Agentic payments, revisited
Almost a year ago, I published the Agentic payments memo.
Since then, it has become the most trafficked piece of content I’ve ever written, is often cited as one of the formative essays in agentic payments, and has become the foundation for Natural, the company I now run.
I’ve always loved financial services, but this last year has been a hyper-accelerant. Most people experience payments as a background utility, but underneath is one of the most important coordination systems in the world.
You can’t exist in the industrialized world without touching payment systems. Financial systems exist everywhere, the systems behind them have to be perfectly reliable, and when they work best, they disappear—becoming just another part of the underpinning of society.
After spending a year building at the frontier of agentic payments, I love payments even more. Existing systems were built for humans in ways that agents won’t operate. Payments need to be rewritten for that reality, and that’s what we’re building at Natural.
Natural is now 365 days old. We’ve raised $40 million, are quickly approaching 20 people, and have built the most comprehensive agentic payments platform. Given how quickly the world has moved, it felt like the right time to revisit the memo format—this time with even stronger opinions and the benefit of actually having built in the category every day.
So consider this V2. Agentic payments, revisited.
What are agentic payments?
By definition, agentic payments are payments executed by agents. I’m not sure how we converged on this term because I think the market is missing an important expansion term by framing the market as agentic payments & banking. You cannot have payments without a source and destination of funds. There are funds in flight, and then there are funds at rest. agentic payments are not just about how agents move money, but also how they hold money.
I would think of the evolution of agentic payments as broken into two major phases: agent-initiated and agent-executed. Today, most of the payments that occur, especially those in commerce, are simply agent-initiated, similar to the way the early code-generation models evolved. Humans are still largely uncomfortable with delegating complete authority to agents to act on their behalf. In fact, I’m the only person at Natural who does not have limits set when it comes to agentic payments. I’m leaning into it as hard as I can.
But we’re perfectly capable of scaling agent-executed payments. The only thing we’re waiting for is more built-in trust and reliability. Agent-executed payments will then be further subdivided into two phases of adoption: unidirectional and bidirectional. Unidirectional payments are where an agent mediates one side of the payment, originator or beneficiary. This is where almost all customers on Natural exist today.
Bidirectional payments, however, are what many think about when they think about “agentic payments”: flows of value exchange where agents anchor both sides of the transaction and there is real-time decision-making about when the payment will happen, to whom, and under what conditions, all without human intervention.
Why we’re here
ChatGPT was launched on November 30, 2022, and in the time since, the way society approaches knowledge work has radically evolved. But LLMs didn’t stop at prompt-response patterns of interaction. They opened the door for a completely new paradigm of computing and interaction, agents.
The rate at which the modality of work has changed, and continues to change, is staggering. Have you actually forced yourself to pause, take a step back, and think about what the state of the world looked like in 2020? The best companies look nothing like they did then. The way engineering, legal, marketing, and operations teams execute looks radically different.
Engineers now queue agents to work on long running tasks overnight while they sleep, and then wake up to review the work. Think about that statement. Agents are getting increasingly good at long-horizon work.
It’s not an if, but when
Across both fundraises at Natural so far, a common question I got was, “When are agentic payments going to be real? It feels like a market that is not an if, but when.” Well, today Natural moves money on behalf of agents, but we didn’t then. The concern was more fundamental. There wasn’t any material volume. How can you underwrite such a large market on an indeterminate timeline? Is it even going to materialize?
Frankly, I used to joke that I would have loved one more quarter of silence before every payments company decided that they needed an “agentic payments strategy.” But apparently, people were not quite as asleep at the wheel as I had hoped.
We operate under the assumption that 1) agents will execute a majority of global payment volume over the next decade, 2) that the outcome of that is largely monopolistic or duopolistic in nature, and 3) the companies that define that world order will largely be determined over the next eighteen months.
That is not to say that the winner of agentic payments will be doing tens of billions of dollars in TPV over the next twelve months, but it is likely that the preferred provider continues to be the preferred provider in 2027, 2037, and beyond.
Snowballs to visualize network effects
Payments, when they work, are a network-effects-driven business—not just structurally, but in the way their advantages compound through data.
