Two Singapore playbooks, one week apart
At the Global Fintech Fest in Mumbai this week, Singapore showed up with a document that India's own payments authority does not yet have. On September 11, 2026, Monetary Authority of Singapore (MAS) Managing Director Chia Der Jiun told the conference — by video, after falling unwell before travel — that MAS and the financial industry had already published SAFR, Safeguards for Agentic Finance at Runtime, as an industry white paper. A day earlier, National Payments Corporation of India (NPCI) non-executive chairman Ajay Kumar Choudhary had told the same audience that NPCI was still "examining the protocols that may be required to identify and authorise digital agents" on the Unified Payments Interface. Singapore has the document. India has the position. Since the two countries are actively linking their payment rails, that gap is not academic.
What makes SAFR notable is not its ambition but its restraint. Published July 3, 2026 by MAS alongside financial institutions and fintechs under the MAS BuildFin.ai initiative, SAFR sets out three runtime checkpoints for any AI agent acting on a financial system: establishing the agent's identity and authority, evaluating each proposed action against controls before execution, and keeping an audit record after the fact. Crucially, MAS states outright that the paper "does not constitute regulatory guidance or supervisory expectations" — it is an industry standard institutions can adopt, not a rule they must obey.
The steelman for a harder mandate
Singapore itself supplies the counter-example to "voluntary is always better." Its Online Safety Code for Social Media Services, issued by the Infocomm Media Development Authority (IMDA) under the Broadcasting Act and in force since July 18, 2023, is a binding designation regime: platforms like Facebook, Instagram, TikTok, X and YouTube must meet statutory obligations on content moderation, minor protections and annual safety reporting, backed by fines of up to S$1 million and the threat of service blocking. That mandate did real work. On March 31, 2026, IMDA placed both X and TikTok under "enhanced supervision" after finding that Singapore-linked child sexual exploitation material cases on X jumped 120% year-on-year (33 cases in 2024 to 73 in 2025), and that TikTok hosted 17 cases of terrorism content shared by Singapore-based accounts — the first time such content had been detected there. Those are not hypothetical harms, and a purely voluntary code, with no designation list and no enforceable reporting duty, would have given IMDA far less leverage to demand fixes on a deadline. Where measurable harm to children and public safety is already occurring at platform scale, a statutory floor with real penalties is the proportionate tool, not an overreach.
Where the mandate model is slow
But the timeline is the tell. The legislative basis for online safety designation dates to 2022 amendments; the Code took effect in 2023; and it still took until March 2026 — nearly three years — for IMDA's enforcement machinery to catch and act on the CSAM spike it had ostensibly been designed to prevent. SAFR, by contrast, moved from a named industry initiative to a published runtime standard in months, precisely because it didn't have to clear a legislative process, a public consultation, or a designation list before financial institutions could start building against it. Agentic AI in payments is moving fast enough that a three-year mandate cycle would still be drafting definitions of "agent" while banks were already deploying them.
This is the case for matching the instrument to the technology's maturity, not to the regulator's comfort. Where harm is already observable and platforms have shown they won't self-report accurately — IMDA's own numbers on X and TikTok make that case — hard mandates with fines and disclosure duties are earned and proportionate. Where a technology is still being defined by the institutions using it, as agentic finance is today, an industry-co-designed standard like SAFR can set a workable floor faster than any statute, with the credible threat of hard regulation held in reserve if adoption stalls or abuse emerges.
The lesson for NPCI, and the risk of copying blind
NPCI is not a private platform choosing whether to self-regulate; it operates UPI as public payments infrastructure serving hundreds of millions of users, which is a stronger case for a floor set by rule than for anything purely voluntary. But NPCI does not need to choose between MAS's speed and IMDA's teeth. It can borrow SAFR's three-checkpoint architecture — identity and authority, pre-execution evaluation, audit trail — as a technical baseline now, published quickly and revised as agentic payments actually deploy, while reserving statutory backing for the reporting and liability questions that matter once agents are moving real money at scale. What NPCI should not do is wait for a mandate-first framework modeled on the Online Safety Code's multi-year arc, because by the time it arrives, the agents it's meant to govern will already be live.