TL;DR: AI agents are beginning to transact autonomously, but moving money is only part of the problem. Machine commerce also needs verifiable identity, delegated authority, payment permissions, settlement, and audit trails.
Key takeaways:
- A machine payment is a chain of delegated trust. Before an AI agent moves money, systems must confirm who it represents and what it is authorized to do.
- Identity and authority are not the same thing. An AI agent can prove who it is without proving it has the right to spend a company’s money.
- Machine-readable permissions are essential for agent payments. Agents need enforceable limits, such as spending caps and approved merchants.
- Machine commerce will produce huge volumes of low-value transactions. Agents may pay each other fractions of a cent for compute and data, which strains the infrastructure built for human payments.
- Autonomous commerce needs dispute resolution after settlement. When agents disagree about what was delivered, participants need records of who acted, what permissions they held, and what was paid.
- Verifiable receipts make disputes provable. Cryptographic signatures and shared ledgers give all parties one independently verifiable record of what happened.
- Machine commerce is a trust problem first. Before agents can transact at scale, they need verifiable identity, authority, permissions, settlement, and accountability.
AI agents are beginning to cross an important line—transacting with each other.
Until recently, most AI systems could search, compare, recommend, and perform other functions before a transaction. However, the final click or approval is still required by a human.
That’s changing rapidly. In July 2026, Visa (NASDAQ: V) announced that AI agents had completed live purchases with participating European merchants. The agents browsed products, selected items, and initiated purchases on behalf of cardholders.
Mastercard (NASDAQ: MA) is preparing for an even more autonomous agent-powered world. Its Agent Pay for Machines service is designed for systems that transact continuously with one another, including high-volume microtransactions executed without human initiation or approval at every step.
However, getting machines to move money is the final step and a relatively easy problem to solve. The more difficult problem is establishing if and why they are allowed to move it in the first place.
A machine payment is a chain of delegated trust
Imagine a company gives an AI procurement agent responsibility for acquiring cloud-computing capacity.
The company instructs the agent to purchase additional compute as needed, provided the price is at or below a certain level. It also sets a daily limit of $5,000.
The agent then discovers another autonomous system offering the required capacity and begins the transaction.
Before any money moves, several questions need to be answered, such as who the agent represents, whether it’s authorized to conduct the transaction, whether the supplier is approved, what its spending limits are, and whether the selling agent represents the supplier it claims to.
All of these questions must be answered before the transaction is completed and the money moves. Therefore, a machine payment can be summed up as a chain of delegated authority.
Identity and authority are not the same thing
This point is crucial to understand. A machine may be able to prove it is Agent A. However, that does not automatically prove it has the right to spend $5,000 of company money.
Identity tells another system who the agent is and/or who it represents. Authority tells it what the agent is allowed to do on behalf of the owner or user.
Emerging agent protocols already separate these concepts. For example, the Agent2Agent (A2A) protocol allows agents to discover one another, publish capabilities, delegate tasks, and collaborate across platforms. It treats agents much like enterprise applications, with identity and security handled by established security mechanisms.
However, once money or valuable assets are involved, basic authentication isn’t enough. Agents will require a scope of authority when value is at stake. This could include approved merchants, transaction types, spending limits, geographic restrictions, approved assets, expiration times, or conditions requiring human approval.
While humans won’t authorize each transaction individually, we will still authorize the rules governing transactions between agents.
Machine-readable permissions are essential
All of the above determine what permissions systems need to look like.
Consider these instructions:
- “Buy me the cheapest printer paper.”
- “Automatically reorder printer paper whenever stock falls below 20 boxes.”
- “Manage all office supplies. Spend up to $5,000 per month with approved suppliers.”
Each instruction delegates more authority than the previous one. Useful machine-commerce systems need to translate these instructions into limits that another system can verify and enforce.
Mastercard is already aware of this problem in agentic commerce. Its Agent Pay Trust Framework covers identity, intent, controls, execution, and ongoing risk monitoring. It asks essentially the same questions: who is acting, what is authorized, and whether the transaction can be trusted.
Payment is only one part of the transaction
Once authority is established, payment still needs to happen.
Mastercard envisages agents continuously paying other agents for services such as computing, logistics resources, data monitoring, and other machine services.
Some of these transactions may only be worth fractions of a cent, placing very different demands on payment infrastructure than human commerce does today. It’s also highly unlikely humans would be involved in approving every API call, data request, or other microtransaction manually.
Machine commerce will produce enormous numbers of low-value transactions on a continuous basis, and agents will handle most of them.
What happens when agents disagree?
Settlement does not eliminate the need for trust.
For example, if one agent buys 1,000 units of compute and only receives 500, yet the seller claims all 1,000 units were delivered, this creates a dispute. Traditional payment systems already have mature systems for chargebacks, fraud investigations, authentication, and dispute resolution. Autonomous systems will require the same.
With such high volumes of payments occurring and perhaps multiple agents, parties, and protocols being involved, reconstructing what happened could be difficult.
However, it won’t be enough to blame AI agents. Participants will need to be able to determine which agents acted, which organizations they represented, what permissions they had, whether they exceeded those permissions, which instructions triggered actions, whether payments or settlements were completed, and what was or wasn’t delivered.
Machines need verifiable receipts
This is where cryptographic audit infrastructure becomes useful. It can enable agents to sign transactions and reference delegated permissions. It can timestamp transactions, record settlements, and hash records such as support tickets to prove they exist in a particular state without publishing private information.
A blockchain or distributed ledger doesn’t need to store every prompt, internal decision, or API request. It can simply act as an independent verification layer for important events.
With such records in place, the parties to a dispute can verify the same immutable record rather than relying on separate databases maintained by multiple organizations.
Machine commerce is a trust problem first
Payment networks, agent protocols, AI companies, and software developers are already assembling different pieces of the infrastructure required for machine commerce. The remaining challenge is to make all the components work together.
Before autonomous systems can freely transact at scale, they need to establish identity, authority, permissions, settlement, and accountability. Furthermore, all of this must be done so that other machines can verify each stage.
Scalable public blockchains can become part of the infrastructure required to enable all of this across protocols, organizations, and borders.
FAQs:
What is machine-to-machine commerce?
Machine-to-machine commerce occurs when software systems or AI agents autonomously buy, sell, or exchange goods and services with other systems.
Are AI agents already making purchases?
Yes. In July 2026, Visa announced that AI agents had completed live purchases on behalf of cardholders. Agentic-powered payments are not a hypothetical situation – they are occurring now.
Why do AI agents need delegated authority?
An agent’s identity only establishes who or what it is. Delegated authority establishes the next steps—what it is allowed to do, how much it can spend, what type of transactions it can complete, etc.
How could disputes between AI agents be resolved?
Resolving disputes between AI agents will require clear, verifiable records of exactly what happened and when. Cryptographic signatures, timestamps, payment records, and tamper-evident logs could all help provide the required evidence.
Can blockchain support machine-to-machine commerce?
Yes. Machine-to-machine commerce does not necessarily require blockchain technology, but public ledgers can provide timestamped, verifiable records.
In order for artificial intelligence (AI) to work right within the law and thrive in the face of growing challenges, it needs to integrate an enterprise blockchain system that ensures data input quality and ownership—allowing it to keep data safe while also guaranteeing the immutability of data. Check out CoinGeek’s coverage on this emerging tech to learn more why Enterprise blockchain will be the backbone of AI.
Watch | Blockchain + AI: Unlocking Web3’s Future
Be the first to comment