At the simplest level, if I want to exchange value with you, you have to be willing to accept that value. We both have to opt into the same medium of exchange. That agreement is the network. The more people who accept it, the more valuable it becomes, and therefore, the harder it is not to accept. That’s true of human-led payments and it’s true of agentic payments.
What’s unique in this period of payments is that everyone is working on a new network graph from scratch. Agents. Every agent you can identify, underwrite, and trust becomes beneficial to the network, and every new participant inherits the work of the network that came before them.
Markets like this tend to become land grabs quickly.
When I picture them, I picture a snowball rolling down a hill. Snow has just enough surface tension and capillary cohesion to hold together and gather more of itself as it moves. The bigger it gets, the more gravity can work on it. The more gravity works on it, the faster it rolls. And the faster it rolls, the more it collects.
That is what network-driven markets feel like. At first, they look fragile and move slowly. Then they look inevitable. Once they’re moving—accumulating mass and speed—they become basically impossible to slow down.
Superlinear growth in scale
We’re moving toward a world where agents become primary financial actors. For our entire lives, financial actions have largely been taken by humans exerting agency: clicking, approving, initiating, or delegating transactions. Monetary exchange has ultimately been tied back to the physical embodiment of a human, or a set of humans.
There are 8.3 billion humans in the world, and the number of transactions that occur today are largely some derivative of that number—: bounded by human scale, attention, coordination, and time.
Over the next decade, agents are going to become a new economic population, one that greatly outnumbers humans. Agents change the constraint. Because they can interact not only with humans, but with tools, markets, companies, systems, and other agents, the growth in interactions is unlikely to be linear. It is more likely to be superlinear.
As agents proliferate, the potential network graph expands dramatically. Each additional agent does not merely add one more participant to the economy; it creates new possible connections across the existing network. Unlike humans, whose interactions are constrained by embodiment, agents can discover, evaluate, and transact with many more counterparts across the system.
Therefore, the amount of machine-mediated economic activity could be at least a couple of orders of magnitude larger than anything that exists today. Not simply because there are more participants, but because each participant can engage with more of the network, and because the cost of initiating, coordinating, and executing economic interactions may collapse.
What the future looks like
When I think about the future of Natural, I think about walking through the world while value moves between consumers, businesses, and agents autonomously, with very little human oversight, and in a way that becomes almost all-consuming.
Outside of card-present transactions in the physical world, it is unclear to me why any payment will not eventually flow through some agentic system. That may mean execution, observability, authorization, reconciliation, or all of the above. But the work humans do today to coordinate value exchange will increasingly be handled by agents.
This premise is why we named the company Natural.
First, payments should be a natural ability of agents. Something innate. But second, and more gratuitously, if the coordination of value exchange happens in the background at global scale, humans are relieved from spending so much of their time on the mechanics of moving money. We can spend more time in nature, and less time moving integers around screens.
The physical act of moving value from one place to another is not, in itself, the point. It is only useful because of the outcome it enables. Using human time to execute value exchange does not make humans better. It is coordination overhead. And coordination overhead is exactly the kind of work agents are going to absorb.
What becomes even more apparent in this world is how much closer the world feels, and I think this happens for two reasons.
The first is that agentic commerce collapses many of the barriers that make today’s economy feel local by default.
Today, if I want to purchase goods or services, I am far more likely to transact with people or institutions that are physically close to me. That is partly because of localization, partly because of convenience, partly because of trust, and partly because of familiarity. But many of those constraints weaken when the medium of interaction becomes tokens moving between intelligent systems.
If agents are mediating discovery, communication, trust, authorization, execution, and settlement, then language should no longer be a meaningful limiter of commerce. Whether you were taught English, Mandarin, Japanese, or none of the above should not determine your ability to participate in the global economy.
The same is true for payments infrastructure. If a provider can mediate both sides of a transaction synchronously, the logistical barriers that make cross-border commerce feel difficult today—foreign exchange, correspondent banks, delayed settlement, fragmented payment methods—become much less visible to the end participant.
This is obviously an oversimplified view of how economies evolve. But it points toward one of the most interesting ideas in economics becoming far more real: absolute advantage.
Absolute and comparative advantage
A lot of commerce today is rationalized through comparative advantage. Comparative advantage is measured in opportunity cost: can you produce this good or service at a lower relative cost than I can produce it myself?
