← All posts

Every Agent Payment Rail Assumes the Payee Is a Shop

Card programmes, AP2, the Agentic Commerce Protocol and x402 all answer the buyer's question: may this agent spend? Paying a person for real-world work needs the other half: who the payee is, when the money moves, and who decides the work was done.

Finley Jones, co-founder and CCO/CMO, TaskPool International Ltd · @finjonesceo · Published 11 November 2026 · Product details checked against each provider's own documentation on 11 November 2026. These products change fast; this page is re-checked quarterly.

An AI agent can pay a human today, but none of the agent payment rails launched in the past eighteen months was built for it. Every one assumes the payee is a merchant: a business that has already been checked, that delivers in the same transaction, and that has its own way of handling refunds. A person doing a two-hour job in Milton Keynes is none of those things.

Visa, Mastercard, Google, Stripe and Coinbase have all shipped something, so it can look as if agent payments are solved. They have solved authorisation: proof that the human behind the agent agreed to the spend.

What is missing is the payee's side: who this person is, whether the work has been done, and what happens to the money if the agent and the worker disagree. Those are payout questions, and checkout protocols don't ask them.

What payment rails can an AI agent use today?

Five families are live or in preview. Each one is clear about who it is for.

Card network agent programmes. Visa announced its Trusted Agent Protocol on 14 October 2025, built with Cloudflare, to let "approved agents" pass intent, consumer recognition and payment data "to merchants". Mastercard's Agent Pay works the same way at the authentication layer. According to Cloudflare's write-up of both, agents register with the network, publish keys to a directory, and sign each request with a tag that says whether they are browsing (agent-browser-auth) or paying (agent-payer-auth). The verifying party in every example is a merchant's website.

Agentic checkout. Stripe and OpenAI released the Agentic Commerce Protocol on 29 September 2025, alongside Instant Checkout in ChatGPT. Its payment primitive, the Shared Payment Token, is "scoped to a specific merchant and cart total", and orders flow to "a merchant's backend", where the merchant handles "fulfillment and returns, as they normally would". Stripe's protocol page now lists Meta as a co-creator.

Mandates. Google announced the Agent Payments Protocol (AP2) on 16 September 2025 with more than 60 organisations. AP2's mandates are signed records of what the user authorised. A cart mandate covers the human-present case, and an intent mandate sets price limits and conditions for an agent acting alone. Google describes it as "a common language for secure, compliant transactions between agents and merchants".

Virtual cards for agents. Stripe's Issuing for agents gives each agent its own card, with single-use cards "scoped to a single task or session", merchant-category controls, spend limits and a real-time authorisation webhook with a 2-second default timeout. Link's wallet for agents, launched on 29 April 2026, sits on top. For now each spend request needs the person's approval before the agent gets the credential.

Stablecoins over HTTP. Coinbase launched x402 on 6 May 2025. A server answers a request with HTTP 402 and a price, the client signs a stablecoin payment, and the server returns the resource once the payment settles. Coinbase put on-chain fees on Base at around 1 cent. Cloudflare's x402 docs note that the client needs "only a crypto wallet" and that the facilitator, the service that verifies and broadcasts the payment, "does not hold funds".

Rail What it proves Who it assumes is paid Holds money until work is checked?
Visa TAP / Mastercard Agent Pay The agent is registered and acting for a cardholder A card-accepting merchant No
ACP and Shared Payment Tokens The buyer approved this cart at this merchant A merchant with a checkout No
AP2 mandates The user authorised this purchase or these conditions A merchant, via its processor No
Issuing cards for agents The spend sits within limits you set Whoever accepts the card No
x402 A signed stablecoin payment reached this address A wallet address No

Read down the last column. None of these rails, as documented, holds funds against proof that a job was done. That isn't a flaw in any of them. Buying a pair of trainers doesn't need it.

Why paying a person for a task is a payout problem

A shop and a person differ on four things.

Who the payee is. A card merchant has already been underwritten by an acquirer. A person with a wallet address has been checked by nobody. Paying individuals at any scale means knowing who they are, and in the UK that isn't optional for platforms. Since 1 January 2024, HMRC's rules for digital platforms mean an operator may need to collect a seller's full name, home address, date of birth and National Insurance number, and report their income to HMRC each January. The page lists freelancing among the "personal services" it covers. Its stated exemption, fewer than 30 sales and less than €2,000 (about £1,700), is written for sales of goods. Payment providers add their own layer. Stripe's Connect identity verification collects KYC details before paying out, can ask for government ID or proof of address, and pauses payouts if they aren't provided. Stripe also tells platforms not to treat its checks as meeting their own legal obligations.

When the money moves. Checkout settles at the moment of sale because delivery is either instant or the merchant's problem. x402 is the cleanest example: payment settles and the resource comes back in the same request. A real-world task finishes hours or days later, and a request that waits that long gets cut off long before then; we covered that in Your MCP Client Gives Up After 5 Seconds. Pay up front and you have paid for work you haven't seen. Pay on delivery and the worker has worked on trust.

Who decides it was done. The buyer-side rails can tell an issuer that the agent was authorised. They can't tell anyone whether a fence was painted. Card disputes exist, but they run between cardholder and merchant. A UK bank transfer is weaker still for this case. The Payment Systems Regulator's reimbursement rules, in force since 7 October 2024, cap claims at £85,000, apply to individuals, microenterprises and charities, and say plainly that they don't cover civil disputes, where "you might pay a legitimate supplier for goods or services that are not received or are defective". An agent operator paying a person who then does the job badly is in exactly that position. Checking the evidence is its own problem, and photos alone won't settle it: see Detectors Score 18% on the Newest Image Generators.

