Can an AI Pay for Things?
How Autonomous Agents Are Learning to Spend Money.
- What Does It Mean for an AI to Make Payments?
- Three Real Examples of AI Agent Payments Today
- Why Are Stablecoins the Default Currency for Agent Payments?
- What Are the Numbers Right Now?
- What Does This Mean if You Hold Crypto?
- The Risks of Agentic Payments
- Conclusion
- Frequently Asked Questions
- Sources
In April 2025, Mastercard processed what it described as the first real agentic transaction, a payment made not by a person, but by a piece of software acting on their behalf.
A few months later, OpenAI and PayPal connected their systems so that ChatGPT users could complete purchases without ever leaving the chat window. By February 2026, Stripe had enabled AI agents to spend stablecoins on the Base blockchain. In May 2026, Coinbase's CEO told investors that AI agents might eventually outspend humans.
If you are even slightly curious about where crypto, AI, and payments are colliding, you are in the right place. This article breaks down what AI agent payments actually are, shows real examples of how they work in the wild, and explains why stablecoins and crypto rails are emerging as the preferred plumbing for this new kind of commerce.
What Does It Mean for an AI to Make Payments?
An AI agent is software that can take actions toward a goal without a human approving every step. You might tell an agent to plan your quarterly business trip and keep it under $2,000. The agent would research flights, compare hotels, find the cheapest itinerary, and book everything. That last part, booking, involves a payment.
A chatbot stops at the recommendation. An agent goes ahead and pays.
For the agent to actually pay, it needs a financial instrument it can use programmatically. Humans use credit cards and bank transfers, but those tools require identity verification, billing addresses, and human-managed credentials. Software cannot walk up to a bank and open an account. It has no Social Security number.
So what does it use instead? Increasingly: a crypto wallet loaded with stablecoins, connected to protocols that were specifically built for machine-to-machine transactions. This is what people mean by AI agent payments.
Three Real Examples of AI Agent Payments Today
1. Booking a Flight
This one is the most vivid example and, as of early 2026, very nearly real. Sabre, PayPal, and MindTrip announced a partnership to build the first end-to-end agentic travel booking pipeline, combining conversational trip planning, real-time inventory from over 420 airlines and 2 million hotels, and integrated payment, all in a single chat-based experience. The target launch was Q2 2026.
In practice, this means telling an AI assistant, "Find me a direct flight to Chicago next Friday under $350," and having the agent search, select, and pay for the ticket autonomously. You approve a budget up front. You do not sit at the keyboard, clicking confirm on each screen.
Mastercard's Agent Pay is already live on the card-network side of this flow, debuting in the US with Citi and US Bank cardholders in 2025 before expanding to Australia, New Zealand, and Europe in early 2026. Microsoft Copilot is being integrated into Copilot Checkout using this same infrastructure.
2. Buying API Credits
This is the most active slice of agentic payments right now, and it is running almost entirely on crypto rails.
When an AI agent needs to call a paid API, such as a data provider, a blockchain analytics service, or a cloud inference endpoint, it makes sense to pay per request rather than commit to a monthly subscription. Coinbase built the x402 protocol exactly for this. When an agent hits a payment-gated endpoint, the server returns a 402 "Payment Required" code along with payment instructions. The agent pays in USDC, attaches proof to the next request, and receives the service.
As of May 2026, x402 has processed over 165 million transactions. Roughly 69,000 active AI agents are using it. The annualized volume is estimated at around $600 million. Amazon AWS integrated the protocol directly into its Bedrock AgentCore platform, meaning enterprise developers on AWS can give their agents payment capabilities without building any custom infrastructure. Settlement takes about 200 milliseconds and costs a fraction of a cent per transaction.
This speed and cost profile is why this works: a $0.004 API call is completely impractical on card rails, where processing fees alone would dwarf the transaction value. On a low-fee blockchain, it is trivial.
3. Subscribing to a Service
Beyond one-off API calls, agents are increasingly managing subscriptions and recurring commitments on behalf of users. PayPal made a high-profile move in Q4 2025, integrating simultaneously with ChatGPT, Perplexity, and Mastercard Agent Pay. Starting in 2026, PayPal users can buy from merchants and manage services directly through ChatGPT conversations, with the agent handling checkout automatically.
OpenAI's Agentic Commerce Protocol, co-developed with Stripe, powers this flow. Shopify and Salesforce are integrating with it. In September 2025, OpenAI launched Instant Checkout for Etsy sellers through this same system.
Google's Universal Commerce Protocol and Agent Payments Protocol (AP2) take a broader approach, with over 60 partner organizations, including Mastercard, PayPal, American Express, Coinbase, Shopify, and Cloudflare, building to the same standard.
Why Are Stablecoins the Default Currency for Agent Payments?
A few reasons stack up together. First, there is the identity problem. AI agents cannot pass Know Your Customer checks at traditional financial institutions. A crypto wallet address is just a public key on a blockchain. Any software that holds the corresponding private key can transact from it.
Second, there is the microtransaction problem. Card rails have a floor below which payments are not economically viable, and processing fees make sub-dollar transactions nearly impossible. A blockchain transaction on Base or Solana can settle for a fraction of a cent, which means agents can pay $0.01 for a single data request without losing money on the fee.
Third, there is the speed problem. Traditional bank settlements run in batches. ACH takes two to three days. An AI agent executing a 30-step research workflow cannot wait 48 hours between each service call. The x402 protocol clears a USDC payment in approximately 200 milliseconds.
