AI agents are becoming increasingly capable of acting autonomously, yet the financial infrastructure around them remains fundamentally designed for humans.
An agent can reason, browse the internet, write code and execute increasingly complex workflows. Yet the moment it needs to pay someone, access capital, or outsource something it cannot do itself, that autonomy begins to break down rather drastically. The existing financial system assumes that a person or legal entity ultimately sits behind every transaction, and understandably so due to the inherent trust assumptions. AI agents do not fit particularly neatly into that architecture.
This creates an interesting bottleneck. If agents are eventually going to operate businesses, manage workflows and participate independently in the economy, being able to think and act is only part of the equation. They also need the ability to transact in a structured and safe manner.
Payman is building the infrastructure for them to do exactly that.
The easiest way to think about Payman is as infrastructure allowing AI agents to transact with humans. An agent might be capable of completing 95% of a workflow itself, but eventually encounter something it cannot do. Perhaps it needs someone to physically inspect something, verify a piece of information, label data, make a phone call or perform some other task requiring human judgement. Today, that usually means the autonomous part of the workflow stops.
Payman instead gives the agent access to both capital and human labour. Through its API, an AI system can create a task, allocate money towards completing it and source someone capable of doing the work. Once completed, the task can be verified and the worker paid. The important distinction here is that the human does not necessarily initiate any part of this process. The agent does.
That sounds like a relatively small change. I think it is considerably more important than it might initially appear. If agents are eventually going to operate businesses, manage workflows or act on our behalf, they cannot simply be good at thinking; they need to be capable of acting. Money is part of that.
There is an understandable tendency to frame AI as a replacement for human labour. I am not convinced that is the most useful way to think about the transition, at least in the near term. AI is exceptionally good at some things and remarkably bad at others. Humans have the inverse problem. The more interesting system is therefore probably not one in which either operates independently, but one in which agents can dynamically pull humans into a workflow whenever necessary.
Payman effectively turns this into infrastructure. Human workers can make their skills and availability accessible to agents. An agent can then determine what it needs, post the task and select someone capable of completing it. This creates a peculiar inversion of the labour marketplace. Historically, humans have hired humans. More recently, humans have begun hiring AI. Payman is betting that AI will increasingly hire humans, as futuristic and somewhat counterintuitive as that may come off.
The marketplace therefore becomes less about people searching for jobs and more about autonomous software allocating work. That is a fairly interesting idea to me.
The second part of the problem is payments. Giving an AI the ability to decide that money should be spent is one thing. Giving it the infrastructure to actually move that money is another. Payman allows developers to fund an agent and programmatically give it spending capacity. The platform provides infrastructure around funding, task allocation and payments, with transactions able to move through both traditional banking networks and cryptocurrencies.
This matters because the payment itself should ultimately be almost invisible. An agent should not particularly care whether the person completing a task wants dollars in a bank account or crypto in a wallet. It should care that the task costs $50, that the person can complete it, and that payment can be made once the work is fully verified. The underlying rail becomes an implementation detail more than anything else.
This is also where crypto becomes considerably more interesting than another speculative use case. Autonomous software is inherently internet-native. Giving software access to internet-native money is fairly intuitive. But Payman does not need to make a binary bet between crypto and traditional finance. Supporting both allows the agent to operate across whichever financial infrastructure the human on the other side already uses.
There is another problem hiding inside all of this. If an agent hires a human, how does it know the work was actually done?
Payman’s initial architecture uses human verification agents to review completed tasks before payment is released. Combined with reputation data around workers, this creates a feedback loop around reliability and performance. The product shown at the time already demonstrated this basic workflow: an agent creates a task through the API, humans complete it, the work is verified, and payment follows.
It is admittedly somewhat recursive. AI hires a human to perform a task, another human may verify the task, and software coordinates the entire interaction. Over time, presumably more of the verification layer itself can be automated. But that is not particularly important for the initial thesis. The point is to allow agents to complete workflows today, including the portions where AI remains insufficient.
Waiting for models to become capable of doing everything is one solution. Giving them the ability to ask humans for help is another. The latter seems considerably more practical.
The timing is what makes the idea particularly interesting. AI agents are moving from something largely discussed conceptually towards products people are actually trying to build. The constraint increasingly shifts away from whether an LLM can generate an answer towards whether an agent can reliably take actions in the real world. Payments sit directly in that transition.
Payman had already attracted significant early interest, with the original thesis noting a waitlist of more than 10k users. The important part is less the absolute number than what it suggests: developers were already beginning to encounter the problem Payman was designed to solve.
There are obviously several things that need to go right. Agents need to become useful enough that people are comfortable giving them meaningful autonomy. Developers need to want financial infrastructure built specifically around agents rather than simply adapting existing payment APIs. There needs to be enough demand for human intervention to justify a marketplace. And perhaps most importantly, giving autonomous software access to money introduces an entirely new set of security, fraud and regulatory problems.
None of these is trivial.
But the underlying bet does not require believing that AI replaces everyone. Almost the opposite. It requires believing that AI becomes sufficiently capable to coordinate meaningful economic activity while remaining imperfect enough to occasionally need us.
That seems like a reasonably defensible place to start.
If agents become autonomous economic actors, they will need access to the same primitive that coordinates most human economic activity: money. And if they cannot complete everything themselves, they will need some mechanism for converting that money into human effort.
Payman is attempting to build the layer between the two. Perhaps the future of work is not simply humans working for companies or AI working for humans. Perhaps, increasingly, we work for the machines too.