Approve the Policy, Not Every Purchase: Budget Limits for Agentic Commerce
Shopping and procurement agents are arriving with spending controls built in, and the pattern that keeps coming up is approval by default with an optional limit. That is a reasonable starting point. It also raises a design question for anyone running their own agents: is the right control an approval on every purchase, or a budget? They solve different problems, and mixing them up is how teams end up with either an agent nobody can trust or a human who clicks "approve" without reading.
This post is about agentic commerce budget limits: what each control is good for, where the check belongs, and a worked example. We are not summarizing any specific vendor's announcement here. If you are reading about a retailer or card network launching agent profiles, go to their own announcement for what they actually promise, and treat anything secondhand, including this post, with care.
Why one approval per purchase does not scale
An approval is only useful while a person is paying attention. Picture an agent that buys small items from the same vendor all day. The first approval gets a careful look. By the tenth, the request looks identical to the last nine and the approval has become a reflex. At that point the gate is still there, but it no longer tells you anything.
Per-purchase approval also moves the cost to the wrong place. The routine, low-risk spend is the bulk of the volume, and it is what you most want to hand off.
Budgets bound the damage; approvals handle the exceptions
A budget answers a different question: how much can this agent spend, in total, before someone has to look? Routine spend inside the budget goes through without a human. The unusual request is what gets a person's attention.
Conceptually, a policy check ends in one of three outcomes:
- Approve. The request is inside the limits. The agent proceeds.
- Deny. The request breaks a rule, such as a disallowed merchant or an amount over the ceiling. The agent does not proceed, and the reason is recorded.
- Route to a human. The request is allowed in principle but over an approval threshold. A person decides, with an expiry so a stalled request does not hang forever.
For more on that routing, see Agent Spend Approval Workflow.
Where the check belongs
The check runs before the payment call and outside the agent. If it lives in the agent's prompt or its own code, a retry loop, a bad instruction, or a bug can skip it. If it is a separate decision the payment path depends on, the agent cannot talk its way past it.
Paitify is one example of this layer. It evaluates a spend request against a policy before the payment call and returns approve, deny, or human approval, along with the rule behind the decision. It is a policy and audit layer only: it does not move, hold, or custody funds, and it does not sit between your agent and the vendor. Your integration has to ask first, and a check only covers calls that go through it. For how the pieces fit together, see how it works, and for the case against relying on provider caps alone, see Hard Spend Caps Are Becoming the Default.
A worked example
This is a hypothetical, not a report of a real incident. A procurement agent has been approved to buy supplies from one vendor. Someone approved the task once, and the agent then pays that vendor nine times in a row, because a retry loop treats each partial failure as a reason to order again.
With only a single approval at the start, all nine payments are covered by it. Nothing in the setup distinguishes the first payment from the ninth.
Now add two limits. We describe these as a generic pattern; Paitify's policy model has no "session" concept, so with Paitify you would approximate it with a daily limit together with an approval threshold:
- A cap on total spend for the task (in Paitify, a daily limit). Once the running total reaches it, further requests are declined.
- A per-payment approval threshold. Any single payment above it goes to a person.
If each order is, say, 40 and the task cap is 150, the first three payments go through and the fourth is declined, and the decline is in the log with the rule that caused it. The retry loop is still a bug, but it now costs 120 instead of 360, and the declined request is the signal that something is wrong.
Try it
Pick one agent and put its paid calls behind an authorize step. The free tier includes one agent, one policy, and 1,000 authorizations per month. You can read the API reference, or watch the demo video on the homepage to see an agent approved and then stopped. If you want the shorter definition first, start with Agent Spend Authorization.