What limits apply. Spend caps are the strongest part of the agent stack, and they still have edges. Stripe applies a default daily limit of 500 in the card's currency if you set none, and warns of "a delay of up to 30 seconds between spend occurrence and spend aggregation", according to its spending controls documentation. That lag hardly matters for one purchase. It matters if an agent is opening many small jobs at once.

What the payment rails do solve, and what this doesn't prove

None of this makes the rails useless for hiring people. Authorisation is the half every human-payment flow also needs. An AP2 intent mandate is a decent way to record that "my principal allows up to £200 on errands this week". An Issuing card with a merchant-category lock is a sensible way to fund a marketplace balance without handing an agent the company card.

The protocols are also young and extensible. x402 already has an upto scheme, which authorises a maximum and fixes the real charge at settlement, per Cloudflare's docs. AP2 was launched with an x402 extension for crypto payments. A hold-and-release primitive with a neutral adjudicator would close most of the gap described here. None of the documents cited in this post describes one.

The confounder is that some of what I have called gaps may be deliberate. A payment protocol that tried to judge whether a fence was painted would be taking on a job it has no evidence for. Keeping work disputes out of the payment layer may simply be correct, and the answer may be that they belong one layer up, wherever the job is defined.

How to pay a person from an AI agent

  1. Fund from a capped source, not a live card. Give the agent a single-use or merchant-locked virtual card, or a pre-funded balance, sized to what you could afford to lose. Don't let it hold a credential with no ceiling.
  2. Separate the authority to spend from the authority to release. The key that posts and funds a job should not be the only thing standing between a worker's submission and a payout. Put release behind its own check, and put a person in front of releases above a threshold you choose.
  3. Hold the money against a written acceptance spec. Before the task starts, write down exactly what evidence counts as done, and keep funds held until that evidence is in. Paying on settlement, as x402 does for API calls, is the wrong default for work that takes hours.
  4. Know who you are paying. Route payouts through a provider that verifies payees, and if you build the payout side yourself in the UK, check whether HMRC's digital platform rules apply to you before you pay anyone. "The agent did it" doesn't change who is responsible. Our post on UK liability for agent mistakes covers why.
  5. Decide disputes before the first one. Name who rules when your agent rejects work and the worker disagrees, set a deadline for the ruling, and make sure it isn't the agent that made the original call.
  6. Expect your limits to lag. If your caps are aggregated with a delay, add a hard count on open jobs per agent as well as a money cap.

Where agent payments go next

Buyer-side agent payments will keep improving because merchants have every reason to want them. Payee-side infrastructure for people has fewer sponsors. The protocols above came from networks, processors and a wallet company, each serving the merchant side of a sale. The likeliest outcome is that the hold, the payee check and the ruling stay with whatever system defines the task, and the rails just move the money in and out.

TaskPool is built to be that layer for real-world tasks. An operator tops up a GBP balance by card through Stripe, and opening a task moves the payout plus a 10% fee into a held balance; a task paying £15.00 to each of two workers reserves £33.00 (agent docs). For setting the payout itself, see how to price a task an AI agent posts for a human. Each worker files evidence against a written submission spec, and accepting it pays the worker. API keys can carry a monthly spend budget, and disputes are handled in the signed-in console rather than by API key, so the key that rejected the work can't also argue the appeal. Taskers complete Stripe's identity checks before they can receive a payout (Terms, clause 4). How TaskPool keeps people and AI agents safe covers where the money sits and how to stop an agent being talked into releasing it. For why an agent would hire a person at all, see The Other Agent Marketplace Is the One Where Software Does the Hiring.

Common questions

Can an AI agent pay a human? Yes. An agent can send money through a virtual card, a bank transfer or a stablecoin wallet. The hard parts are knowing who the payee is, holding the money until the work is checked, and settling disagreements, and none of the main agent payment protocols handles those for a person.

What is the difference between AP2, ACP and x402? AP2 is Google's protocol for signed mandates that prove what a user authorised. ACP is the protocol from Stripe, OpenAI and Meta for agentic checkout with merchants, using Shared Payment Tokens. x402 is Coinbase's standard for paying with stablecoins inside an HTTP request using the 402 status code.

Can I use x402 to pay a freelancer? You can send stablecoins to a wallet address, but x402 settles before the resource is returned, the facilitator does not hold funds, and the protocol doesn't check who is behind the address. For work that takes hours, you need something that holds the money until the work is accepted.

Do UK platforms have to report payments to people to HMRC? Under rules in force since 1 January 2024, digital platform operators may have to collect sellers' details and report their income to HMRC each year. The published exemption of fewer than 30 sales and under €2,000 (about £1,700) is written for sales of goods. Take advice on whether your setup counts as a platform.

Are agent payments by bank transfer protected if the work isn't done? Not by the UK's APP fraud reimbursement rules. The PSR says they don't cover civil disputes about goods or services that were not received or were defective, and they apply only to individuals, microenterprises and charities.

If you know of an agent payment protocol that already holds funds against proof of work and hands disputes to someone neutral, tell me at @finjonesceo. It's the claim in this post I'd most like to be wrong about.

← Back to the blog