Key takeaways
- In agentic payments, the agent handles discovery, payment and the receipt, while a human sets the budget and the approval rules.
- Card fees of 2.9% plus 30 cents make card rails a poor fit for the small, per-call purchases agents make most often.
- MCP lets AI assistants connect to tools, but the protocol itself has no payment step.
- x402, AP2, ACP, Visa Intelligent Commerce and Mastercard Agent Pay each solve a different part of agent payments and can complement each other.
- Safe agent spending rests on a separate small wallet, an auto-approve limit, human approval for larger purchases and a receipt for every call.
What are agentic payments?
Agentic payments are payments that software agents make within rules a person sets. For decades a human sat at every checkout: reading the price, typing a card number, clicking Pay. In an agentic payment the agent does those steps, and the human decides the budget and the approval rules up front.
The payer can be an AI assistant buying a data lookup for its user, a coding agent paying for one API call, or a business agent ordering from a supplier's agent. The seller can be a person, a company or another agent. A payment counts as agentic when the decision to buy and the payment itself both happen inside the agent's workflow.
Demand is growing quickly. IDC projects agent tasks rising from 44 billion a year in 2025 to 415 trillion by 2030. OpenRouter reported in August 2026 that agent transactions are growing 14 times faster than human digital transactions.
How do AI agents pay?
Most AI agent payments follow the same loop, whatever system sits underneath.
- Discover. The agent searches for something that fits its task, such as a weather lookup, a dataset or a file. It reads the listing for the price, the seller and the inputs it must send.
- Check the rules. The agent compares the price with what it is allowed to spend. Small amounts can be pre-approved; larger ones go to a human.
- Get approval if needed. The person approves on a web page, in the chat or on a phone, often with a passkey.
- Pay. The agent pays from a wallet, token or mandate it was given. It does not need the person's card number.
- Receive the result. The seller's service returns what was bought: an answer, a download link, a stream.
- Keep the receipt. The agent records a transaction id and a receipt, so the person can audit the spend or ask for a refund.
The real differences between systems sit in steps 4 and 6: what the agent pays with, where the money settles, and how anyone proves the payment happened.
Why do AI agent payments need new rails?
Three things break when you hand today's online checkout to an agent.
Price size. Agents often buy small things many times: one API call, one search, one record. Card processing is priced for larger purchases. Stripe's standard US online price is 2.9% plus 30 cents per successful domestic card transaction (Stripe pricing). On a $0.10 charge the fee is about $0.30, more than the sale itself. On a $1 charge the fee takes about a third.
Connection without payment. The Model Context Protocol (MCP) is the open protocol many AI assistants use to connect to tools and data. Remote MCP servers authorize with OAuth 2.1, and MCP itself has no payment step (MCP specification). An agent can reach a tool, yet it has no standard way inside MCP to pay for it.
Trust and proof. A merchant needs to know which agent is calling and who authorized it. The person needs proof of what the agent bought. Card statements and chat logs do neither job well across thousands of small, automated purchases.
What is agentic commerce?
Agentic commerce is the wider market that agentic payments serve. Agents discover products, compare them, order and pay, while sellers publish offers that agents can read. Payments are one layer; catalogs, discovery, approvals, delivery and payouts are the others.
In agentic commerce, a buyer may never visit the seller's website. The buyer's agent reads a machine-readable listing, calls the seller's service and pays in one flow. Sellers who want those sales need listings agents can find, prices agents can understand and a way to get paid per use. Our guide on how to sell an API to AI assistants covers the seller side.
Which protocols handle agentic payments?
Several open protocols and card network programs launched in 2025. They solve different parts of the problem, and these approaches can complement each other.
