Agentic Payments – When AI agents take over payments

A car that pays for itself. A voice assistant that fills shopping baskets, recognises discounts and pays bills. A digital CFO that analyses expenses, manages liquidity and optimises payment dates – without human intervention. What sounded like science fiction just a few years ago is now becoming reality. The rise of so-called agentic payments marks the beginning of a new era in payment transactions: payments that think for themselves.

This development is more than just a technological gimmick. It represents the transition from rule-based automation to true autonomy – a change that is reshaping the foundations of finance. Whereas previous systems only execute commands, agents act independently within defined limits. They recognise patterns, analyse contextual information and make decisions that were traditionally reserved for humans. Humans only specify the goal – such as a budget, a policy or a framework – and leave the operational implementation to a digital counterpart.

From humans to machines

The basic principle of agentic payments is delegation: transactions are no longer initiated by humans, but by a digital agent that has been given a mandate in advance. This mandate defines what is permitted, what the limits are and in what context decisions may be made.

Such agents can not only initiate payments, but also control entire value chains. They recognise price changes, automatically negotiate contracts or search for the best payment method from various providers. The key difference to traditional automation lies in the ability to make decisions based on situational assessment – i.e. not on ‘if-then rules’, but on probabilities, empirical values and target optimisation.

Technologically, the interaction is complex. Open APIs enable access to accounts, contracts or accounting systems. Machine learning models process the data, recognise patterns and generate decisions. The transaction itself then takes place via existing payment infrastructures such as SEPA Instant, card or wallet networks, and in some cases also via blockchain-based smart contracts. This creates a multi-layered system in which intelligence, infrastructure and trust are closely intertwined.

First live scenarios and the new role of platforms

What was still being discussed in theory in 2023 is now productive. PayPal, Stripe and Mastercard have introduced the first live integrations in the US, where payments are initiated directly from AI systems. Users can make purchases in ChatGPT without visiting a website or manually entering payment details. The AI agent compares prices, selects providers and then initiates the transaction.

Technically, this is done using new protocols: Google developed AP2 (Agent Payment Protocol), an open standard, while Stripe and OpenAI are focusing on interoperability between AI systems and payment networks with the Agentic Commerce Protocol (ACP). Anthropic, in turn, has established the Model Context Protocol (MCP), which standardises the exchange of data and intentions between agents. This infrastructure allows payments to be processed securely and transparently directly from the context of an AI application.

At the same time, Visa and Mastercard are experimenting with their own frameworks, such as the Trusted Agent Protocol, which regulates the authentication and liability of machine actors. This is shifting the centre of power in payment transactions away from apps and interfaces towards intelligent interfaces where transactions become invisible – but take place everywhere.

From automation to autonomy

The difference between an automated process and an autonomous agent is subtle but crucial. Automation means that systems execute predefined commands – such as a standing order or a monthly salary payment. Autonomy, on the other hand, means that the system independently decides whether, when and how a payment makes sense.

An example: an agent recognises that a supplier's price has fallen by five per cent. It checks whether the contract can be adjusted, automatically negotiates new terms, triggers a modified order and makes the payment – all within defined parameters. The human receives only a confirmation and an audit-proof log.

In Industry 4.0, such mechanisms are already leading to automated value streams. Sensors in production facilities report wear and tear, trigger orders and initiate payments as soon as goods are delivered or services are rendered. The payment process thus becomes invisible, but an integral part of operational business.

Trust through ‘Know Your Agent’

The more autonomous payments become, the more important trust becomes. It is no longer enough to identify the person initiating a payment – you also need to know the identity of the machine executing it.

The new KYA (Know Your Agent) principle therefore adds a machine dimension to the classic KYC. Each AI agent is given a unique digital identity that is cryptographically secured. Tokenisation replaces sensitive payment data, audit trails document every decision, and governance dashboards allow mandates, limits and authorisations to be managed centrally.

Visa, Mastercard and identity providers such as Trulioo are working on certification standards that ensure that only authorised agents can access payment networks. This creates a multi-layered trust system that combines technological autonomy with regulatory traceability.

Regulation between innovation and risk

Legally, agentic payments are breaking new ground. In the European Union, several pieces of legislation overlap: PSD3 and Payment Services Regulation, the AI Act and the GDPR. They define the scope within which AI agents are allowed to operate.

The open question of liability remains central. If an agent makes a mistake – such as an unauthorised payment or an incorrect assessment – the question arises as to who is responsible: the user who commissioned the agent, the provider of the payment interface or the developer of the AI model? The European Commission is currently discussing whether AI agents should be classified as ‘supervised digital agents,’ which would create a new category between humans and machines.

In addition, there are data protection requirements. Agents process personal and financial data and must remain GDPR-compliant. ‘Privacy by design’, real-time auditability and the obligation to provide explainable decisions are becoming compliance standards. BaFin and the EBA emphasise that despite automation, humans remain responsible – the ‘human in the loop’ principle should also apply to AI-based payments.

Economic potential and new risks

The economic impact could be enormous. According to studies by BCG and McKinsey, the global market for agent-based payments could reach a volume of up to two trillion US dollars by 2030. More than half of global e-commerce could then be handled by autonomous agents.

For companies, this means greater efficiency, fewer points of friction and completely new business models. Banks could offer agent-as-a-service to ensure regulatory security. Merchants could fully integrate payments into AI-based shopping experiences.

But the more interconnected the systems are, the greater the risks. New fraud vectors are emerging – from manipulative prompts to deepfake agents. There is also the threat of a concentration of power among a few large tech providers who control protocols, interfaces and identities. For European banks, this raises the strategic question of whether they want to be active architects of this ecosystem – or mere infrastructure suppliers.

Industries in transition

For the banking sector, the issue is doubly relevant: technically and strategically. Banks are increasingly losing direct contact with end customers because AI agents are taking over the interaction. At the same time, they can reposition themselves as a trust layer – i.e. as providers of liability, auditability and regulatory security.

Companies also benefit. In the B2B sector, complex processes such as ‘procure-to-pay’ or ‘order-to-cash’ can be fully automated by agents. They identify open items, prioritise payments and take advantage of discounts in real time. For treasury departments, this opens up the possibility of granular liquidity management. Insurers, in turn, are testing models in which AI agents check claims, evaluate image data and trigger payments immediately – in seconds instead of days.

Challenges and the way forward

Despite all the progress made, key challenges remain. Technical integration is complex, and many core banking systems are not yet sufficiently API-enabled. There is also a lack of a common European regulatory framework that sets binding rules on liability, oversight and transparency. In addition, there is the ethical question: how much control are people willing to relinquish? Trust is not created by technology alone, but by transparency. Users must understand why an agent has made a decision and have the option to revoke it.

For companies, banks and payment providers, this means that the introduction of agentic payments is less of an IT project and more of a governance task. It requires new forms of oversight, auditing and responsibility.

Conclusion: Trust is the new currency

Agentic payments are not a distant future trend, but the logical continuation of open banking, instant payments and embedded finance. They make payments invisible, but at the same time omnipresent. Humans are not disappearing from the process – they are moving to a higher level: from executors to designers.

Those who start opening up API architectures, ensuring data quality and establishing governance structures today will set the direction tomorrow. Because the question is no longer whether machines should be allowed to pay – but who we trust to do it right.