AI agents in banking and what they mean for your merchants and their customers' payments

• Agentic AI is increasingly being used to analyze payments and execute them - selecting routes and completing transactions in real time, with risk controls applied as part of the flow.
• Consumer trust remains a major constraint, shaped by concerns around control, consent, and whether automated decisions can be overridden.
• Traditional payment stacks struggle under this model, with rigid routing and tightly coupled systems creating fragmentation across channels.
• Orchestration becomes essential, giving agents a central layer to operate through while maintaining control and auditability across the payment flow.

Don't have time to read more now? Sign up to our newsletter to get the latest insights directly in your inbox. 

...The one that answers balance queries, helps you find a statement or reset your password, and bounces you to a human when it can't cope.

That kind of AI is useful, but it’s fundamentally reactive - it waits for a question before responding and is stuck within preset limits. It doesn’t move beyond the initial prompt and doesn’t really follow up with anything meaningful.

However, a new class of AI is emerging: agentic systems built not just to respond, but to act.

These are systems designed to pursue a defined outcome,, selecting tools, linking decisions, and executing actions until the outcome is reached or until a human intervenes.

Execution is the critical difference. This is no longer old-school AI advising on payments, but agentic AI actively selecting routes, applying real-time risk controls, executing transactions, and recording decisions in real time - all within the same moment of interaction.

Even so, most people haven’t quite caught on to how much this is changing. In our survey of more than 3000 people across the UK and US, just 12% in each market said they were “very familiar” with the idea of AI handling payments automatically.

So while the technology is advancing quickly inside payment systems, public awareness is still playing catch-up.

That’s why it’s worth taking a look at how agentic AI is reshaping payment flows, what that means for merchants and their customers, and why more flexible, orchestrated infrastructure is becoming essential. And that’s exactly what we’ll be exploring today...

Despite low awareness, agentic AI is already operating inside banks today, with its impact most visible in decisioning, customer experience, payments, and the day-to-day role of teams.

Here’s how payments typically work today: a customer selects a product, authenticates, confirms the transaction, and then waits for it to complete. Each step introduces a potential point of drop-off, while each routing decision was written into a configuration file weeks ago by a human.

The change becomes clearer when you look at card-present and card-not-present payments side by side…

Card-present

In card-present environments, this capability moves directly into the terminal or device:

• Agents embedded in POS systems can pre-fetch tokenized credentials, evaluate device health, and negotiate routing before the card even taps.
• Risk scores are generated in milliseconds and immediately trigger approvals or step-up prompts.
• When connectivity drops, agents keep transactions moving in offline mode, capping limits, queuing verification data, and reconciling automatically once the connection is restored.
• Multiple agents can run in parallel across a single transaction, covering device health, PCI scope, KYC checks for high-value sales, and routing across issuers, acquirers, A2A rails, and wallets.
• That same decisioning logic can then be applied consistently across devices, locations, and form factors, reducing variation between individual terminals and the environments they sit in.

Card-not-present

Online, the change is even more pronounced:

• Agents evaluate cart risk, device signals, chargeback history, and issuer policies simultaneously, building a more complete view of each transaction as it happens.
• Authentication becomes adaptive, with 3D Secure applied in real time based on approval likelihood and risk.
• When payments fail, AI agents automatically reattempt them through alternative routes or updated credentials, keeping the journey moving rather than bringing it to a stop.
• Recurring and installment arrangements, including renewals, retries, and proration, are managed continuously in the background.
• Every step is logged, from who initiated the charge through to the rules applied and how tokens moved.

The result is a payment flow that moves faster, with improved accuracy and clearer explanations than anything achievable with static rules. But it raises an uncomfortable question...

When software is allowed to execute payment decisions, who is accountable for each action it takes?

Our survey data gives a clear signal on where consumers expect the answer to land:

• Loss of control was the top concern (59% US, 49% UK)
• Not being able to stop or override decisions followed closely (54% US, 50% UK)
• Security concerns remain high, with nearly half worried about systems being hacked or manipulated
• Payments happening without clear consent is a major concern across both markets
• Only 5% said they had no major concerns 

Skepticism here is close to universal, and that carries implications for both how banks deploy agents and for the governance infrastructure they need to build around them.

If agents are acting, banks need to be able to prove what each one did, why it did it, the data it relied on, and that a human could have stopped it at each point.

That is where the challenge moves beyond policy, into infrastructure.

Here's the problem: most traditional payment stacks weren't built for this.

Many still rely on fixed configurations, which makes it harder to respond in real time. Routing is often set in advance, switching between providers isn’t always possible mid-transaction, and different systems don’t always share the same rules across card-present, card-not-present, and account-to-account payments. As a result, agents struggle to make consistent decisions across the full flow.

The deeper issue is coupling. In most legacy setups, the checkout, processor, fraud tools, and routing logic are tightly bound together, so changing one part can mean reworking the rest. An agentic environment needs the opposite: each component independently addressable, so agents can call the right tool at the right moment without the whole stack moving with it.

This becomes more visible as systems expand over time.

As banks and merchant networks scale, the stack often becomes more fragmented. The more disconnected the infrastructure, the less room agents have to operate effectively, and the harder it becomes to trace what actually happened across a transaction.

The answer isn’t to replace what’s already there: in most cases, the individual parts of the payment stack work, they just don’t work together in a coordinated way. 

That gap is where complexity and cost build up, and it's what Aevi’s orchestration layer is built to close.

Orchestration sits between the agent and the underlying payment infrastructure. It decouples the checkout device from the processor, so merchants can switch acquirers without changing hardware, and banks can route across cards, wallets, local payment methods, and A2A rails from a single set of consistent rules.

With that logic sitting centrally, agents can make routing decisions in real time based on factors like cost, fraud risk signals, approval rates, or network availability, all without hard-coded dependencies.

Aevi’s orchestration platform is central to making agentic AI in banking manageable at scale. It provides the structure that sits between agents and the underlying payment infrastructure, enforcing scope, logging every action, and applying consistent rules across channels so agents can operate safely within defined boundaries.

This is where our consumer data gets really instructive.

• Around 45% expect AI to manage everyday payments within the next decade
• But only 20% (UK) and 32% (US) are comfortable letting it choose how they pay
• Over 75% of over‑65s actively oppose AI making payment decisions
• Fewer than 1 in 4 trust any organization to use AI to manage their money

Taken together it’s clear that AI is expected to play a bigger role, but not to be left to operate without oversight.

That's also how the most effective agentic AI in banking deployments actually work. Agents handle the volume (routine routing, retry logic, fraud triage, compliance checks), while humans handle the exceptions (disputed merchant history, cross-jurisdictional conflicts, vulnerable customer cases). The agent surfaces the case; the human makes the call.

Agentic AI in banking is already moving into live environments across fraud, underwriting, customer experience, and the payment flows that connect them.

The challenge now is infrastructure. Payment stacks need to keep pace with systems that can make and execute decisions in real time, while still remaining controlled and observable across channels.

And that all comes down to design. The stacks that hold up are built for flexibility and governed for accountability, with orchestration that lets agents do their job without breaking the payment experience for the people on the other end of the transaction.

If you're starting to think about how agentic systems could operate safely within your own payment stack, now is the right time to have that conversation. Talk to our team to explore what orchestration makes possible!