| System | Backed by | What the agent pays with | Where it settles | Typical fit |
|---|---|---|---|---|
| x402 | Created by Coinbase (2025); x402 Foundation at the Linux Foundation since April 2026 | A signed payment sent in an X-PAYMENT header after an HTTP 402 reply | Onchain, in stablecoins such as USDC on chains such as Base | Pay-per-request APIs for crypto-native builders |
| AP2 | Google, with 60+ merchants and financial institutions | Signed intent and cart mandates | Depends on the payment method; a crypto extension integrates x402 | Agent shopping with a clear record of what the user allowed |
| ACP | Stripe and OpenAI | A scoped, logged payment token passed to the merchant | Card rails | Instant Checkout in ChatGPT |
| Visa Intelligent Commerce | Visa | Agents identified through the Trusted Agent Protocol | Card rails | Agent checkout at card-accepting merchants |
| Mastercard Agent Pay | Mastercard | Agentic tokens bound to one agent | Card rails | Agent checkout at card-accepting merchants |
| Qatom | TODAQ Micro Inc. | An agent wallet (twin) holding USD-TDN | Between twins, with no gas and no public ledger | Per-call APIs, data and services sold to agents |
x402 is an open protocol created by Coinbase, announced on 6 May 2025 and governed since April 2026 by the x402 Foundation under the Linux Foundation, that revives the HTTP 402 "Payment Required" status. The server answers with a price and accepted tokens, the client signs a payment and retries, and a facilitator verifies and settles onchain (Coinbase). Independent 2026 research found real x402 volume far below raw settlement counts and security gaps in payment facilitators, since mitigated; see x402 in 2026.
AP2, the Agent Payments Protocol, is an open protocol from Google announced on 16 September 2025. An intent mandate records what the user lets the agent look for and spend, and a cart mandate approves a specific purchase. A detailed intent mandate with price limits lets the agent buy without a second approval (TechCrunch).
ACP, the Agentic Commerce Protocol, is an Apache 2.0 open standard co-developed by Stripe and OpenAI and announced on 29 September 2025. It powers Instant Checkout in ChatGPT (Stripe).
Visa Intelligent Commerce uses the Trusted Agent Protocol so merchants can verify the agents that call them. Mastercard Agent Pay issues agentic tokens bound to one agent that carry who authorized it (RisingWave). The card-based options suit physical goods at checkout. For a deeper side-by-side, read agent payment protocols compared.
How do you keep an AI agent's spending under control?
The controls matter more than the protocol. Whatever system you use, look for these five.
- A separate, small budget. Give the agent its own wallet and move only a little money into it. On Qatom the agent twin's balance is a hard ceiling, and the agent never touches your primary wallet or card (move funds to your agent).
- An auto-approve limit. Purchases under a limit you set go through without asking.
- Approval for larger purchases. Above the limit, nothing is paid until you approve, in the chat or by passkey or phone app (approve a purchase).
- A receipt for every call. Each purchase should return a transaction id you can quote in a refund request.
- Care with retries. A retried paid call is a new charge, so check the last receipt and the balance before retrying (retry safely).
Our guide on how to let an AI agent make purchases walks through each control step by step.
Where does Qatom fit in agentic payments?
Qatom is an all-in-one agentic commerce platform. Sellers turn APIs, data and services into catalog items, called storefronts, that AI agents discover, order from and pay for. Buyers' agents purchase the same way, inside spending limits the buyer sets, with a verifiable receipt for every call.
Buyers connect any MCP client, such as Claude, Grok or Muse, to the Qatom master MCP at https://mcp.m.todaq.net/mcp and search with catalog_search (connect your AI). Storefronts are discoverable in the assistants and channels listed on qatom.ai. The round trip runs from buyer agent to Qatom twin to seller agent in one call, with payment, delivery and a receipt returned together.
Money inside Qatom is USD-TDN, a US-dollar token where 1 USD-TDN equals 1 USD. RBC holds the reserve funds backing USD inside Qatom and supports bank on-ramp and off-ramp transfers. Each twin is built on the TODA file protocol, so every holding is a self-verifying file and a receipt can be checked anywhere.
On cost, Qatom charges a 0% platform take rate, no gas and no transaction fees, and transaction cost does not scale with the value or volume of a payment. Settlement takes under one second. Prices can go as low as 0.001 USD-TDN per call. Plans are Builder (free), Pro ($20 a month), Max ($300 a month) and Enterprise (pricing), with hosted, VPC, on-premises and air-gapped deployment. Developers can start with the Qatom developer docs.
Qatom is headquartered in Toronto. Building on TODAQ research since 2017, it has been in production since 2023 on the TODA File Protocol and the ADOT Protocol (about us).
Next steps
- Compare the main standards in agent payment protocols compared.
- See why small payments need different rails in micropayments for APIs.
- Browse every buyer and seller step in the Capabilities Index, or contact us with questions.