The industrial revolution made this obvious. Humans no longer had to spend most of their time producing the basics of survival. Factories emerged, specialization increased, and the middle class could focus on narrower forms of work. Specialization improved outcomes because different people, companies, and regions had comparative advantages in producing different things.
But absolute advantage tells a different story. Comparative advantage asks, “Who is better positioned to do this relative to their alternatives?” Absolute advantage asks, “Who is definitively best at this?” Who can produce the highest quantity, at the highest quality, at the lowest cost?
When the world feels compressed, and the barriers to trade fall, comparative advantage starts to feel less important than absolute advantage. Today, I might work with the best accountant I can find in San Francisco. But what happens when that restriction disappears? What happens when I can work with the best accountant, period, at the same cost and with the same relative friction?
You could reasonably say that the best human accountant is still constrained by time, capacity, and their own internal limitations. That is true if the work is still fundamentally rooted in human labor.
But what if the best accountant is an agent? What if it is publicly accessible to anyone, regardless of geography, language, currency, or prior knowledge? All of a sudden, comparative advantage matters much less. What matters is absolute advantage.
That future can feel scary. It implies a significant amount of retraining across almost every imaginable job. But it also contains a real form of egalitarianism: everyone gets access to the same best-in-class option.
The best accountant. The best designer. The best lawyer. The best financial advisor. The best engineer. The best doctor. The best teacher. Not the best one you can find locally. Not the best one you can afford because of where you live or who you know. The best one, available to everyone.
That is what agentic payments make possible. Not just faster transactions, or smarter routing, or better observability. Those are the primitives.
The larger shift is that agents turn value exchange into ambient infrastructure. They make commerce feel less like something humans manually coordinate and more like something the world simply does.
What’s true, at least fallibilistically
Rails
There still seem to be diverging opinions on the foundational set of “rails” that will exist for agents. I’ve seen everything from “stablecoins are the only option for agents” to “agents are still going to use cards” to “agentic payments won’t materialize at all”.
Agents will almost certainly use a combination of modalities to exchange value over a shorter time period. At Natural, we think about this from a product perspective through a hub-and-spoke model. While we'd like to push agentic payments toward a closed-loop model that settles on our internal ledger, we're still a ways away from being able to control a meaningful portion of the volume that way.
Therefore, we have a bunch of products that serve as spokes to the core network.
Today that means paying and requesting funds from people who don’t yet exist on the network, issuing cards to be processed by another PSP, or accepting card payments issued from another bank to ingest into a merchant account. In all of these examples, Natural only mediates one side of the transaction, but they support meaningful flows where agents are transacting today.
Over time, with greater network density, we can push payments toward settling natively on our ledger.
Margin compression
Because agentic payments may globalize faster than any market in payments history, and because agents seek high throughput, fast settlement, and low cost, we are very likely to see margin compression across the payments industry.
This is likely to be offset by the fact that, as mentioned before, the number of transactions that occur is likely to be many orders of magnitude larger than what exists today. That being said, the pricing and structure of agentic payments will likely look different than the historical payments rate sheets we know today. Markets generally tend to price more efficiently as they advance up the technological curve, and the same will be true of this market.
What will be interesting to watch is whether some of the very large existing incumbent financial services providers can reinvent their business models in ways that may look like short-term cannibalization of their top line (in order to preserve long-term flexibility and maneuverability in the market). It's easy as a private company to reposition for a future that you feel like is coming, but much harder as a public company, where you are bound to the reporting of quarterly performance. Public opinion is a ruthless barometer.
Speed of settlement
As mentioned in the first memo, what is most required of agentic payments is high performance, low latency, and fast settlement times. Agents do not operate on the same time scale as humans. They operate on the time scale of machines. They are completing actions and tasks in a matter of seconds or minutes instead of days or weeks.
The latency introduced by existing payment modalities hampers and even threatens the ability for an agent to do what it needs to do at an acceptable speed. If an agent's task is dependent on a downstream service where a payment mediates the exchange of value required to complete that action, then that payment must settle synchronously or as close to real time as possible for the agent to be able to recognize its full value.
Any payment system that introduces friction beyond that is likely to be made irrelevant and obsolete to agents.
