When I started out in the software industry, one phrase came up again and again: “software is eating the world”. It was a concise way of saying that software was taking over and reshaping one industry after another.

Today we can confirm it was true. You only have to look at how the top of the S&P 500 has changed: from oil companies and banks to technology firms like Apple, Meta, and Google.

Now we are witnessing a new revolution, one that may touch even more industries than the last. AI is here to stay, and it is here to transform the world.

There is a clearly rising curve of trust in AI. Little by little, as the models improve and we gain experience using them, we grow more confident and tear down small walls. A year ago, giving an agent access to our email and calendar felt reckless. Today it is normalized, part of the daily routine.

We already trust AI with our information. The next challenge? Trusting it with our money.

AI Agents and Payments: The Gap at the Edge of Autonomy

What is an agentic payment?

An agentic payment is a financial transaction initiated autonomously by an AI agent, without direct human action at the moment of execution, to pay instantly for a resource, API call, piece of content, or service required to complete a delegated task.

AI Autonomy and Payment Execution

As we delegate more and more tasks, we keep running into friction in processes that were designed by and for people. Here is one of the clearest: what happens when an agent needs to pay for a resource in order to finish the task it was given?

We will tear down this wall soon, letting agents handle a budget so they can act more autonomously. But are we ready for that?

The difference between an agent mishandling your inbox and an agent making a wrong payment is enormous. A misfiled email is fixed in seconds; a bad transfer is not always so forgiving. This new era needs new protocols and new controls to let AI agents pay on behalf of a person or a company. All of it is being built right now.

Current Payment Infrastructure Was Built for Humans, Not AI Agents

Part of the problem is that the payment system we have was designed for humans. User accounts, email verification, card forms, captchas, SMS confirmations. Every layer of that scaffolding assumes there is a patient person with fingers on the other side.

For the agentic world, much of that simply doesn’t make sense. Why force an agent to create an account on every service it wants to query just once? Why ask it to verify an email? The natural logic is different: pay instantly, and you get the resource. Pure pay-as-you-go, no account, no friction.

This also opens an interesting door for content creators. There is a growing group of media outlets, forums, and tools that don’t want LLMs consuming their content for free and draining their traffic. Pay-as-you-go offers them an elegant way out: when they detect that the caller is an agent, instead of blocking it, they charge a few cents for access. The potential customer is no longer just a service with an API: it’s any content reachable on the web.

Trust, Risk, and Responsibility in Agent-Initiated Payments

None of this works until we solve the hard questions first:

1. Who is liable when an agent makes a fraudulent payment?

Linking an agent to a payment account is technically possible today. What we still have to update are our processes, both on the regulatory side and the operational one. Picture an agent that makes a fraudulent payment: who is responsible? Perhaps the person who assigned the task never asked for anything fraudulent at all.

2. How do you scope an agent’s payment risk?

So before we hand an agent the keys to an account, the first thing we need to define is its level of risk. This is new territory, which means restrictions should start conservative. Like in every other process, agents will have to earn our trust by proving, payment after payment, that they can spend responsibly. That implies strict limits, business rules, and real-time monitoring: autonomy with a tight leash, not a blank check.

3. How do you prevent an AI agent from being tricked into a fraudulent payment?

There is a security dimension too. Every automated payment is a potential attack surface. How do we make sure the agent pays the right recipient, the right amount, for the right resource? How do we keep malicious content from tricking an agent into overpaying, or paying the wrong party entirely?

These aren’t footnotes. They are the real work.

Where Agentic Payments Are Heading

Let’s be honest: this isn’t easy today. The infrastructure isn’t ready yet, the standards are still emerging, and trust is built slowly. But the trajectory is unmistakable. Just as we learned to delegate inbox management, we will learn to delegate bounded, well-controlled payments.

My own bet is that within the next 6 to 12 months we’ll see a far more natural integration between agents and payment platforms like Stripe, along with standards that let an agent detect that a resource requires payment and settle it inline, without breaking its flow. Protocols like the Machine Payments Protocol – a revival of the old, almost forgotten HTTP 402 “Payment Required” – point in exactly that direction: giving machines a standard way to pay for what they need.

At Devengo, we’re getting ready for that future. We believe instant transfers have a natural role to play in the agentic world, and we want to be there when agents really start reaching for the wallet.