Infosys Ltd. rolled out Topaz agent assistants at two regional rural banks participating in a National Payments Corporation of India sandbox extension, automating know-your-customer follow-ups and customer grievance triage while keeping payment switches and unified payments interface rails outside the experimental stack. Bank officials said the agents summarise missed document fields for account openings and draft first responses to complaints about failed Aadhaar-enabled payment system transactions, tasks that previously backlog branch managers during harvest-season rushes.
What runs inside the sandbox
The sandbox isolates large language model prompts from live switch authorisation paths, a condition NPCI imposed after earlier fintech pilots accidentally logged sensitive personal data in third-party analytics buckets. Infosys hosts inference in Indian data centres and maps agent actions to predefined workflows—send SMS reminders, schedule field officer visits, escalate dispute tickets—rather than allowing open-ended tool use on core banking databases.
Topaz agents read policy PDFs and circulars the banks upload nightly, citing paragraph numbers in drafts human staff approve before messages leave the bank’s domain. No autonomous debits or credit limits change without a signed-off rule engine step.
Why rural banks need help
Regional rural banks serve farmers and micro-enterprises in districts where smartphone banking adoption outpaced staff training during the pandemic. Customer queries spike when subsidy credits fail or when KYC reverification deadlines approach; agents reduce average handling time but do not replace field visits required for land-title collateral.
NABARD inclusion metrics reward digital transaction growth, yet branches still face physical queues during government benefit distribution weeks. Automating paperwork chases frees officers to focus on loan restructuring conversations that algorithms cannot conduct responsibly.
Regulatory guardrails
The Reserve Bank of India’s fintech department expects banks to log every agent suggestion and human override for audit, mirroring requirements for algorithmic credit scoring pilots. NPCI’s extension runs through December unless banks demonstrate zero leakage of full primary account numbers into model training sets.
Privacy advocates warn that even redacted transcripts can re-identify villagers in small populations; Infosys said agents mask names in logs used for model tuning and delete raw chat after 30 days unless litigation holds apply.
Competitive context
Tata Consultancy Services Ltd. and Wipro Ltd. market similar agent platforms to urban commercial banks; Infosys targeted rural lenders because Topaz bundles with Finacle core installations many already run. Smaller fintech startups offer voice bots in local languages Infosys plans to add after Hindi and Marathi pilots complete.
Public-sector banks watch the sandbox outcomes before national rollouts; union leaders demand staff headcount protections, so vendors pitch agents as augmenting rather than replacing tellers—a messaging line regulators repeat in press briefings.
What success looks like
Metrics include reduced average grievance closure time and fewer abandoned account-opening applications, not model benchmark scores. If harvest-season queues shorten without increasing payment dispute chargebacks, NPCI may widen the sandbox to cooperative banks that clear the same data-localisation checklist.
For India’s payments rail operator, the lesson is narrow: agentic AI can sit beside UPI’s plumbing if it never touches authorisation keys—a design constraint other markets are still debating while Pune and Nagpur district branches test it on Monday morning queues.
Language and literacy gaps
Branch staff said farmers preferred voice notes in local dialects when agents summarise circular changes; Infosys is recording approved scripts rather than letting models improvise legal terms. Training modules show tellers how to reject agent drafts that cite outdated interest-rate caps—a failure mode testers triggered deliberately in red-team exercises last month.
Regional rural bank chairpersons asked NPCI to extend sandbox hours through kharif loan disbursement peaks when foot traffic doubles; regulators have not committed beyond the December review date but acknowledged seasonal staffing pain in meeting minutes published Friday.
Infosys executives told investors the rural rollout would not move revenue needles this quarter but serves as a reference win for public-sector bids closing in December.
Branch managers in Nagpur district said average grievance acknowledgement time fell from two days to under six hours in pilot weeks, a metric NPCI will publish in sandbox closure reports if dispute rates stay flat through the harvest season.