AI Agent Payments in Business: Fraud, Rules and Control
VirtuaBroker explains at MERGE Madrid 2026 how to let AI agents pay supplier invoices without nasty surprises: changed bank accounts, hidden instructions, duplicate payments, spending limits and human approval
20min · Full recording from 29/10/2026 at MADRID Stage. Also available on YouTube.
AI agent payments in business: fraud, rules and control
Overview
What happens if an AI agent pays on behalf of your company while you sleep and something goes wrong? In this MERGE Madrid 2026 talk, VirtuaBroker, a Spanish international payments fintech, explains how to prepare a company so that agents can pay supplier invoices automatically. Using a fictional case of payments to suppliers in several countries, it walks through the most common failures, from fraud through a changed bank account to duplicate payments, and the controls it recommends before automating.
What you'll learn
- From instructing to delegating: what changes when an agent pays instead of a member of the team
- Fraud: why scams aimed at employees will target agents too
- Changed bank account: the control that stops you paying an impostor
- Hidden instructions: why a document must be treated as data and never as an order
- Technical failures: retries that duplicate payments and agents sharing a balance
- Getting started: budget, thresholds, daily report and human review
Session summary
From instructing to delegating: VirtuaBroker compares agent payments with autonomous driving, which has arrived in stages. Until now, management trusted its finance team and asked questions if something looked odd. With an agent, that trust has to be replaced by rules, all the more so because payments today are executed in seconds and a mistake is hard to undo.
Fraud moves to agents: the talk recalls the classic scam in which someone poses as a senior executive and asks an employee for an urgent payment, which works because nobody dares to question the order. The company argues that agents will be the next target.
Rules that seem sufficient: the fictional case starts with four rules: pay only approved invoices, only to known suppliers, keep a minimum balance and ask for human approval above a certain amount. According to VirtuaBroker, they are not enough.
Four failures and their controls: the first is an invoice below the threshold whose bank account has changed, sent from an email that looks legitimate. The control is that any change to already verified data requires human review. The second is hidden instructions inside the invoice, such as invisible text ordering payment without checks. The rule is that documents are data and never instructions. The third is technical: if the bank is slow to respond and the agent retries, the payment can go out twice, so every payment must be unique. The fourth is several agents working in parallel on the same balance without coordinating and spending more than planned, which calls for reserving the funds.
How to set it up: the recommendation is for the agent to know suppliers and amounts, but for a separate system to verify the rules and route anything unexpected to a person. It also proposes a report every morning with what was paid and what was held, a maximum budget per period even if the account holds far more, and starting with strict rules that are relaxed as everything proves to work. Its advice is not to automate everything from day one.
VirtuaBroker's role: the company places itself as the last piece, execution. It lets agents connect to its system, adds its own controls and offers international payments and collections through local payment systems in different countries, which it says are faster and cheaper than traditional banking.
Watch the full talk
Watch the full recording on MERGE's YouTube channel, with VirtuaBroker on how to automate supplier payments with AI agents without losing control.
FAQs
Can AI agents pay a company's invoices?
According to VirtuaBroker, it is already possible, provided what the agent can and cannot do is clearly defined.
What are the risks?
Those the talk highlights are fraud through a changed bank account, hidden instructions in documents, duplicate payments caused by retries and several agents spending the same balance.
Which controls are recommended?
Human review for any change to verified data, treating documents as data, unique payments, reserving funds, a maximum budget and a daily report.
Should everything be automated from the start?
No, according to the company. It recommends starting with strict rules and human oversight and relaxing them gradually.