AI agents for internal tasks. AtOM automates the technical coordination between NPCI and hundreds of partner institutions whenever the network needs to change.
The Problem AtOM Actually Solves
Anyone who has watched a bank roll out a new UPI feature knows the friction involved: emails between technical teams, manual sandbox testing schedules, physical sign-offs and asynchronous rollouts across institutions. AtOM replaces this with software agents that negotiate directly with each other. When NPCI releases a specification update, partner bank agents automatically validate their local configurations, run test cases and generate cryptographically signed readiness certificates, compressing what used to be a multi-institution coordination headache into a machine-orchestrated workflow. Built on the JSON-RPC 2.0 messaging standard, AtOM currently focuses on UPI but was designed to extend across NPCI’s other rails, including RuPay, FASTag and Bharat Connect.
What AiNxt Actually Offers
AiNxt takes a different approach, an open-source, Bring-Your-Own-Model platform that lets banks and fintechs build their own AI agents without depending on any single vendor’s proprietary system. The suite includes AiNxt OS for managing agent execution, AiNxt Code for embedding agent-building into existing developer tools, a command-line interface for technical teams and an Enterprise layer for governance and compliance monitoring. Because it’s BYOM-compatible, regulated institutions can run these agents entirely within their own private infrastructure, an important detail given India’s data localisation requirements.
Why the Auditability Matters
Every message exchanged through AtOM carries a digital signature, creating a traceable record that regulators and internal audit teams can inspect after the fact. This matters because introducing autonomous decision-making into a regulated payments network raises an obvious question: what happens when something goes wrong? NPCI’s answer, for now, is cryptographic traceability rather than pre-emptive liability rules, a meaningful safeguard, but not the same as knowing who’s actually responsible when an agent gets something wrong.
A Separate, Bigger Question: Agents Spending Your Money
Running in parallel to these operational tools is a more consequential, still-unannounced framework called the Unified Agent Protocol, which would let AI agents make small autonomous purchases on a consumer’s behalf without per-transaction approval. This would work by adapting two existing UPI features: UPI Circle, which currently lets one person delegate payment authority to another, repurposed to authorise an AI agent instead and Reserve Pay, which ring-fences funds, currently capped at ₹10,000 for up to 90 days, for future automated debits.
This is a fundamentally different risk category than AtOM’s backend orchestration. AtOM operates between institutional systems with cryptographic safeguards; UAP would involve consumer money moving based on an AI agent’s judgment, opening up questions around prompt injection attacks, model hallucination, and what happens if an agent buys the wrong thing or the wrong quantity.
The Liability Gap Nobody Has Filled Yet
This is the honest crux of the story: current regulatory guidelines don’t specify who bears financial responsibility if an operational agent falsely certifies a system change under AtOM, or if a consumer-facing agent under the future UAP framework executes an incorrect purchase. Whether that liability falls on the model provider, the app developer, the sponsor bank, or the infrastructure operator remains undefined and this is arguably the single most important gap standing between these platforms and full-scale trust.
Part of a Broader Push
These launches didn’t happen in isolation. NPCI also unveiled FiMi Banking, described as India’s first sovereign compact AI model built for retail banking, developed with Google using the Gemma architecture and piloted with HDFC Bank, alongside open-sourced safety benchmarks for evaluating financial AI. A separate showcase, Token Nxt, featured 20 fintechs demonstrating fraud detection and quantum-safe cryptography tools. Taken together, this represents NPCI’s deliberate strategy of building open, publicly accessible AI infrastructure rather than letting private players like Mastercard, Visa or specific fintech-OpenAI partnerships control how agentic commerce develops in India.
What Remains Unpublished
Several practical details are still missing from public disclosures: which banks and payment service providers are actually piloting AtOM in production, what licensing model governs the open-source AiNxt repositories, the cryptographic key infrastructure managing agent identities and any quantitative data on how much faster partner onboarding has actually become. These aren’t minor omissions, they’re the details that will determine whether this is a operational leap or largely aspirational infrastructure at this stage.
What to Watch
The real test will be adoption. Watch for whether major UPI sponsor banks and popular payment apps like PhonePe, Google Pay and Paytm actually integrate AtOM for their change certifications, whether developer activity picks up on AiNxt’s open-source repositories, and crucially, whether the RBI issues formal circulars establishing liability boundaries for autonomous agents before the Unified Agent Protocol moves from concept to consumer product. Until those liability questions get answered, AtOM and AiNxt represent a genuinely useful backend upgrade, while the more ambitious vision of AI agents spending your money autonomously remains, for now, a work in progress.
About the Author
