Regulation
Agent payments under the EU AI Act
What the AI Act's oversight, logging and risk-management duties mean in practice when your agents can move money.
The EU AI Act doesn't mention "agent payments" anywhere. But several of its obligations land squarely on any business that lets an AI system act on money. This isn't legal advice — talk to your counsel — but here's how we map the main duties to day-to-day controls.
Human oversight
The Act expects people to be able to understand, monitor and override AI systems where the stakes are high. For payments, that translates into approval thresholds, the ability to pause an agent instantly, and approvers who see enough context to make a real decision.
Record-keeping
Systems need to log events automatically so their behaviour can be traced. A payment record should capture the request, the rules applied, the risk assessment, any human decision and the outcome — with timestamps and in a form that can't be quietly edited.
Risk management
You're expected to identify foreseeable risks and put measures in place. For agents that spend, the obvious ones are:
- Overspending or duplicate payments
- Paying the wrong or a fraudulent merchant
- Prompt injection that tricks an agent into making a payment
- Loss of the agent's credentials
Transparency to the people affected
If an agent pays a supplier or refunds a customer, think about how that party can find out an AI system was involved and who to contact if something went wrong.
Where to start
Most teams already have the raw material: a payments processor, a finance approval process and some logging. The gap is usually connecting them per agent. Give every agent an identity, run every payment through a policy, and keep one record that ties it all together.