How AI Will Affect Payments in the Next 10 Years Research Summary
The next decade could change the way we buy, pay and borrow. For banks and payment providers, the biggest question may be who controls the customer's purchasing decision.
For decades, the payments industry has concentrated on making transactions faster, cheaper, safer and more convenient. We've moved from cash to cards, from physical terminals to digital wallets, and increasingly towards instant account-to-account payments.
Artificial intelligence is about to introduce another significant change, although I suspect its greatest impact will have relatively little to do with the actual movement of money.
The real change is in who makes the decision to transact.
Until now, payment systems have largely responded to decisions made by people. We decide what to buy, where to buy it and how to pay. The payment infrastructure then takes care of executing that decision.
AI agents are beginning to change this relationship. They can search for products, compare alternatives, negotiate prices and, increasingly, complete purchases within parameters established by their users.
Consider something as ordinary as booking a business trip. Today, I might compare flights, check hotel availability, consider my preferred airlines and loyalty programmes, and decide which card offers the best combination of rewards and foreign exchange costs.
An AI agent could potentially handle that entire process. Give it a budget, travel preferences and permission to transact, and it could make those decisions without requiring me to work through each step.
That sounds convenient. But consider what happens to the businesses that currently depend on influencing those decisions.
The airline, hotel, booking platform, bank, card issuer and payment provider may all find themselves competing for the attention of an AI agent rather than a human customer.
And that has some interesting commercial implications.
How significant could agentic commerce become?
The projections are substantial, although we should be careful about treating forecasts as inevitable outcomes.
Research by Celent for Tietoevry estimates that AI agent-initiated transactions could account for approximately 17.5% of e-commerce activity across five major European markets by 2035, representing around €191 billion in transaction value.
McKinsey has suggested that the broader global market for agentic commerce could reach between $3 trillion and $5 trillion by 2030.
These estimates measure different things, and neither should be confused with the value of payments revenue available to banks and payment providers. Nevertheless, they indicate the scale of activity that could eventually move through AI-assisted purchasing channels.
There are already signs that consumer behaviour is changing.
Research cited by Santander suggests that 39% of consumers are using AI tools for product discovery, with adoption exceeding 50% among Generation Z.
Discovering products through AI is, of course, very different from allowing an AI agent to spend money independently. There is a considerable gap between asking an assistant to recommend a television and authorising it to purchase one.
Closing that gap will depend on trust, security and the practical controls available to consumers.
But the direction is worth paying attention to.
What happens to cards and account-to-account payments?
One of the more interesting questions is whether agentic commerce will accelerate the shift away from traditional card payments.
Celent projects that account-to-account (A2A) payments could increase from approximately 24% of European e-commerce transaction volumes today to nearly 40% by 2035.
Cards are expected to continue growing in absolute transaction volumes, but their relative share could decline.
There are good reasons why A2A payments may become increasingly attractive in an AI-driven environment.
An AI agent doesn't necessarily need the familiar checkout experience designed for a human customer. It needs a reliable way to initiate an authorised payment, receive confirmation and complete the transaction.
Where instant payment infrastructure is available, an account-to-account transfer can potentially achieve this without relying on the traditional card acceptance process.
For merchants, lower acceptance costs and faster access to funds could make A2A particularly attractive.
We can already see how consumer payment behaviour changes when convenient alternatives become widely available.
Brazil's Pix is perhaps the clearest example. It has demonstrated how quickly an instant payment system can become part of everyday commerce when it combines accessibility, convenience and broad merchant acceptance.
Spain's Bizum and Portugal's MB WAY provide further examples of consumers becoming comfortable with account-based digital payments.
However, I wouldn't write off cards quite so quickly.
Card networks have spent decades developing capabilities that extend well beyond payment authorisation and settlement. Fraud protection, dispute management, consumer credit, international acceptance and established liability arrangements all have considerable value.
These capabilities become especially relevant when the person initiating a transaction is replaced by an autonomous software agent.
The question is whether those capabilities remain tied to the card payment itself or become services that can operate across different payment methods.
That distinction could have a significant influence on the competitive position of card networks over the next decade.
Credit could become detached from the payment method
Credit is another area where I expect some interesting developments.
Today, the relationship between payments and credit remains heavily influenced by the products through which credit is delivered.
Credit cards provide a familiar revolving credit facility. Buy Now, Pay Later (BNPL) providers offer financing at checkout. Banks provide overdrafts and other lending products linked to customer accounts.
But why should the availability of credit depend on which payment instrument a customer happens to use?
If an AI agent understands the customer's financial position, purchasing requirements and available financing options, it could potentially select the most appropriate credit facility independently of the payment method.
A customer might make an instant account-to-account payment while financing the purchase through a separate credit arrangement.
Celent estimates that credit delivered through A2A payment products could reach approximately €35.7 billion by 2035, equivalent to around 7.4% of the projected pay-later market.