Liability and fraud
There will be unique shifts in liability and fraud vectors with agentic payments, but largely, I think the market has overblown the novelty and fear introduced when it comes to agents moving money.
At Natural, we categorize liability into three main areas today:
- Pay-in/pay-out Regulation E disputes
- Card disputes, which touch networks
- Intra-network agentic disputes
The first two categories largely follow existing paths of procedure and liability. The third category, where agents are settling between each other, is where things are largely new.
For us, today, the stance is that we can't be responsible for poorly constructed agents. We're a financial services company, and we're really good at payments infrastructure. We're going to honor your agent's request exactly as instructed. If you're unhappy with the outcome of your agent, that's on you, but if we deviate from your agent's request, then we'll absorb the full liability.
This is likely the right stance to take because it rests on the premise that labs and foundation models are going to continue to improve. If the performance of agentic payments only gets better with the acceleration and proliferation of AI, then the risk here quickly approaches zero as models become more and more performant as a whole.
It's best to bet on the fact that the genie is out of the bottle when it comes to AI. If your product is going to erode with a new model release, then you've got no long-term defensibility. If your product is going to get better with a new model release, then you are correctly positioning yourself on the wave to benefit from compounding advantages.
Limits and controls
The limits and control plane of agentic payments are what makes them valuable today and is likely to be what holds them back from being highly autonomous in the future. The way I think about the evolution of limits and controls is like this: today, people may be largely uncomfortable with agentic payments and therefore may say, "I want a human review for everything over $100." You see that agent running in production for enough time, and then you say, "Okay, it's never made any mistakes. Only human review if it’s over $500." You see that continue, and now you’re over $2,500, then over $10,000. At some point, you see agentic payments run in production for so long that you come to the conclusion, "I don't need to review these payments at all."
This is a two- to five-year lagging version of the way software engineers have evolved with generative code. Very early on, engineers felt a deep need to review every single line of output. Then, Cursor put out Tab so engineers could just look at recommendations and pre-fill the next line. That continued into more robust code blocks and files, and now full projects and PRs are generated by AI with limited human oversight. While it feels like fully autonomous payments are far off in the distance, I argue that they're much closer than you think.
There’s a reason that Boris Cherny at Anthropic recommends that every few months you delete your Claude.md. It's not because he wants you to relearn what is most valuable in your file. It's because that file is actually holding back your ability to do what it is that you want to do, because the skill and intelligence of the model is being limited by its instructions.
Where we’re still challenged
Identity
On Natural, agent identity is a derivative of legal identity. In markets that are heavily regulated like payments, the winners won’t be the ones that move fastest around regulation; they will be the ones that turn regulation into a moat. However, even if you push the burden of identity up from agents to verifiable legal identity, there are still real challenges that seem underdiscussed and unsolved.
As an example, an Agent ID on Natural is a stable ID. That is to say that there’s one UUID generated that is supposed to link to one “agent”. The analogous concept in humans would be something like an SSN in the United States. In human use cases, an SSN works because that identity is tied to your physical body. You, as a human, are stable. When we tie this same concept to agents things get tricky. I can ascribe an ID to an agent, but an agent is just an amalgamation of code.
If I change the underlying code, at what point is that no longer the same agent? 20% different? 50% different? Is it always the same identity? If I’m the only one acting for or with this agent maybe this problem doesn’t seem so big. This identity only has to mean something to me. But when we have agents acting in a greater ecosystem where there is some expectation about the actions that agent is likely to take things get tricky.
If I’ve permissioned an agent that is owned by someone else to do an agreed-upon set of actions on my behalf, and that agent construct has entirely changed, have we broken the shared construct that we’re operating under? I think so.
For the last couple of years, traditional software vendors have had little reason to add additional friction to the action plane of agents. However, as agents become increasingly consequential actors in the world, there will be a strong push toward understanding who is actually responsible for certain actions. Maybe the first indicator that this is real is the recent Anthropic Mythos shutdown by the US government.
Authority and trust
If identity based on some random UUID is hard to hold onto, what should bound an agent's risk and ability to transact in a larger ecosystem is its reputation and trust within that ecosystem. The market needs to develop some central authority for analyzing and providing feedback on those IDs such that there's shared knowledge on the reputation and risk of that agent.
