Spend7

Agentic payments need a check before the money moves

AcademySpend7 Content Team8 min read
Bundled fibre optic cables, representing the machine-speed traffic behind agentic payments
Bundled fibre optic cables, representing the machine-speed traffic behind agentic payments Photo via Unsplash.

TL;DR: key takeaways

  • An agentic payment is one where software decides to pay, usually in advance and at a speed no human reviews.
  • AP2 mandates, Agent Pay tokens and network trust signals answer identity and authorisation questions. They do not answer whether this payment fits your budget.
  • Place the pre-payment decision after the amount and merchant are known and before settlement, inside the payment tool rather than the prompt.
  • An unknown baseline should be reported as unscored, never as an allow.

A procurement agent is told to buy software licences up to a set amount. It finds a vendor, agrees a price, pays, and moves on to the next task. Nobody reviews that transaction before the money leaves. Multiply that by a few hundred agents running across a company, and you have the subject of this piece: agentic payments. The phrase is used loosely, so this article pins it down. It covers what the term means, what the current networks and protocols actually establish, what they leave open, and where a pre-payment decision belongs in the flow.

What makes a payment agentic

A payment is agentic when the party deciding to pay is software. Three things tend to change at once.

First, the buyer is a process rather than a person at a checkout page. Second, the authorisation was given in advance, sometimes as a written policy days before the purchase. Third, the speed is wrong for a human review queue. A research agent buying market data at a fraction of a cent per call can make thousands of purchases in an hour, and none of them will look unusual on its own.

That combination breaks assumptions that card fraud tooling was built on. A card network sees a token, an amount and a merchant. It was never designed to answer whether the buyer intended this particular purchase, or whether the buyer is an agent that has been stuck in a loop for twenty minutes. The question has moved from "is this card valid?" to "was this agent entitled to make this payment, now, in this amount, with this merchant?"

The rails are arriving in layers

Three announcements from the last eighteen months show the shape of the market.

Google announced the Agent Payments Protocol, AP2, in September 2025 as an open standard for agents paying on a user's behalf. Its central idea is the mandate: a signed record of what the user authorised. Coverage describes intent, cart and payment mandates, and reports a launch group of more than sixty companies. Secondary sources disagree on some structural details, so the mandate schema should be checked against the specification itself before anyone builds on it. Google's AP2 launch, as reported by The Paypers is a reasonable starting point.

Mastercard launched Agent Pay on 29 April 2025. Its Agentic Tokens extend the tokenisation already used for mobile contactless payments and card-on-file. Each token stands in for card details and can be used only by a verified agent acting for the cardholder, within pre-approved parameters. Mastercard's Agent Pay launch, as reported by Maginative gives the consumer framing.

Visa's announcement from its Payments Forum in June 2026 described Visa Intelligent Commerce, a platform meant to give agents the trust, controls and connectivity to discover, start and complete purchases. Visa is enriching tokens with more detail on transaction type, where the token is used and who is paying, and attaching a token assurance signal built from provisioning and behavioural history. Visa's own release is the primary source for those details.

Read together, the networks are building identity, tokens and trust signals. Those answer two questions: is this agent allowed to transact at all, and is this token genuine? They do not answer the questions a finance team actually asks on a Tuesday afternoon: should this particular payment happen, given what this agent usually spends, what our policy says about this merchant, and whether the agent is retrying something that already succeeded?

What each layer proves, and what it does not

LayerWhat it can establishWhat it cannot establish
Agent identity and verificationWhich registered agent is actingWhether this payment fits your budget
Mandate or tokenThe user authorised something of this shapeWhether the cart still matches the intent at settlement
Network trust signalToken provenance and behavioural history at network levelYour merchant policy, your per-agent caps, your retry state
Pre-payment decisionWhether this intent passes your rules and behaviour checks nowThat the merchant is honest, or that fraud is impossible

Table: each layer answers a narrower question than the one a buyer usually assumes it does.

The last row is the one teams miss. A valid mandate and a valid token say that a payment was authorised in principle. They do not say that the agent has now attempted the same purchase forty-one times, or that a decimals bug turned a small price into a large one. Those are facts about your flow, and only a check that runs in your flow can see them.

Where the pre-payment check goes

The order of operations matters more than the tooling. The decision has to happen after the agent knows the amount, the merchant and the category, and before any money moves. If it happens earlier, the check is scoring a guess. If it happens later, it is a report, not a control.

It also has to live outside the agent's reasoning. An instruction in a system prompt saying "do not spend more than a set amount" is something the agent can reason its way around, especially after a long context or a plausible-sounding reason to make an exception. A check inside the payment tool, which the agent calls but cannot edit, does not depend on the agent agreeing with it.

