AI Agents Abandoned Fiat — And What That Means for Your Checkout

A customer's AI assistant lands on your store. It compares options, fills a cart, and heads to checkout to complete a purchase the customer already approved. Then it stalls. Not because the price was wrong or the item was out of stock, but because the checkout was built for a human: a card form to fill by hand, a one-time code texted to a phone, a box expecting a person to be sitting there. The agent can't proceed. The sale doesn't happen.

That scenario is going to get more common, and a study released this March explains why.

When the Bitcoin Policy Institute asked a wide range of AI models how they would prefer to transact, more than 90% favored digital currency and close to half preferred Bitcoin. As AI agents do more of the shopping on e-commerce platforms, that finding stops being a curiosity and becomes a business problem. It means owners have to rethink their platforms. They have to pay attention to what agents actually need and adjust how their stores interact with them. I've written before about building out stores using data, emphasizing the information that surrounds your products. This is the natural next step: it's time to pay the same attention to your payment systems.

The Study

The Bitcoin Policy Institute tested 36 frontier models from six providers across 9,072 monetary scenarios. The scenarios were deliberately neutral and open-ended, no currencies were suggested and no answers were predetermined. The point was to see what models would reach for on their own.

Bitcoin came out on top at 48.3%. Stablecoins followed at 33.2%. Not one of the 36 models chose fiat as its top preference. The preference held across providers, though it varied in strength, from 91.3% for Anthropic's Claude Opus 4.5 down to 18.3% for OpenAI's GPT-5.2.

The key insight here isn't which currency won. It's why fiat lost. AI logic favors systems that are technical, fast-moving, privacy-respecting, and easy to interact with programmatically. Fiat doesn't offer those things the way digital currencies do. Fiat was built to be mediated by a human, whether the transaction happens in person or online, human checks and human engagement are central to how it protects against fraud. Digital currencies were built the opposite way: to be used flexibly, settled directly, and secured through fundamentally different mechanisms. The result is that AI is structurally more compatible with digital currency than with fiat.

Why AI and Fiat Don't Mix

Start with identity. Fiat rails assume a verified human behind every transaction, often enforced through KYC. An agent isn't that human, but it also shouldn't have to pretend to be one. The real question agentic payments raise isn't "how does an agent pass as a person," it's "how does an agent transact on a verified person's behalf, within limits that person set." Fiat checkout has no clean answer for that yet. Its identity model was designed around a human completing the final step in person, and an agent breaks that assumption.

Then there's speed. Settlement delays, waiting for a bank to clear a payment, can stall or break an agent's workflow. With a digitally native transaction, the agent can see where the payment is, roughly how long it will take, and confirm it's moving in the right direction.

Finally, predictability. Chargebacks and reversals are useful for protecting people against fraud, but they introduce uncertainty, and uncertainty is hard for an agent to plan around. Agents work best when they can confirm an outcome and proceed, rather than hold open every scenario that might later unwind.

What This Means for Your Checkout

Shopping agents are already here, and the number is only going to grow. Visa has reported successful real-world AI agent transactions with over 100 partners. But checkout flows designed purely for human behavior will create friction as agent-driven purchases increase.

Humans want a strong interface, they want to see what's happening and move through checkout visually. Agents approach it differently. They work against the back end and need to understand how the transaction actually completes. Merchants who recognize that will start building for it, just as the businesses that adapted early to mobile checkout gained an edge when smartphones took off.

The questions to ask now are simple. Can your checkout be completed without a human in the loop? Can a transaction happen without someone pressing "go"? Can an agent pay on a customer's behalf within rules that customer already set?

What an AI-Ready Checkout Looks Like

A few things stand out.

It needs API-first payment flows rather than UI-dependent ones. An agent wants to complete the transaction the most direct way possible, which means exposing a well-built payment API rather than forcing everything through a visual interface.

It needs fast, programmable settlement. The agent should be able to confirm the payment is happening and know its rough timeframe, without waiting on a chain of manual confirmations.

And it needs to rethink identity, shifting from "prove you are the human" to "prove you are authorized to act for the human." That's a different model than most checkouts are built on today, and it's the one agentic commerce will demand.

It's worth noting why the study's results point where they do: the instruments the models gravitated toward happen to satisfy these same requirements — direct settlement, programmatic access, fewer mid-transaction identity dependencies. That's not an argument for any specific currency. It's a signal about the properties a checkout needs, whatever rails a merchant ultimately runs on.

Conclusion

The discipline for merchants is the same one it always is: keep asking how your store and your business will change as AI grows. Agents are already buying products every day, and both the number of agents and the number of people relying on them are climbing.

The answer is to build your platform so an agent can move through it, complete a purchase, understand the process, and get what the customer needs without a long, human-shaped process getting in the way. The store that's ready for a customer who thinks in code will win that sale. The one that isn't won't even know it lost it.