For banks, this creates an opportunity to rethink how credit is offered.
Rather than relying entirely on traditional card products or standalone BNPL propositions, banks could make financing available through APIs that allow authorised agents to assess and select suitable credit options.
Naturally, affordability assessments, customer consent and regulatory obligations would still apply.
But the commercial opportunity is clear enough.
Banks with established lending capabilities could compete for financing opportunities across a much wider range of payment transactions.
The challenge will be making those capabilities accessible at the point where the purchasing decision is being made.
The real challenge is deciding who has authority to pay
There is, however, a fundamental issue that needs to be resolved before autonomous payments become commonplace.
How does a payment provider know that an AI agent has legitimate authority to initiate a transaction?
With traditional payments, the industry has developed numerous ways to establish that a transaction has been authorised.
We have PINs, passwords, biometrics, transaction authentication, tokenisation and increasingly sophisticated fraud detection systems.
But these mechanisms were largely designed around the assumption that a human customer is involved in making or approving the payment.
What happens when the customer has delegated that authority to software?
Imagine authorising an AI agent to purchase office equipment for your business, subject to a maximum transaction value of AED 5,000.
The agent identifies a suitable supplier and initiates a payment for AED 4,500.
How does the bank know that the agent is acting within its authority?
What happens if the agent selects the wrong supplier, purchases an unsuitable product or misunderstands the original instructions?
And who is responsible if the agent is manipulated into making a fraudulent payment?
These questions extend beyond conventional payment authentication.
They concern delegated authority, identity, consent and liability.
One approach being developed is the Agent Payments Protocol (AP2), which uses cryptographically verifiable mandates to establish what an agent has been authorised to do.
The underlying principle is sensible.
An agent should be able to demonstrate not simply that it has access to a payment instrument, but that it has permission to undertake a particular transaction within defined conditions.
Those conditions might include spending limits, approved merchants, product categories, expiry dates or requirements for additional customer approval.
The payment infrastructure would then verify that the proposed transaction complies with the mandate before executing it.
This is an important distinction because AI agents and payment systems operate in fundamentally different ways.
AI systems can interpret instructions, evaluate alternatives and make decisions based on probabilities.
Payment systems, by contrast, need precise instructions and predictable outcomes.
A payment either meets the conditions for execution or it doesn't.
Bringing these two environments together safely will require clearly defined controls between the agent making the decision and the infrastructure moving the money.
Fraud could become a much bigger problem
The payments industry has invested enormous resources in fraud prevention, and AI is already being used extensively to identify suspicious transactions.
Unfortunately, criminals have access to many of the same technological advances.
AI can be used to create convincing fraudulent communications, generate synthetic identities and automate increasingly sophisticated attacks.
The Global Anti-Scam Alliance estimated that consumers worldwide lost approximately $442 billion to scams in 2024.
Meanwhile, Juniper Research projects that the market for AI-powered fraud detection could reach $39.1 billion by 2030.
Agentic commerce introduces additional risks.
An AI agent might be manipulated through fraudulent product information, compromised websites or malicious instructions embedded in material it encounters while completing a legitimate task.
The agent could then initiate a transaction that appears technically authorised but was influenced by information the customer never intended it to follow.
Instant payments add another complication.
When funds move in seconds, there is very little time to identify and stop a fraudulent transaction before the money reaches the recipient.
This is particularly important for account-to-account payments, where the protections and recovery mechanisms available to customers may differ significantly from those associated with card transactions.
Fraud prevention will therefore need to operate at the speed of the transaction itself.
That means understanding the identity and authority of the agent, the context of the purchase and the behaviour surrounding the transaction.
It also means deciding when additional customer authentication is necessary.
An agent purchasing a routine subscription from an established supplier may require very little intervention.
The same agent attempting to transfer a substantial sum to an unfamiliar recipient should probably trigger additional checks.
Getting that balance right will be essential.
Too much friction and customers will question the benefit of using AI agents. Too little control and the risks could become unacceptable.
Where does this leave the banks?
I think banks face an interesting strategic choice.
They can continue concentrating primarily on the payment products customers use today, or they can start considering how their capabilities will be accessed when AI agents increasingly influence purchasing decisions.
That doesn't mean abandoning cards, digital wallets or existing banking channels.
Those products will remain relevant for a considerable time.
But it does mean recognising that the customer interface may no longer be the primary place where payment decisions are made.
Historically, banks and payment providers have invested heavily in keeping their products visible and accessible to customers.
Card branding, mobile banking applications, wallet positioning, loyalty programmes and merchant partnerships have all played a part.
In an agent-driven environment, some of that influence could move towards the platforms controlling the purchasing process.
An AI agent may select a payment method based on cost, speed, available credit, consumer protection or other criteria established by its user.
If a bank's products cannot be accessed and evaluated through those systems, they may simply be excluded from consideration.
This raises some practical questions for banks.
Can their payment products be accessed securely through APIs?
