AI agents are beginning to interact directly with payment and commerce systems. In agentic commerce, software can select an option and initiate a transaction within rules set by a user or business. An agent can prepare an order, initiate payment, and pause when human approval is required. In some workflows, one agent may transact directly with another. That broader Agentic Commerce model is becoming practical because several technical layers are developing at the same time.
Agents can use external tools more reliably, transaction permissions can be defined more precisely, and new protocols are connecting agents with merchants and payment providers. 24/7 settlement networks add another option for automated transactions. Payment systems must still record who authorized each transaction, what was approved, how funds moved, and what evidence remains.
AI can now move from recommendation to execution
Earlier AI systems were mainly used to identify products and suggest actions. Completing an economic task required separate transaction infrastructure. Tool-enabled agents can already handle parts of a transaction workflow. They may retrieve structured offers, apply user-defined criteria, and pass selected information to an external system.
This makes a different type of interaction possible. Consider a company booking cloud capacity. Its policy might restrict the agent to approved providers, set a monthly limit, and require a manager’s confirmation for amounts above a certain threshold. Within those constraints, the agent could compare available offers and prepare the purchase.
For businesses, the same model can apply to recurring and policy-based activity. An agent could:
- Purchase approved cloud or data services
- Renew software within a department budget
- Reorder inventory from an authorized supplier
- Initiate a treasury transfer between operating entities
The agent’s action has defined limits and leaves an authorization record.
Commerce protocols are creating a common language
A conventional website is designed for a person navigating pages and forms. An agent works more reliably when the same information is available in a structured format: the offer, its current price, the accepted payment method, and the status of the transaction. Emerging commerce and payment protocols are beginning to provide this machine-readable layer.
Google’s AP2 and UCP, the Agentic Commerce Protocol developed by OpenAI and Stripe, and agent-payment initiatives from Visa and Mastercard approach the problem from different directions. The initiatives cover different layers of the transaction process. Their scope includes merchant connectivity, records of user intent and authorization, and mechanisms for identifying trusted agents.
This work matters because autonomous transactions require a shared format for several questions:
- What did the user or business authorize?
- Which agent is acting under that authority?
- What limits, merchants, or categories are permitted?
- How can the instruction be revoked or challenged later?
A payment credential confirms that the system can access funds. Separate authorization data is needed to show whether the user approved the merchant, amount, and purchase category.
Payment infrastructure is becoming software-readable
Most payment systems still assume a person is present at checkout or that an instruction moves through a bank’s existing operational process. An agent may initiate activity outside that pattern, for example, at any hour or across several service providers. This puts pressure on providers to expose reliable APIs and return transaction status without delay. The resulting records must also be detailed enough for reconciliation and investigation.
Existing payment methods will remain part of agent-led transaction flows. A company may use cards for software subscriptions and bank transfers for larger supplier payments. The orchestration layer would choose the method and apply the relevant controls before execution.
Financial institutions will need to link the agent’s instruction to the order and the resulting payment. Fees, ledger records, refunds and settlement data should remain traceable to the same transaction. The harder problem may be what happens when execution is incomplete. An agent might submit the same order twice or receive a payment confirmation without a corresponding order confirmation.
Stablecoins add a 24/7 settlement option
Stablecoins may be most useful for agents operating across borders or purchasing digital services, particularly when payments are frequent, automated, and too small for conventional cross-border fee structures.
Software may need to purchase one API call, a small dataset, a model inference, or a short period of computing capacity. For low-value automated payments, conventional cross-border fees may exceed the economic value of the transaction. Blockchain-based settlement is one possible alternative, depending on network costs and the surrounding compliance infrastructure.
The settlement layer still requires custody, liquidity, sanctions screening, accounting, redemption, and fiat conversion. Providers such as Coinspaid connect blockchain settlement with the compliance, liquidity, accounting, and fiat systems required for business use.
What still has to be solved
Production deployment depends on explicit controls over permissions, signing, and human intervention. Institutions need to restrict what the agent can sign, how much it can spend, and when a person must intervene. A production-ready system may require:
- Scoped and revocable credentials
- Per-transaction and cumulative limits
- Verified merchant or counterparty identities
- Separation between the AI model and transaction signing
- Complete audit and reconciliation records
- Human approval for exceptions or higher-risk activity
The first viable deployments are therefore likely to be narrow and observable. Known counterparties, limited values, and repeatable tasks give institutions a controlled way to measure fraud, false declines, operational workload, and cost per completed transaction.
Agentic commerce is becoming technically possible because the separate components are starting to connect. An agent can pass a structured instruction to a payment flow that preserves its authorization limits and produces a usable transaction record. Institutional adoption will depend on whether the full transaction flow can handle authorization, execution failures, revocation, and disputes reliably.
About Author
