The AI risk no bank can manage alone

The most unsettling failure in digital payments may not be a black screen. Imagine a fraud-detection model that remains online but suddenly flags thousands of legitimate transfers as suspicious. Banks may see functioning servers while customers see rejected payments, delayed salaries and complaints no single institution can explain. This is a hypothetical scenario, not a reported incident. But it raises a real question: who tests what happens when several financial institutions encounter a shared technology problem at once? There is a warning from outside AI. In July 2024, a defective update from cybersecurity company CrowdStrike disrupted Windows systems worldwide. Australia’s central bank reported that some payment services and access channels were affected, although the country’s systemically important financial market infrastructure suffered no material impact. The incident was not an AI failure. It did, however, show how a problem originating at one technology supplier could spread through many institutions without any bank first running out of capital. That distinction matters in Southeast Asia. At the Basel Committee and International Conference of Banking Supervisors meetings, held in Bali from September 28 to October 1, Indonesian authorities highlighted cyber threats, fraud, AI and dependence on third-party technology. At an October 5 SEACEN conference, Bank Indonesia emphasized trustworthy AI, resilient and inclusive payments, and regional cooperation. These meetings did not announce the exercise proposed here. They did frame a pressing question: what, exactly, should financial authorities measure when they say a system is resilient? This is not a case for pretending supervisors have ignored technology risk. The Basel Committee’s 2021 operational-resilience principles call for mapping critical operations, dependencies and recovery arrangements. Its December 2025 third-party-risk principles extend attention beyond conventional outsourcing relationships. And a 2025 Financial Stability Board report proposes monitoring AI suppliers for criticality, concentration and substitutability. The foundations already exist. Joint crisis drills are not new, either. In April 2026, the Netherlands’ central bank and other authorities ran an exercise involving payment and capital-market participants, testing how they would communicate and decide during sector-wide disruption. A 2025 IMF working paper modelled shocks to common service providers and payment networks. The question for the AI era is how to build on such approaches when a shared model goes wrong without a conventional outage. An unavailable provider is relatively easy to understand: calls fail, services time out and backup systems are activated. A faulty model can be less obvious. A fraud screen may remain available while rejecting an abnormal share of genuine payments. An automated system may flood human reviewers with alerts. If several providers use the same underlying model, data feed or cloud dependency, the disruption could appear in different institutions at roughly the same time. That is a plausible risk, not evidence that banks across Southeast Asia actually share one model or would respond identically. The region already has a useful reference point. Indonesia’s BI-FAST rules on fraud management explicitly distinguish confirmed fraud from false positives and specify how payment participants respond to alerts or rejected messages. This does not establish that BI-FAST uses AI for these decisions. It establishes something more modest but important: screening errors and their consequences are operational questions that payment authorities must take seriously, not merely problems for a vendor’s software team. A proportionate next step would be to add a shared-model malfunction to existing multi-institution drills. Supervisors, payment operators and participating banks could use synthetic transactions to simulate two shocks: a common technology service becoming unavailable and a fraud-screening component staying online but sharply increasing false alerts. The exercise should test both failure detection and the real ability to fall back to independent procedures. No genuine customer payments would need to be interrupted. The results should be reported through customer-facing measures, not just server uptime. How long would it take participants to recognize a common cause? What share of legitimate payments would be delayed or blocked? How quickly could staff switch to validated fallback processes without inviting more fraud? Would backlogs be cleared before they interfered with payroll or merchants’ settlement? And could customers receive consistent, timely explanations instead of being passed from bank to payment operator and back again? Such drills would also expose an uncomfortable trade-off. A single powerful technology provider can offer better cybersecurity, redundancy and specialist expertise than many institutions could build independently. Fragmenting every system to avoid common suppliers might raise costs and make security worse. The objective is not forced technological self-sufficiency. It is to discover which supposedly separate safeguards actually depend on the same upstream service, and how much real substitutability exists under stress. Coordination must respect institutional boundaries. In Indonesia, Bank Indonesia’s payment-system and macroprudential mandates sit alongside the Financial Services Authority’s banking-supervision responsibilities. Other countries organize oversight differently. SEACEN could provide a forum for anonymized scenarios, shared recovery metrics and lessons learned, rather than assuming that a single regional regulator or entirely new rulebook is necessary. A September 2026 CPMI-IOSCO consultation on critical third-party services to financial market infrastructures adds urgency, although its scope is not every AI application used by a bank. AI promises faster and more accurate financial decisions. Yet a bank can be financially sound while customers struggle to move money because an external technology chain is misbehaving. The next generation of resilience tests should ask more than whether each institution can recover on its own. It should ask whether payments can keep moving while several institutions discover, explain and repair the same failure together. The author works at Bank Indonesia and writes in a personal capacity. The views expressed are his own and do not represent his employer.