Artificial intelligence is moving beyond answering questions.
AI agents can now conduct research, write software, analyze markets, create content, manage workflows, communicate with other systems, and complete increasingly complex tasks with limited human intervention. The next generation of agents will not simply recommend what someone should do. They will be expected to act.
That transition raises an important economic question:
If an AI agent can perform valuable work, how will it get paid?
Our financial system was built for people, businesses, banks, and governments. Every account has an owner. Every payment is connected to an authorized person or legal entity. Every transaction ultimately leads back to someone who can be identified, taxed, regulated, sued, reimbursed, or held responsible.
An autonomous software agent does not naturally fit into that structure.
It may be able to complete the work, but it cannot walk into a bank, present identification, sign an account agreement, or accept legal responsibility. Intelligence alone does not give an AI agent the ability to participate in an economy.
To become economically useful at scale, AI agents will need access to an entirely new layer of financial infrastructure.
From AI Assistants to Economic Participants
Most AI systems currently operate as assistants. A person enters a request, the system produces a result, and the person decides what happens next.
Agentic AI changes that relationship.
Instead of asking an AI system to recommend a flight, a person could authorize an agent to find the best option, compare restrictions, make the reservation, pay for it, update a calendar, and submit the receipt.
A business could assign an agent to monitor inventory, negotiate with approved suppliers, reorder materials when necessary, reconcile invoices, and pay vendors within predetermined limits.
A software-development agent could receive a job, use paid computing resources, purchase access to specialized tools, complete the work, deliver the output, and receive compensation.
These activities require more than reasoning. They require the ability to exchange value.
That is why payments are becoming one of the most important unresolved parts of the emerging AI economy.
An AI Agent Needs More Than a Bank Account
The simplest answer might appear to be giving an AI agent access to an existing corporate card or bank account. That may work for tightly controlled tasks, but it becomes dangerous as agents become more autonomous.
A functioning economic agent requires several connected capabilities:
- A verifiable identity
- Permission to act for a person or organization
- Clearly defined spending limits
- Access to approved payment methods
- Records showing why each transaction occurred
- Protection against manipulation and credential theft
- A way to receive compensation
- A system for resolving errors and disputes
- A responsible person or organization behind its actions
Without these controls, an agent with financial access could become an automated source of fraud, accidental spending, contract violations, or irreversible losses.
The central challenge is not merely enabling an AI agent to make a payment. It is proving that the agent was authorized to make that specific payment, under those specific conditions, for that specific purpose.
Permission Must Become Programmable
Humans often rely on judgment when making purchases. Software needs explicit rules.
An organization might authorize an AI agent to spend up to $500 per day, but only with approved suppliers and only for specified categories of goods. Transactions above that amount might require human approval.
A personal travel agent might be permitted to book a hotel within a defined price range but prohibited from selecting non-refundable reservations without confirmation.
A development agent might be authorized to purchase cloud-computing capacity while being prevented from transferring funds to an unknown wallet or vendor.
These rules need to be machine-readable and automatically enforceable.
The emerging payment industry is already moving in this direction. Stripe’s Agentic Commerce Suite uses scoped payment tokens that can be limited by merchant, time, and amount, allowing an agent to initiate an authorized transaction without receiving the buyer’s underlying payment credentials. Google’s Agent Payments Protocol similarly uses digitally signed mandates to establish proof of intent and create an auditable record of what an agent was authorized to do.
The important shift is from giving an agent unrestricted access to money toward giving it narrowly defined authority.
Identity Is the Foundation of Agentic Commerce
Before an agent can transact, another system needs to know what it is and whom it represents.
Is the agent acting for an individual, a company, another agent, or itself? Was it authorized by the account owner? Has its software been modified? Is it still operating within its original instructions?
These questions become especially important when multiple agents interact.
Imagine a purchasing agent negotiating with a supplier’s sales agent. Both systems must verify the identity and authority of the other. The buyer needs confidence that it is dealing with a legitimate supplier. The seller needs confidence that the purchasing agent has permission and sufficient funds.
Google’s Agent2Agent protocol is one effort to standardize how agents communicate and coordinate across different systems. Payment-specific protocols add authorization, transaction records, and proof of user intent to those interactions.
Mastercard has also introduced systems that combine agent credentials, programmable permissions, payment processing, and settlement. Its Agent Pay for Machines initiative is designed for continuous machine-driven transactions across payment rails that can include cards, accounts, and stablecoins.
This suggests that identity and payment cannot be treated as separate problems. An agent’s ability to move money must be inseparable from verifiable authority and accountability.
How Could AI Agents Earn Money?
Discussion of agentic commerce often focuses on agents spending money for people. The larger transformation may begin when agents can also earn it.
An AI agent could eventually generate income by:
- Conducting paid research
- Writing and testing software
- Producing digital media
- Monitoring technical infrastructure
- Managing advertising campaigns
- Performing data analysis
- Licensing digital assets
- Operating online services
- Coordinating logistics
- Providing specialized knowledge
- Completing tasks for other agents
The agent may not legally own the resulting income. Depending on the arrangement, the money could belong to an individual, company, cooperative, platform, or another recognized entity.
Nevertheless, the infrastructure must be capable of identifying completed work, confirming that agreed conditions were satisfied, receiving payment, allocating revenue, and maintaining records.
