Why agentic payments need guardrails

What are agentic payments?

According to Ghali Bennani Laafiret, CEO and co-founder of Ralio, Agentic payments are payments initiated by non-deterministic AI agents, not humans or pre-configured software. Think: Claude processing your invoices and paying them automatically, or an AI procurement agent ordering supplies when inventory runs low.

This is an important shift at a trust and execution level. We moved from humans instructing payments to software executing payments via API (deterministic, rule-based) and now we're entering stage three: AI agents initiating payments autonomously.

The three categories of agentic payments

Ghali breaks down agentic payments into three distinct categories:

1. Consumer payments: Agent e-commerce where AI browses websites and completes purchases on your behalf
2. Agent-to-agent payments: Microtransactions where agents pay for tools, content, or services they need to complete tasks
3. Business payments: AI agents processing invoices, managing payroll, and handling treasury operations

The 5x payment problem

On Griffin Uncut, Ghali shares a real example with us - an AI agent tried to process an invoice payment. The payment failed. The agent, trying to be helpful, retried it. No confirmation came back. It tried it again. And again. And again.The agent ended up making the payment five times.

As it never got anything back that confirmed that the payment went through, the natural next step was to keep trying. This is a clear example of what can happen when you give AI agents access to payment infrastructure without proper guardrails.

Traditional payment systems assume deterministic behavior. Software follows rules: if X, then Y. You can predict, test, and validate every possible outcome. AI agents don't work that way. They're trained on human behavior patterns, but they don't always act like humans. They have personalities‍—‌some are overly helpful, others are cautious. Ask the same agent the same question 100 times and you'll get mostly the same answer, but not always.

"We don't solve for humans. We solve for agents," Ghali says.

What a trust layer actually does

A trust layer sits between agent builders and payment providers, abstracting complexity on both sides. The trust layer provides:

• Identity and authorisation: Which agents are permitted to make payments and under what conditions
• Spend limits and guardrails: Hard-coded rules that can't be overridden by an agent’s creativity
• Audit trails: Complete visibility into why an agent made each payment decision
• Anomaly detection: Catching the 5x retry problem before it becomes a £50k mistake
• Deterministic execution: When an agent decides to pay, the actual payment execution is rule-based, not based on probability

The trust layer helps you make a non-deterministic payment deterministic. AI can help build workflows and make decisions, but anything involving actual money movement should be deterministic.

Testing the trust layer

Ghali says Ralio runs its own payroll and invoice payments through its platform to test it. They've encountered edge cases like agents converting GBP to pence and back, accidentally multiplying amounts by 100, amongst others. In test environments, these are fixable bugs. In real life, it could be catastrophic.

Why this matters now

Most B2B payments still run manually. Some use software. Almost none use agents. But that's changing fast. As AI adoption grows, businesses will leapfrog from manual processes directly to autonomous agents, skipping traditional automation entirely. The window for building proper infrastructure and guardrails is now, before agentic payments go mainstream. Watch the full Griffin Uncut episode with Ghali Bennani Laafiret on building the trust layer for agentic payments, why timing matters when building for the future, and the three seasons of agentic payment evolution.

Griffin on Youtube: youtube.com/@griffinbank

More information on Griffin’s MCP server here: docs.griffin.com/docs/guides/griffin-mcp-server

More information on Ralio: ralio.co