AI will not fix fragmented payments infrastructure

Artificial intelligence is becoming increasingly capable of making sophisticated decisions about payments. It can identify patterns across huge datasets, assess risk in real time, model likely transaction outcomes and recommend how a payment should be handled. But there is an uncomfortable question underneath that progress: what happens if the infrastructure cannot act on those decisions? 

That is where the next challenge for AI in payments lies. The industry is investing heavily in intelligence, but many payment environments still depend on static routing rules, manual intervention and infrastructure that is tightly coupled to individual providers. Adding more sophisticated AI on top of that architecture does not automatically create intelligent payments. In some cases, it simply gives businesses greater intelligence about constraints their infrastructure still prevents them from overcoming. 

AI can provide the intelligence, but the infrastructure still determines whether a business can act on it. 

Intelligence is only useful if infrastructure can act 

Payments have become remarkably simple from the consumer’s perspective. A card is tapped, a wallet is opened or a button is pressed and the expectation is that the transaction will complete almost instantly. Behind that experience sits an increasingly fragmented ecosystem of gateways, processors, acquirers, fraud services, token providers, wallets and alternative payment methods. 

According to McKinsey’s 2025 Global Payments Report, the global payments industry generates $2.5 trillion in revenue from approximately $2 quadrillion in annual value flows, supported by 3.6 trillion transactions. AI can potentially make sense of that complexity by considering issuer behaviour, historic acceptance rates, transaction cost, fraud indicators, network conditions and customer context before determining how a payment should be routed. The problem is that identifying the best decision and executing it are two different things. 

If moving a transaction to another provider requires manual intervention, AI cannot deliver genuine real-time optimisation. If payment credentials are locked inside one provider’s environment, identifying a better processor does not necessarily mean the business can move easily. If adding a new payment rail requires another lengthy infrastructure project, intelligent decisioning is still constrained by the architecture underneath it. 

The next phase of AI in payments therefore depends as much on infrastructure as it does on algorithms. This infrastructure gap is explored in the recent white paper The Orchestration Imperative: Turning Payment Complexity into Strategic Advantage, which examines how payments orchestration is evolving from basic connectivity into a broader control and intelligence layer. 

Connectivity should not be confused with orchestration 

This distinction becomes clearer when looking at how payment orchestration is being deployed today. Research by PYMNTS Intelligence among 110 US companies with annual revenues of $10 million or more found that 89% had implemented failover or backup routing. Yet only 41% had adopted automated dynamic routing, while 55% still made routing decisions manually and 68% relied on manual intervention to switch providers during outages. 

That suggests many organisations have built connectivity and redundancy without yet creating an environment capable of continuously making and executing intelligent payment decisions. Connecting several processors gives a business options, but orchestration provides the control layer through which those options can be selected, optimised and adapted as conditions change. 

A truly intelligent environment should be able to decide where a transaction has the strongest chance of success, respond automatically if conditions change and retry or reroute appropriately when a payment fails. That requires more than APIs between different providers; it requires an architecture capable of translating intelligence into action. 

AI needs control, not just data 

Data is usually presented as the fundamental requirement for effective AI, but in payments, control is just as important. Consider payment tokenisation: tokens allow sensitive payment credentials to be represented securely, but token portability and control can become another source of dependency when credentials remain tied to a particular provider. 

The same PYMNTS Intelligence research found that 93% of companies lacked full control over their payment tokens and credentials, meaning just 7% had full token control. It also found that introducing new payment rails remained a major source of friction. That creates an important limitation for AI-led payments. 

An intelligent system may conclude that another provider, rail or payment method offers a better outcome, but that recommendation has little operational value if the organisation cannot make the change easily. This is why the infrastructure conversation around AI has to extend beyond computing power and models. Businesses also need the freedom to act on the intelligence those models produce. 

Payment intelligence will move beyond static rules 

Much payment routing today still relies on relatively fixed logic. A transaction might be sent according to geography, currency, cost or predetermined provider priority. AI creates the potential for a much more dynamic model. 

Routing decisions could increasingly account for how particular issuers are behaving, historical acceptance patterns, current network performance, fraud indicators and the context surrounding an individual transaction. Rather than waiting for performance to deteriorate and then changing a rule manually, systems could identify shifts much earlier. AI could also simulate the effect of routing changes before they are implemented or recommend adjustments within predetermined commercial and risk boundaries. 

The objective should not be autonomous decision-making for its own sake, but controlled intelligence that improves outcomes while preserving security, explainability and accountability. That distinction will become increasingly important as AI takes on a more active role in commerce. 

Agentic commerce raises the stakes 

Agentic commerce introduces another layer of complexity. As software agents begin searching, selecting and potentially purchasing on behalf of consumers or businesses, payment systems will increasingly need to process transactions initiated by machines rather than directly by people. That creates questions not only about how the transaction is authorised, but about the context surrounding it. 

What has the agent been permitted to buy? Which payment instrument should it use? Are there restrictions on value, merchant, geography or frequency? How should the transaction be routed if several methods are available? 

Payment infrastructure will need to interpret those permissions and execute the appropriate decision at machine speed. This is where fragmented architectures become particularly problematic, because many payment environments were built around relatively linear transaction journeys and fixed provider relationships. Agentic commerce may therefore expose infrastructure weaknesses that were less visible when humans remained directly involved in each purchasing decision. 

Better AI will not compensate for incomplete infrastructure 

There is evidence that greater orchestration maturity already correlates with better payment outcomes. In further PYMNTS Intelligence analysis of payment orchestration maturity, 78% of companies with all five core orchestration capabilities reported transaction-completion gains of at least 2%, compared with 10% of organisations with only three or four capabilities. 

That does not mean orchestration alone guarantees higher conversion, nor should findings from a relatively small US sample be treated as universal. The research does, however, illustrate an important principle: individual capabilities become more valuable when they work together and the same principle applies to AI. A sophisticated model sitting on top of fragmented infrastructure will always be limited by what that infrastructure allows it to control. Better prediction does not solve provider lock-in, better analytics do not make manual failover automatic and better recommendations do not make an inaccessible payment rail suddenly available. 

The organisations that gain most from AI in payments will be those that create an environment in which intelligence can be translated into action quickly, securely and within clearly defined controls. The future of payments will unquestionably become more intelligent. But intelligence alone will not fix fragmentation. The competitive advantage will belong to organisations whose infrastructure is capable of turning better decisions into better payment outcomes.