AI agent payments — agents that pay invoices, settle expenses or buy on someone's behalf — moved from concept to infrastructure in 2025. Visa launched Intelligent Commerce and Mastercard launched Agent Pay in April 2025, and Google introduced the Agent Payments Protocol (AP2) that September. Payment networks are building ways to identify agents and tie their transactions to real users. Inside your company, you still need to decide what your agents may pay, to whom and when.

Two kinds of AI agent payments

  • Business payments — an accounts payable agent pays supplier invoices, or a procurement agent places orders.
  • Agentic commerce — an agent buys goods or services on a customer's behalf, using credentials issued for that purpose.

Network-level programmes help with the second: verified agents, tokenised credentials and spending controls set by the user. Company-level controls matter for both — especially the first, where the money is yours.

The risks

  • Paying the wrong party — a fraudulent invoice, or changed bank details.
  • Paying the wrong amount — duplicates, wrong currency, wrong units.
  • Exceeding authority — amounts beyond what anyone agreed the agent could handle.
  • Manipulation — prompt injection in an invoice or email that steers the agent.
  • No trail — nobody can show who authorised a payment.

Controls for AI agent payments

policy "supplier-payments" {
  action = "pay_invoice"

  deny             when vendor not in approved_vendors
  deny             when vendor.bank_details_changed_within_days < 30
  allow            when amount <= daily_limit(vendor)
  require_approval when amount > daily_limit(vendor)  approvers = 2
}

Beyond the rule itself:

  1. Approved payees only. The agent can pay suppliers on a maintained list — never add new ones.
  2. Bank detail changes are a hard stop. Changes go through your existing verification, not the agent.
  3. Limits per payee, currency and day. Totals, not just individual amounts.
  4. Two-person approval above a threshold. Rare, but required. See AI agent approval workflow.
  5. Separation of duties. The agent that matches invoices shouldn't also approve exceptions.
  6. Complete records. Every payment attempt with decision, rule version and approvers in an audit trail.

Rolling out a payments agent

Start with a narrow supplier set and low limits. Dry-run the rules against several months of past invoices — including month-end — before going live. Review held and denied payments weekly with finance, and expand coverage only when the data supports it.

Where Gaten fits

Gaten checks every payment action your agents attempt against your rules — approved payees, limits, two-person approval — and records each decision for finance and audit. It complements the controls payment networks provide. See sample rules.

Frequently asked questions

Can AI agents make payments?

Yes. Payment networks now offer programmes for agent-initiated payments, and companies use agents for accounts payable. Strict limits and approvals are essential.

What is agentic commerce?

Purchases made by AI agents on a user's behalf, using credentials and controls designed for agents, such as Visa Intelligent Commerce, Mastercard Agent Pay or Google's AP2 protocol.

What controls should an AI accounts payable agent have?

Approved payees only, no bank-detail changes, limits per payee and day, two-person approval above a threshold, and a complete record of every payment decision.