Fourth, there is the audit problem. When an agent spends money autonomously, someone needs to be able to verify every transaction after the fact. On-chain records are permanent and public. There is no arguing about what happened. Every payment is timestamped, addressable, and verifiable.
Here you go:
| Credit Card / Fiat | Stablecoin / Crypto | |
|---|---|---|
| Identity | Requires KYC, name, SSN, and address | Public key address — no ID needed |
| Settlement speed | 2–3 day ACH or batch settlement | ~200ms on Base (x402 / USDC) |
| Microtransactions | $0.004 per API call: not viable | Sub-cent txns at near-zero fee |
| Availability | Business hours; downtime possible | 24/7, no banking hours |
| Audit trail | Statements; chargebacks possible | Immutable on-chain ledger |
As of April 2026, 98.6% of machine payments settle in USDC, according to a Keyrock report. Circle's stablecoin is essentially the default settlement currency of the nascent agentic economy.
What Are the Numbers Right Now?
The honest answer is: small but growing very fast.
Keyrock, a crypto trading and investment firm, estimated that AI agents settled over $73 million across roughly 176 million transactions on blockchain rails between May 2025 and April 2026. Visa alone processes $14.5 trillion annually. So the volume is negligible in relative terms.
AI agents influenced $67 billion in global Cyber Week 2025 sales, representing about 20% of all orders, according to Salesforce data. Retailers with AI agent capabilities saw 13% sales growth during that same period, compared to 2% for those without. McKinsey projects that US business-to-consumer agentic commerce will reach $1 trillion in annual revenue by 2030.
What Does This Mean if You Hold Crypto?
AI agent payments are being built primarily on crypto rails. The growth of the agentic economy is a direct driver of stablecoin usage, Layer 2 transaction volume, and on-chain infrastructure more broadly. Coinbase's CEO literally said agents could outspend humans; a meaningful signal about where value is flowing.
There is also a personal security dimension. As AI agents gain payment autonomy, the importance of your own key management goes up, not down. An agent wallet is a hot, software-controlled credential that can transact without per-transaction human approval. If that system is compromised, funds can move. This is precisely why keeping your core holdings in self-custody cold storage matters more as the agent economy grows, not less.
The flip side is opportunity. DeFi protocols that are compatible with agent workflows, blockchains with fast, low-cost settlement, and stablecoin infrastructure are all positioned to benefit from this shift.
The Risks of Agentic Payments
Regulation is lagging badly. MiCA in Europe, the US GENIUS Act, and the EU AI Act are all expected to take effect around mid-2026, but none of them directly address autonomous machine-to-machine transactions, or who is legally liable when an agent payment goes wrong. That grey zone will produce confusion.
Concentration risk is real. As of April 2026, 98.6% of on-chain machine payments use USDC. That is a single stablecoin issuer holding a significant share of the agentic economy's financial plumbing. Any disruption to Circle, the company that issues USDC, would have outsized effects.
Smart contract risk is not hypothetical. When an agent interacts with a DeFi protocol or a programmable payment contract, bugs in that contract can drain funds. The agent will not notice a malicious contract if its policy rules do not specifically check for it. This is an area where the tooling is immature.
Consumer trust is still low. Only 29% of UK consumers and about 16% of US consumers currently say they trust AI to make payments autonomously, according to survey data. Adoption will move slower than the infrastructure build-out suggests, at least in consumer markets.
Conclusion
AI agent payments are not a future concept. The trust layer is still being built. Every major player, from Mastercard to Google to Visa, is working on the same problem, which is how consumers can be confident that an AI agent spent money the right way on their behalf.
The crypto angle will not go away. Card networks are fighting to stay relevant in this space, and they will succeed in consumer markets. But for machine-to-machine payments, sub-cent API transactions, and 24/7 global settlement, the crypto simply fits better.
Frequently Asked Questions
| Can AI agents really spend money on their own? | Yes, though within the user's limits. AI agents can initiate and complete payments autonomously using agent wallets and protocols like x402 or Mastercard Agent Pay, but they operate inside predefined spending rules. |
|---|---|
| What currency do AI agents use to pay? | Mostly USDC, a dollar-pegged stablecoin. As of April 2026, it accounts for roughly 98.6% of on-chain machine payments. Fiat card rails (via Mastercard and Visa) are also used for consumer-facing agent transactions. |
| What is x402? | An open payment protocol created by Coinbase and co-governed with Cloudflare. It revives the HTTP 402 status code to let AI agents pay for APIs and services in USDC over HTTP instantly, with no accounts or subscriptions required. |
| Does this affect my existing crypto holdings? | Not directly, but it matters for context. The agentic economy is a driver of stablecoin and Layer 2 adoption. It also reinforces why self-custody cold storage for your core holdings is important. |
| Is Mastercard Agent Pay the same as x402? | No. They operate on different rails. Mastercard Agent Pay uses card network infrastructure with agentic tokens. x402 uses stablecoins on-chain. Both solve the same problem from different starting points. |
| What happens when an agent payment goes wrong? | On-chain transactions are irreversible. This is why spending limits and whitelisted destination addresses are critical. Policy rules must catch errors before the transaction hits the blockchain. |
| Is it legal for an AI to make purchases? | The agent acts on behalf of a human, who is ultimately responsible. Regulation is still catching up; no jurisdiction has directly addressed autonomous machine-to-machine liability as of mid-2026. |
| Should I put my crypto in an agent wallet? | No. Agent wallets are for operating funds only, funded with small amounts for specific tasks. Core holdings belong in self-custody hardware storage where no software agent can access them. |