An agent that transacts in a predictable way, in line with the way it says it's going to transact, needs to have a corresponding strong reputational score, and vice versa. Whether this is a delegated and open-source reputation library or some centralized corporate underwriting based on consortium buy-in, I'm not sure which will materialize. Over the next two to five years, we need to push further into understanding the risk and reputation of agents based on their IDs.
Use-cases
At Natural, we get exposure to the most advanced founders building at the front of the market. And while the market has radically evolved since the last version of this memo was written last year, we’re still incredibly early. Stepping outside of the payments and tech bubble, we’re still relatively far away from even approaching the “Early Majority” portion of the adoption curve. While I certainly think agentic payments will have its ChatGPT moment, we’re in a market that is aggressively building for a future that we all collectively believe will exist.
From the driver's seat, while every month we see more and more exciting use cases and serious payments flows, the market is still relatively tractable on a spreadsheet. Every week, I have more conversations with founders who come into the conversation saying something along the lines of, “I’m planning to build something, and I know I want it to be on top of agentic payment rails, but I’m not sure exactly what I want to build yet.” Those kinds of statements didn’t happen this time last year. Not even close. But that market will still take another year for those founders to work through their early trials and tribulations (think about where Natural was a year ago, it was just an idea!).
I’m excited for the crop of companies that are built now that agentic payments infrastructure is closer to being sound, but we’re still early in the types of things that are being built.
Types of payments
I often see agentic payments discussions push toward micropayments. Which makes sense, I guess. Agents need to make a number of important API calls in order to execute on the tasks that they’ve set out to do, and often people want to do that on a pay-as-you-go basis. But to me, this feels more like a feature than a full product.
Micropayments, definitionally, are a subset of payments as a whole. And I find the premise that those companies are going to be able to push upstream to large transactions to be optimistic at best. Internally, I think about this as, “How do you build a company that is as globally important as J.P. Morgan is?” J.P. Morgan moves, “over $10 trillion on 60 million transactions across more than 200 countries and territories in 120 currencies” per day. That’s a staggering number.
It’s much easier to go from a position of structural importance to the way the economy operates down to supporting micropayments than it is to start with small nominal values and push your way up the value chain towards large transactions.
The competitive landscape
Since writing the original memo, this is one of the sections that has evolved the most. When starting Natural, there were very few companies competing to win the market, and many of the incumbent financial services companies hadn’t expressed too much interest. Since then, the number of startups competing for the space has grown exponentially.
Rather than comment on all the companies that now exist, I’m going to comment on the ones I get asked about frequently, or the ones that I think I should get asked about more often.
Stripe
The classic reference. Frankly, I’ve been impressed by how quickly Stripe is rotating. Patrick is one of my favorite CEOs, so I’m happy to see him eyes-wide-open regarding this market. That being said, I’m still rooting for Natural to absolutely crush the market. It’s all in the name of friendly competition.
The race Natural is playing is one of speed, unification, and focus against Stripe. We’re building a unified payments stack from the ground up with no hygiene problems, and getting to build everything from an agent-native, first-principles approach.
Stripe is having to unify all of the products they built or bought over the last decade. Those products have different leadership teams, different architectures, and different surfaces. They have a lot of distribution and product advantages, but having to stitch together multiple products is hard, and changing what you’re known for is not easy.
From the outside, it feels like Stripe is making a lot of different bets on agentic payments. Maybe that's the right strategy, maybe that's the wrong strategy. It doesn't feel like one unified bet, and maybe that's the problem. As someone who is watching from the outside, I would love to see Stripe say, "Here's our agentic payments product and strategy, and here are all the things you can do with it," and not the piecemeal approach that different teams are taking today, which feels relatively fragmented.
Airwallex
The dark horse. I think the ethos behind the way Jack approaches scaling regulatory moat is very similar to the way I think about it at Natural. It’s the hard path. It’s the most expensive path. It’s the most time-consuming path. But measured over a long enough time period, it’s the path that leads to the most durable advantage.
I think Airwallex's problem, at least perceptively from the outside, is that it feels like they’re doing a lot across a lot of different verticals. Multi-product marketing mixes are hard, but preventing product scope and bloat from growing is even harder. I’m not sure exactly who Airwallex is for, and maybe that’s the problem.
