AI Goes Shopping
Agents can buy, now what?

The conversation with Matt Destephe, Head of AI Lab, and Duncan Steblyna, SVP Product at dLocal, quickly moved beyond chatbots and copilots. Because something more interesting is already happening. AI is starting to act on our behalf, and that includes spending money.

When the shopper is software

This isn’t entirely hypothetical anymore. AI agents can already find products, make choices and complete transactions on someone’s behalf. Adoption is still early, but the payment infrastructure around them is moving very fast.

And that raises some pretty important questions. Who gives the agent permission to spend? How much can it spend? How does a merchant know it is legitimate? And what happens when something goes wrong?

That middle layer is getting crowded with payment networks and technology companies working on ways to identify agents, prove what a customer authorized and put limits around what an agent can do. But underneath all of that, a payment still has to happen.

That creates an interesting contrast, because the way we shop could change dramatically while the local reality of making a payment stays surprisingly familiar.

Who gives the AI permission?

As agents become more capable, one of the biggest questions may have less to do with intelligence and more to do with trust.

Today, a person chooses something, reaches checkout and approves the payment. Agentic commerce starts to break those moments apart.

You might tell an agent today to buy a flight next month if the price falls below US$ 500. The decision to spend and the transaction itself could happen weeks apart, without you sitting at checkout. So payment systems need to understand more than whether the card or account is valid. They need to understand who the agent represents, what the person actually agreed to and whether this particular purchase falls within those instructions.

That is why identity, intent and permission are becoming such an important part of the conversation around agentic payments.

Faster doesn’t always mean easier

There’s another consequence of AI that gets less attention: the better AI gets at doing the work, the more important it becomes to define the work properly.

If AI can produce something in minutes, a vague instruction doesn’t disappear. It can simply produce the wrong thing faster. Payments could face a similar problem. Giving an agent the ability to spend is one thing. Giving it the right context, permissions and limits (and knowing when a human needs to step in) is another.

There’s also what Matt calls “cognitive debt.” As more decisions and processes are automated, companies risk ending up with systems that work very efficiently but that fewer people can properly explain.