Moving money faster has compressed the time banks have to answer a more complicated question: How should this particular payment move?
A transaction can arrive through online or mobile banking, an API or, potentially, an artificial intelligence agent. The bank may have several ways to send the money, including batch and instant rails. Before choosing one, it still has to consider fraud, risk, customer requirements, fees, timing and whether the recipient information is correct.
Nilesh Dusane, global head of institutional payments at AWS, said the underlying obligations haven’t disappeared as payment options have expanded.
“You need to do KYC [know your customer], you need to ingest payment instructions, you need to do risk management, you need to figure out the fees and then ultimately route the payment over an appropriate rail,” Dusane told PYMNTS.
The challenge is getting those functions to work within the time available for a particular transaction. An instant rail may be available without being necessary for a payment that isn’t time sensitive. A transaction that does require immediate settlement gives the bank far less time to complete the checks and routing decisions that precede it.
Cross-border payments add another set of variables. An originating bank must determine whether beneficiary information is valid and whether factors such as processing windows and time zones will affect the transaction. It also can’t necessarily answer every question itself.
“There are certain things a bank can do on their own, but then there are certain things which they have to rely on the beneficiary bank” to provide, Dusane said. He added that beneficiary-validation coverage isn’t ubiquitous worldwide. A provider suited to verifying recipients in Europe may differ from one with strong coverage in Japan.
Payment Infrastructure Has to Support the Decision
Infrastructure in this setting extends beyond compute and storage. Banks also need the data and decisioning capabilities required to conduct fraud checks, assess risk and perform KYC while a transaction remains actionable.
Keeping a system online doesn’t ensure that every payment can be completed through its original route. Dusane cited an ACH payment that misses its processing cutoff. Where the payment qualifies, an orchestration layer could direct it to FedNow instead.
Banks don’t necessarily have to rebuild their payment operations all at once to add those capabilities. Dusane said institutions can begin with a defined use case, such as instant payments, corporate payments or tokenization, and use an orchestration layer to direct eligible transactions. Institutions with systems reaching end of life may face broader migrations, while others can modernize individual functions.
Artificial intelligence widens software’s role in payments: it now reaches across fraud analysis, payment operations and, through agents, the initiation of transactions themselves.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
One approach Dusane discussed is the payment foundation model. Rather than using a language model designed principally around text, a payment model can treat individual transactions as tokens and analyze their order and relationship.
“What you really need is the ability to tokenize individual transactions, look at the sequence of what those transactions are, and then do the best you can to predict the next transaction,” he said.
The sequence provides additional context for fraud detection. A plane ticket followed by a hotel transaction and restaurant spending, for example, can establish a pattern. Subsequent activity can then be evaluated in relation to what came before it rather than solely against predetermined rules.
Scale and context are where AI earns its place. Banks must keep up with a steady stream of rule updates from payment schemes — changes that have traditionally been interpreted and implemented manually.
A financial institution operating in several countries can encounter different messaging standards, file formats, data residency requirements and clearing rules. Those requirements can also change monthly or quarterly.
Dusane said banks have traditionally had employees retrieve updated rule files, determine which changes apply to the institution and translate them into system requirements. AI agents can assist with reading successive versions, comparing changes with earlier requirements and developing specifications for corresponding system updates.
“Agents are doing that, looking at those files in context,” he said, while describing a process that retains “the human in the loop.”
Localization is necessary even when institutions use common technology across markets. Each implementation still has to accommodate the payment rails, standards, data requirements and operating rules of the jurisdiction where it runs.
For AWS, the work described by Dusane centers on the underlying cloud infrastructure and tools used by financial institutions to connect channels, service providers and payment rails. The payment decisions themselves remain functions of the bank’s customers, controls, use cases and operating requirements.
For banks, the expanding set of payment options is an opportunity. More rails and instruments give them more ways to serve customers — and they now have more tools to act on that, from generative AI for building personalized experiences to payment foundation models that bring intelligence to fraud, risk and routing. The opportunity belongs to institutions that can assemble the information each payment requires and act on it while the payment is still in motion — turning speed into better, more personalized service.
Add as Preferred Source
