Agentic commerce will transform shopping, but not before difficult challenges are solved. One problem we see all the time is that an AI agent will make an exquisitely personalized recommendation, and the buyer’s intent will be strong. Yet, the transaction still fails.
The AI doesn’t choose wrong. Rather, the payment step required something the agent couldn’t provide.
That failure is a structural flaw. And it reveals that two separate engineering problems are being conflated: building an agent that finds the right product and building a system that can reliably complete the purchase.
The broader pattern is consistent: fragmented tooling leads to execution failures. Payment fragmentation is a specific instance of that problem, and one of the costliest, because it sits at the end of the conversion flow, directly at the moment of transaction.
The solution requires enterprise leaders to decide which of those problems they are solving because the payment infrastructure required for each is different, and choosing the wrong one late in the build is expensive to fix.
BNPL Wasn’t Designed for Agents
Here’s the problem: The most commonly used payment model at checkout today is BNPL (Buy Now, Pay Later). In a traditional BNPL transaction, a consumer applies for short-term credit at the point of purchase, receives a real-time approval decision, and completes the transaction through a provider like Affirm. The model was designed for a specific context: a human being, present at a screen, with a few seconds to wait. If something goes wrong, he or she can respond by filling out a form, accepting the terms, and being redirected to a third-party flow.
Agents cannot do any of those things. There is no human in the loop to handle an exception. The flow either completes or it does not. In a human checkout, that means six in ten consumers find another way to pay. In an agentic flow, it means six in ten transactions fail at the payment step — silently, without recovery.
Businesses deploying agents are discovering this limitation in production, where the cost of rearchitecting a payment integration is significantly higher than modeling it correctly from the start.
Certainty is Essential
The answer is to devise an agentic commerce system that knows, before initiating a transaction, whether it will complete. Not probably. Not usually. With enough reliability to design around.
Card-linked installments, which operate on existing card rails, are one emerging solution. They help prevent payment failures because the credit decision is made at the time of card issuance. The agent is drawing on available credit, not triggering a real-time lending decision. Authorization rates above 85% mean the system can be designed with high confidence in completion.
Just as important, card-linked models self-select for consumers with established credit and available capacity — the same consumers who index toward higher-AOV purchases and repeat buying behavior. Card-linked BNPL is a high-quality, low-risk growth engine.
Involve Legal and Compliance
Most discussions about how AI agents will handle payments are happening in product and engineering teams, but legal and compliance teams need to be involved as well. The type of payment system selected for an AI agent decides which laws and rules apply, which customer data is collected, and who is responsible if something goes wrong when an AI makes a purchase without human intervention.
With card-linked payments, the rules are already known and limited because they’re built on the existing credit card network. The compliance team doesn’t have to learn new rules or take on new responsibilities. They just have to manage a payment type they already know. That’s what makes it possible to use at the huge scale that AI systems are built for.
So, the real question for a CIO or General Counsel isn’t “Which payment model looks best in a demo?” It’s: “Which payment model can we actually control and follow the rules on when an AI is making purchases at machine speed?”
Three Questions Before You Build
Before finalizing the payment layer in an agentic commerce deployment, three questions determine whether the infrastructure choice will hold.
1. What is the authorization failure rate of your intended payment model under agent-initiated conditions, and what does a failed transaction cost in a flow with no human fallback? A high failure rate at checkout can kill conversion and revenue, because in an agentic flow, there’s no human to change payment methods, so a failed transaction becomes a lost sale and a negative customer experience.
2. Does your payment infrastructure require any step — application, approval, redirect, consent — that your agent cannot execute autonomously? If a human is needed, the purchase will stall at the last mile, derailing a high-intent sale.
3. If your enterprise attempts to automate BNPL credit flows on behalf of consumers, which regulatory frameworks apply, who owns that exposure, and does your compliance function know it is happening? Automating BNPL without understanding the regulatory environment can open the company to legal liability and reputational damage, especially if compliance teams don’t know that AI agents are making credit decisions on behalf of customers.
There are no universal answers. A company with its own credit card, direct credit relationships, underwriting data, or even its own BNPL or installment payment options faces different constraints than a specialty retailer that lacks a credit infrastructure and relies more on third-party BNPL providers. The payment decision should follow those specifics, not from which BNPL provider has the best commercial terms or the most recognizable brand.
Whichever path the merchant chooses, the agent is only as reliable as the infrastructure it runs on. Payment certainty is not a feature. It is a precondition.