But if Airwallex wins payments over the long run it’ll be because they’ve been quietly compounding structural advantages that seemingly no one wants to do (other than Natural). I respect it.
Coinbase
You have to hand it to Brian for undoubtedly marking its place in agentic payments with x402. It’s smart, well-executed, and well-marketed. Whether that’s the bet that prevails in agentic payments over the long run, I have separate thoughts on that, but I’ve been impressed.
Regardless, I’m still under the belief that x402 has a brand and reputation problem. If you’re considering the question, “how do I move this $100k payment” I don’t think of x402. It’s the beauty and harm of the crypto community at large. They cult, in a good way, around early payment protocols, but that same behavior turns off certain sophisticated payments players.
But maybe more harmful, when you allow a large number of agents access to paid services without a strong understanding of identity, what happens when some actors start abusing the services they’re accessing?
Should agents be able to query the OpenAI Chat Completions API with harmful requests? Send unregistered marketing SMS through Twilio without 10DLC registration? All without some strong connection to those merchants and no easy way to revoke access?
Maybe the right long-term bull case on Coinbase is that, eventually, agentic payments proliferates down into everyday consumers, and no one has done better at activating that community than Coinbase. They turned something fringe into a mainstream brand in a way that feels approachable.
Visa & Mastercard
I’m going to lump these into the same category because they face similar challenges and have similar opportunities. Where they’re doing a great job is quickly moving to support the agentic payments that will be card-led. Like I said earlier, this market will exist and those payments will still touch the major networks. For how large both of these companies are, I’m impressed to see the degree to which they’re getting in the weeds on experimental flows with relatively new companies.
The challenge for both is that if the market does push towards stables-based or closed-loop payments, then both risk major payment volume deteriorating from their top line over the next two decades. Here’s a radical idea: if they can coordinate bringing interchange rates down with issuing banks for agent-initiated transactions to compete with the cost structure of alternative rails, then they’ll lock in their market capture.
Regardless of their ability to succeed in changing the cost structure of card transactions, I think the interesting value proposition and positioning of networks is their ability to mediate the identity and trust layer of agentic payments. They're not perfectly positioned to do it today, but if they spend serious time making that their top priority, then the major networks can make sure that they lock in at least their network fee for providing comfort and stability to both sides of a payment that the actors are reputationally sound.
The banks
I’ve met nearly all of them at this point, and while there are some internal champions that are very much aware of how disruptive this market will be, largely they haven’t moved fast, and you may argue are not in a position to move fast.
With the way the regulatory market is currently working, if I were the executive of a top bank, I would be seriously worried that 1) we don’t make the right bets internally to make sure we capitalize on the massive disruption about to occur and 2) we miss the strategic external bets that pose serious downstream balance sheet risks if we fail to work with the right partners. That may be because they choose to work with a different bank, or even more problematically for incumbent banks, they choose to pursue their own charter.
We're in a unique regulatory period where de novo charters are being approved at a rate not seen in the last two decades. Some of these charters will absolutely flail and go up in flames, because they're not adept enough at actually running a nationally chartered bank. However, some companies that are approved for charter pose a real existential threat to both the major global banks as well as the long tail of regional banks and credit unions.
Where Natural sits in this
Natural is building the foundational payments stack for AI agents.
Agentic solutions only work well when the user can go, “Can you do this?” and the answer is yes, recursively—always thinking about something larger and more complex. Because of that, there was no option not to build everything. Any way that an agent wants to store, move, or process payments, Natural does or will support.
We made a bet on agentic payments a year ago, when no one was talking about it. We knew that we needed to own that entire stack. Natural builds and operates the primitives directly: ledgering, money movement, multi-bank settlement, multi-currency, fraud and compliance, agent identity and observability, and more.
Within all of those products, we have the advantage of getting to build them under a unified architecture and share core platform components like agent observability, agent identity, risk and underwriting, and compliance.
Developers can compose these primitives to power any sort of agentic payment flow they’re looking to build. It’ll take time before the scope of what’s being built at Natural is truly felt by the market, as we’re still early in the adoption curve, but we’re working at the front of the most consequential market in financial services.
The writing is on the wall, and there isn’t a better place to be than Natural. Come join us at /careers.