A workable check takes the intent and returns allow, flag or deny, with the reasons attached. Deterministic limits come first: a per-transaction ceiling, and a rolling window so the last 24 hours count rather than the time since midnight. A calendar-day cap has a known hole, where an agent spends its daily allowance just before midnight and again just after. Behavioural signals come second, scored against that agent's own history rather than against the fleet average. Spend7's risk-check API returns one of three decisions: allow, flag or deny. A cap can be configured as decisive (deny) or advisory (flag), which is useful when a rule is new and you want to see what it would have done before it blocks anything.

Two rules keep the check honest. The first is that an agent without enough history is reported as unscored, not allowed. Pretending an unknown is normal is exactly how an unfamiliar payment gets waved through. The second is that only payments which were not denied should count against a cap. Otherwise a burst of refused attempts consumes the allowance and locks out the next legitimate payment, which is a failure that looks like the control working.

A method you can run this week

Start with inputs. For each agentic payment you want to govern, list what the agent knows at the moment it pays: its identifier, the amount in minor units for the currency in use, the merchant, the category, the payment rail, and the quote or mandate it was given. If any of those are missing, the check should say so. Filling the gap with a default is the most common way a control quietly stops being one.

Next, set the deterministic limits before you rely on any behavioural scoring. A per-agent rolling cap, a per-payment ceiling, and a merchant or category allowlist cover the cases that history cannot. They are cheap to explain to an auditor and cheap to test, which matters more than it sounds. A rule you can state in one sentence is a rule someone can check.

Then route the decisions. Allow settles. Deny stops and writes a record of why. Flag is a step-up: it goes to a person or to a second rule, and the payment does not settle silently. A flag that is logged and then paid is the most expensive form of having a control, because it produces a complete record of a loss that the system saw coming.

Finally, read the log after about two weeks. Most teams find the same dull and useful things. The daily cap never came close to binding. One agent trips off-hours checks constantly because it runs an overnight batch. And there was exactly one retry loop that nobody had noticed. Tune the thresholds to what happened. Raise the limits that never bind, and leave alone the one that caught something.

Mistakes that show up in real flows

Treating a mandate as a budget. A mandate records what was authorised. It does not stop the same agent from acting on that authorisation again and again. Keep a rolling cap on top of it.

Sharing one identifier across a fleet. Per-agent baselines are meaningless if twelve agents report under one name. You get one blended history that describes none of them, and an unusual agent hides inside the average.

Setting decimals by habit. USD uses two decimal places, USDC uses six, and many ERC-20 tokens use eighteen. Get the minor units wrong and every cap you set is off by orders of magnitude, in whichever direction hurts.

Assuming the network's signal replaces yours. A network can tell you a token is genuine. It cannot know your budget, your supplier list or your retry state. Treat its signal as one input among several.

Logging flags without a route. If nothing reads the flag queue, the flag is decoration. Decide who sees a flag, how quickly, and what happens if no one responds.

What this does not promise

A risk decision is not a guarantee against fraud, and it is not a settlement. It is an input to a payment decision that you control. Expect false positives early, particularly for agents with short histories, and expect that a determined bad actor will sometimes look ordinary. Tune the thresholds against your own log, keep a human path for flagged payments, and measure the rate of each decision over time so that a drift in behaviour shows up as a number rather than a surprise.

Where to go next

The AP2 payment safety guide covers what a mandate proves and the checks that matter when you verify one. Agent payment rails compared sets the rails side by side, which helps if you are choosing between them. For per-agent caps in practice, read how to set x402 spend limits. Agent mandate governance is the next read on the policy side. The agentic commerce overview explains what breaks when the buyer is software, and the API reference has the request shape for the risk check.

Frequently asked questions

What are agentic payments?
Agentic payments are payments that an AI agent initiates and completes on behalf of a person or business. The agent finds a product, agrees a price and moves money, often without a human reviewing the individual transaction. The authorisation is usually given in advance, through a policy, a mandate or a token scoped to a purpose.
How do agentic payments differ from card payments?
A card payment assumes a person is present, or at least that a person chose the purchase. A card network sees a token, an amount and a merchant, and it cannot tell whether the buyer was a person who meant to buy or an agent that was looping. Agentic payments need controls that check the agent's intent and its history, not only the card's validity.
Does AP2 stop an agent from overspending?
Not on its own. An AP2 mandate records what the user authorised and lets a merchant verify it. It does not limit how often the agent acts on that authorisation. Pair mandate checks with a rolling cap and a per-payment ceiling so repeated payments are bounded.
Can a network's trust signal replace a pre-payment check?
No. A network trust signal can show that a token is genuine and that it has a sound behavioural history at network level. It cannot know your budget, your merchant policy or whether your agent is currently retrying the same call. Those rules sit with you, and they need to run before the transfer.

Score a payment before it settles

Spend7 returns allow, flag or deny in one call, with the signals that produced it. The free tier covers a single agent, its spend caps and its full decision log.

Keep reading