In short (TL;DR)

Not today on its own. EU payment services rules (PSD2) and strong customer authentication require that payments be authorized by a human. For an AI agent to initiate and approve payments itself, There is currently no mechanism in European law. A legal and reasonable division of labor already exists today: AI prepares, compares, and verifies payments, while humans review and approve. Card networks are simultaneously building protocols that could change this picture in the coming years.

What does the law say today?

Authorizing a payment is a personal process in the EU. PSD2 and Strong Customer Authentication (SCA) rules require that the payment order be confirmed by the payer: a person who authenticates themselves with at least two factors. Taylor Wessing's analysis summarizes the situation: there is no special regime for agent-based payment models, the same rules apply as for everyone.

In practice, this means three things:

  • An AI agent cannot be a „payer” – that role belongs to a human or a process clearly authorized by them;
  • Standing authorizations given to the bank (for example, direct debit) are permitted because they are backed by a person's given and limited mandate;
  • If someone gives their bank passwords to an AI, it doesn't make the payment a legal authorization, but it does violate the bank's terms and conditions and leaves them liable for any damages.

What are card networks and technology companies building?

The interest is huge because the stakes are high: McKinsey estimates that agent commerce could affect $3-5 trillion worth of transactions by 2030. Google announced the AP2 agent payments protocol in September 2025 with more than 60 partners (including Mastercard, PayPal, Adyen, and American Express), and Visa is developing its Intelligent Commerce program. Their shared idea: to give the agent a cryptographically verifiable, human-bounded mandate that each party can verify.

IMF 2026 Analysis honestly addresses unresolved issues: who is responsible if an agent makes a mistake; how to link an agent's actions to the authority given by a person; what happens when thousands of agents behave in the same way at the same time. Until these answers are available, the person remains a validator.

What division of labor is already in effect today?

What the market calls the „man in the chain” model is simple and legal in a small business:

  • AI reads incoming invoices, checks details and compares with orders;
  • AI compiles a payday list and highlights anomalies: new bank account, double bill, unusual amount;
  • the person reviews the list and confirms the payments at the bank themselves;
  • Every step leaves an audit trail, so you can later see who decided what.

This way you get the benefits of automation without bending authorization rules. We describe how such a workflow looks in accounting in a broader sense in the article agentic financial management and on the invoices side in the article AI accounting.

Checklist before trusting AI with your finances

  1. Never share bank passwords or PIN calculators with any tool.
  2. Keep approval in human hands: AI proposal, Your signature.
  3. Set amount limits and a double approval threshold, for example, two approvers starting at 1,000 euros.
  4. Require an audit trail: what the agent did, based on what, and when.
  5. Always check bank account changes in a separate channel: this is also invoice fraud the most effective step against.

Summary

An AI agent cannot make payments independently in the EU today, as PSD2 requires a human for authorization and there is no special regime for agents. It is permissible and beneficial to let an AI prepare payments and have a human approve them: this model works, keeps accountability clear, and prepares you for the time when protocols like AP2 will make authorized agent payments possible in Europe as well. This article is a general overview, not legal advice.

Try Bilnex's free invoice processing - AI reads and prepares, you decide.

FAQ

Can an AI agent initiate payments in the EU?

Not independently: PSD2 and strong customer authentication rules require that a payment be authorized by a human, and there is currently no mechanism in European law for an AI agent as a payer. AI can prepare payments, but a human must confirm.

What is agent trade?

A model where an AI agent searches for products, compares offers, creates an order, and initiates payment within the limits set by a human. McKinsey estimates that this could impact $3-5 trillion worth of transactions by 2030; Google's AP2 protocol and Visa Intelligent Commerce are the first major attempts to standardize it.

Is a direct debit or standing order an „AI payment”?

No. Behind direct debit and standing order is a human-given, limited, and revocable mandate registered by the bank. This is exactly the kind of clearly authorized mandate that agent payment protocols are trying to create for AI.

How to use AI in payments today?

Let AI read incoming invoices, check details, compare with orders, and create a payment day list with any discrepancies highlighted. A human reviews the list and confirms the payments at the bank — this way, the authorization remains legal and the audit trail intact.