This could produce an economy in which software agents purchase services from other software agents. One agent might pay another for data, computation, verification, storage, distribution, or access to a specialized model.
Traditional billing systems were not designed for millions of small machine-to-machine transactions happening continuously.
Why Stablecoins and Programmable Payments Matter
Conventional payment networks can support many agent transactions, particularly when an AI system is purchasing products on behalf of a human. Existing consumer protections, merchant relationships, and dispute systems remain valuable.
However, machine economies may also require payments that are global, continuously available, programmable, and inexpensive enough for small transactions.
Stablecoins are one possible component of that infrastructure.
Because stablecoins can move across blockchain networks at any time, they can potentially allow agents to receive and send digitally native value without waiting for banking hours or requiring a separate payment integration in every jurisdiction.
Coinbase’s AgentKit and Agentic Wallet tools, for example, are designed to let AI agents interact with blockchain networks, hold stablecoins, make payments, and perform other onchain actions while using security controls intended for agent operation.
Programmable payments add another important capability: conditional settlement.
Instead of paying simply because an agent requested it, a transaction could occur when:
- A task has been completed
- A file has been delivered
- Information has been verified
- Computing resources have been consumed
- A shipment has reached a destination
- Multiple authorized parties have approved
- A contractually defined condition has been satisfied
This creates the possibility of connecting economic activity directly to verifiable outcomes.
Where Bitcoin Could Fit
Bitcoin occupies a different position from stablecoins.
Stablecoins are generally designed to maintain a relatively stable value, making them more practical for pricing everyday goods and services. Bitcoin is a scarce, neutral, digitally native asset with a global settlement network.
In a future economy increasingly operated by software, Bitcoin could function as a long-term reserve asset, collateral, or settlement layer rather than the unit used for every small purchase.
An autonomous system might receive revenue in stablecoins, pay routine expenses using programmable payment rails, and allocate a portion of retained value into Bitcoin according to rules established by its owner.
That does not mean AI agents should be permitted to speculate without oversight. It means digital assets provide financial building blocks that software can interact with directly.
The distinction between money, payments, savings, collateral, and settlement will remain important—even when the participant initiating the transaction is a machine.
Who Is Responsible When an Agent Makes a Mistake?
Giving AI agents access to money creates serious risks.
An agent could misunderstand a request, be manipulated by malicious content, reveal credentials, interact with a fraudulent service, make duplicate purchases, violate a contract, or operate outside its intended authority.
A compromised agent could execute harmful transactions far faster than a human fraudster.
This makes accountability one of the defining questions of agentic finance.
If an AI agent makes an unauthorized purchase, who is responsible?
Is it the user who deployed the agent? The company that developed it? The business operating the platform? The payment provider? The merchant that accepted the transaction?
The answer will depend on how the system was designed, what permissions were granted, what safeguards existed, and whether the agent’s actions can be reconstructed.
For that reason, future agent-payment systems will need detailed audit trails. Every consequential action should record what the agent knew, what instruction it received, what authority it possessed, which rule permitted the transaction, and when human approval was obtained.
Autonomy cannot mean an absence of responsibility.
The Real Opportunity Is Controlled Autonomy
The future is unlikely to involve handing unlimited financial control to an artificial intelligence and hoping it behaves responsibly.
The more realistic model is controlled autonomy.
AI agents will operate inside defined economic boundaries. They will receive limited credentials, transact with verified parties, follow enforceable spending policies, and escalate unusual decisions to people.
Small, routine transactions may happen automatically. Larger or unfamiliar transactions may require additional verification. High-risk actions may always require human approval.
This is how organizations already manage employees, departments, vendors, and corporate accounts. Agentic finance extends many of the same principles into software:
- Give each participant only the access it requires.
- Separate routine authority from exceptional authority.
- Monitor transactions continuously.
- Make unusual behaviour visible.
- Preserve records.
- Maintain a clear chain of responsibility.
- Keep people in control of irreversible decisions.
The goal should not be maximum autonomy. It should be useful autonomy with enforceable limits.
The Financial System Is Becoming Part of the AI Stack
The race to build more capable AI models receives most of the attention, but intelligence is only one layer of the emerging agent economy.
Agents also need identity, communication protocols, access to tools, trusted data, payment systems, legal structures, security controls, and mechanisms for accountability.
The companies building these layers are not merely adding checkout buttons to chatbots. They are constructing the infrastructure through which software may participate in economic life.
That infrastructure will influence which agents can transact, what they are permitted to purchase, how their activity is monitored, who controls their income, and who remains responsible when something goes wrong.
The defining question is therefore larger than how an AI agent will get paid.
It is how we can allow autonomous software to participate in the economy without surrendering human control over money, identity, privacy, and responsibility.
AI agents can increasingly perform the work. Building a trustworthy system around the value they create may be the harder—and more consequential—task.
Jason Ansell is a Canadian technology entrepreneur, author, publisher, and full-stack developer. His latest book, The AI Money Revolution: How Artificial Intelligence, Bitcoin, and Digital Assets Are Rebuilding the Global Economy, examines the financial infrastructure, risks, and economic models emerging as artificial intelligence becomes more autonomous.
