SWIFT announced on Thursday that a consortium of 23 global banks will begin live testing of a new AI-powered payment routing system beginning in the second week of April, with Deutsche Bank, Standard Chartered, and BNY Mellon among the confirmed first-wave participants. The system — developed over the past eighteen months at SWIFT's La Hulpe research facility in collaboration with a small number of machine learning specialists from its partner institutions — uses a transformer model trained on three years of anonymised settlement data to assign routing probabilities across available correspondent banking paths in real time, replacing the current rule-based routing logic that has underpinned SWIFT's messaging infrastructure since the 1990s.

The commercial case is straightforward and the failure rate problem is real. SWIFT's own data indicates that approximately 6% of cross-border payment messages initiated on the network encounter at least one routing failure — a bank along the correspondent chain that cannot execute the next leg of the transaction because of liquidity constraints, cut-off time mismatches, compliance holds, or technical error. Each failure introduces delay and often cost: the sending bank typically must re-initiate through an alternative routing path, a process that in the worst cases can add twenty-four to forty-eight hours to a transaction that should, in principle, settle the same day. SWIFT's 40% failure reduction target implies bringing that 6% rate to approximately 3.6%, which at current message volumes of 53.3 million per day would translate to roughly 1.3 million fewer failed transactions daily across the network.

The technical architecture SWIFT has described operates as a pre-routing intelligence layer rather than a replacement for existing messaging infrastructure. When a payment instruction is submitted, the system evaluates, in milliseconds, the liquidity positions of the relevant correspondent banks as inferred from recent message flow patterns, the time-zone and cut-off schedules for each potential routing path, historical failure rates for each path for that currency pair and payment type, and current compliance screening queue depths at intermediate banks. It then ranks available paths by predicted success probability and presents the top three to the sending bank's automated systems, which can be configured to accept the highest-ranked recommendation without human intervention or flag for review based on threshold rules.

The competitive context for this announcement is the continued growth of the blockchain-based settlement networks that SWIFT's member banks have been investing in at the same time as they remain committed to the traditional messaging infrastructure. JPMorgan's Kinexys platform, which settled more than $2 billion per day in 2025, operates on a fundamentally different architecture — assets move directly on a permissioned ledger rather than through correspondent banking chains, eliminating the failure modes that the new routing AI is designed to address. mBridge, the central bank digital currency platform operated by a consortium of Asian central banks, offers similar direct settlement guarantees for participating currencies. SWIFT's response — improving the existing system rather than rebuilding on ledger technology — reflects a deliberate institutional judgment that the correspondent banking model remains viable at scale, and that AI can extend its operational life by removing friction that currently makes it vulnerable to displacement. Whether that judgment proves correct will depend on whether a 40% reduction in failure rates is sufficient to retain the clients who are simultaneously expanding their investment in SWIFT's competitors.