How to evaluate agentic AI payments for secure, reliable transactions

Zeb Winzenried, Senior Director, Testing Services at Applause

AI agents are moving beyond researching products and filling baskets to making purchases on customers’ behalf. Gartner predicts that, by 2028, AI agents will be involved in 90 per cent of B2B buying. Meanwhile, Bain & Company estimates that agentic commerce could account for 15 to 25 per cent of US ecommerce sales by 2030.

Forrester research, however, found that three-quarters of consumers are uncomfortable with agentic payments, even when they can set spending limits and rules in advance. Only 6 percent are comfortable allowing an agent to make payments autonomously. For agentic commerce to gain wider acceptance, consumers need confidence that they can authorise an agent, define its boundaries and maintain control of their finances.

This presents a new testing challenge for payment providers. They need to know which agents are authorised to make purchases, what customers have allowed them to do and whether they stay within those limits. A payment could be approved even when an agent selects the wrong product or acts beyond its authority. Testing must therefore cover the entire purchase flow, not simply whether the payment goes through.

Adapting payment infrastructure

Most authentication processes, including strong customer authentication (SCA) and 3D Secure, assume the customer is present to approve a payment. Providers need ways to authenticate the agent, connect its payment requests to the permissions the customer has set, and apply appropriate risk checks.

Zeb Winzenried

An agent may need to communicate with a merchant’s systems, a payment provider, a bank and other agents to complete a purchase. A problem at any stage could interrupt the transaction. Providers must check that these systems can exchange the information needed to complete a transaction and that they follow the customer’s instructions at every stage.

For example, an agent shouldn’t get around a spending limit by retrying a payment or substituting an item. Withdrawing permission should also stop an operation already underway. When a customer asks to cancel an order, the merchant needs to receive that instruction before the goods are dispatched. Purchases above the agreed threshold should still require the customer’s approval.

Why real-world testing matters

Testing an AI agent is more complicated than simply checking whether a conventional payment works. The same prompt can produce different results, making it difficult to compare an agent’s response against a single expected outcome. A purchase may also involve multiple tool calls, API connections and decisions. A failure at any stage can interrupt the transaction or result in an incorrect purchase.

Human evaluators can play a crucial role in assessing whether an agent handles requests safely, stays within its assigned role and completes tasks accurately. They should also be able to trace the agent’s decisions, identifying which tools it used, what information it relied on and whether each action matched the customer’s request. This helps teams understand why an agent might have made an incorrect purchase and where changes are needed.

For payment providers, this evaluation must extend to complete purchase journeys. They need to test these journeys with real users, in the markets they serve, using real bank accounts and payment instruments, including debit cards and digital wallets. This comprehensive testing allows teams to evaluate each of the payment stages an agent encounters and how it interacts with banking and merchant systems. It also helps ensure that the agent follows the customer’s instructions and delivers the intended outcome.

Laying the foundations for safer agentic payments

A good starting point is to create an agentic risk register. It should record how customers give consent, where evidence is held, how spending limits are enforced and how quickly an agent’s authority can be withdrawn. Teams should also identify who is responsible for authentication and disputed transactions, and what needs to be tested.

The register should inform ongoing testing and monitoring. A fault in an agent or one of the systems it relies on could affect thousands of purchases before anyone notices. Providers need to identify problems quickly, intervene before more transactions are affected and test whether agents can be manipulated into ignoring instructions or bypassing safeguards.

Ultimately, customers need confidence that agents will act within the limits they have set. Providers must demonstrate that purchases are properly authorised, that customers can withdraw permission and that failures are addressed promptly. Effective testing will help build that confidence while protecting customers, revenue and reputation.

spot_img
spot_img

Subscribe to our Newsletter