Can they support delegated authority and transaction-specific consent?
Can they offer credit independently of a particular payment instrument?
Can they provide fraud protection, identity verification and dispute management across different payment methods?
And perhaps most importantly, can they bring these capabilities together without being constrained by the traditional separation between cards, account payments and lending?
Many banks still manage these businesses through separate organisational structures and technology platforms.
That has been workable when customers interact with each product through established channels.
It becomes less practical when an AI agent expects to access different financial services as part of a single purchasing process.
There may also be commercial opportunities beyond providing the payment itself.
Banks could offer trusted identity services, delegated authorisation, transaction risk assessment, financing and payment routing capabilities to merchants and agent platforms.
Some of these services could become independent sources of revenue.
Others may simply become necessary capabilities for remaining competitive.
Either way, banks will need to decide where they want to participate in the transaction.
The competitive battle is already developing
The major payment networks and technology providers are clearly paying attention.
Visa, Mastercard and American Express are developing capabilities intended to support AI-initiated transactions, drawing on their existing networks, tokenisation infrastructure and experience in managing payment credentials.
Stripe and Adyen are positioning their platforms to support agent-driven commerce through APIs, payment orchestration and integrations with AI systems.
Technology companies are also developing the infrastructure through which AI agents will discover products, interact with merchants and initiate transactions.
For the established payment networks, the opportunity is to remain relevant as trusted infrastructure providers even when the customer no longer interacts directly with their traditional products.
For payment service providers, the opportunity is to become the infrastructure through which agents execute transactions across different payment methods.
And for banks, there is an opportunity to provide the financial services, trust and regulatory infrastructure that autonomous commerce will require.
But there is also a risk.
If the platforms controlling AI agents become the primary interface between customers and merchants, they could gain considerable influence over which financial services are selected.
We've seen elements of this before with digital wallets and online marketplaces.
Agentic commerce could take that separation between the customer relationship and the underlying payment provider considerably further.
What should banks be doing now?
I don't think banks need to make speculative investments based on predictions of how many transactions AI agents might initiate in 2030 or 2035.
The adoption forecasts are uncertain, and there are still unresolved questions around liability, regulation and consumer acceptance.
But several of the capabilities required for agentic commerce already make commercial sense.
Modernising payment architecture, improving API connectivity, supporting instant payments and bringing card and account-based payment capabilities closer together are worthwhile investments regardless of how quickly AI adoption develops.
The same applies to improving digital identity, transaction authentication and fraud prevention.
Where I would encourage banks to think more carefully is in how these capabilities are organised and made available.
A bank that can provide payments, credit, identity verification and transaction controls through a coherent set of services will be better positioned to participate in emerging agent-driven commerce models.
A bank that continues to treat each of these as an isolated product may find integration considerably more difficult.
There is also a question of commercial ownership.
If an AI platform controls the purchasing decision, who owns the customer relationship?
Who determines which payment method is used?
Who receives the transaction data?
And who earns the revenue from providing the financial services that make the transaction possible?
Those are questions worth examining now, even if the eventual answers remain uncertain.
Looking towards 2035
Predicting the payments industry ten years ahead is never straightforward.
We've seen technologies that were expected to transform payments struggle to achieve meaningful adoption. We've also seen developments such as Pix demonstrate how quickly established behaviour can change when the right combination of infrastructure, regulation and customer convenience comes together.
AI agents will almost certainly become more capable over the coming decade.
Whether consumers will trust them with substantial financial decisions, and how quickly that trust develops, remains to be seen.
My expectation is that adoption will be gradual.
Initially, AI agents will assist with research and product selection while customers retain control over payment authorisation.
As confidence improves, consumers and businesses may become comfortable delegating routine purchases within clearly defined limits.
More autonomous transactions could follow once the necessary controls, liability arrangements and regulatory frameworks are sufficiently established.
The pace will vary across markets and use cases.
But for the payments industry, the more interesting question is already becoming apparent.
If the customer no longer chooses the payment method directly, what determines which provider gets the transaction?
For banks, card networks and payment providers, understanding the answer may be considerably more important than predicting exactly how many transactions AI agents will initiate.
Sources and research
- Tietoevry / Celent (2025), research on the future of European payments and agentic commerce.
- Accenture (2026), analysis of agentic commerce and payments.
- International Monetary Fund (April 2026), research on agentic AI and payment infrastructure.
- Santander (2026), research and commentary on the future of agentic commerce.
- Capgemini Research Institute, World Payments Report 2025.
- Juniper Research, projections for AI-powered fraud detection through 2030.
- Global Anti-Scam Alliance (2025), research on global scam losses.
- McKinsey & Company, estimates of the potential global agentic commerce market.
Note: Market forecasts represent estimates based on the assumptions and methodologies of the respective research organisations. Actual adoption and transaction volumes will depend on technology development, consumer acceptance, regulation and market conditions.