Before an agent can buy: write its spending authority like a job description
AWS Marketplace now lets AI agents pay for products on the company's AWS bill. Here is how to design what an agent may buy, from whom, and who answers for it.
On 19 November 2026 AWS announced that AWS Marketplace "now supports agentic payments over the x402 protocol": AI agents can buy select products, pay per request in US dollars, and the charges settle on the customer's existing AWS bill. AWS adds that customers "can set spending limits, approved sellers, and guardrails before agents transact." Those settings are the easy part. The hard part is deciding what to put in them, and that is a question about your operation, not about the cloud console: which work justifies a purchase, who owns the budget, and what happens when the limit is reached.
A spending limit is a number. Authority is a decision.
Most companies already have spending rules for people: who can approve what amount, which suppliers are on the approved list, which purchases need a second signature. Those rules rarely live in one place. Part sits in a policy document, part in the ERP, part in the head of the finance lead who "just knows" a request looks wrong.
An agent cannot read the head of the finance lead. When it gets the ability to pay, whatever is not written down becomes either "allowed" or "blocked" by default, and neither default is what the business actually meant. A limit of 500 dollars a day says nothing about whether the agent should spend those 500 dollars on a data feed for a customer case or on a tool nobody asked for.
So before turning on payments, write the agent's spending authority the way you would write it for a new hire in purchasing: what it buys, for which work, from whom, up to how much, and who answers for it.
The authority matrix: one row per purchase type
List every kind of purchase the agent could make. One row each. Seven columns are enough:
| Column | Question |
|---|---|
| Purchase type | One noun: "API calls to an address-verification service", "a dataset subscription" |
| Work it serves | The process step that justifies it: "verify delivery address before dispatch" |
| Approved sellers | The named sellers allowed for this row, and nobody else |
| Unit limit | Maximum price per request or per item |
| Period limit | Maximum per case, per day and per month |
| Budget owner | The person whose budget this spend is charged to |
| Over-limit path | What happens at the limit: stop, queue for approval, or switch to a cheaper fallback |
Two rules make the matrix work. First, no row without a process step. If you cannot name the work a purchase serves, the agent should not be able to make it. Second, no row without a budget owner. A charge that lands on "the cloud bill" belongs to nobody, and nobody notices when it doubles.
- 01Case needs inputA step in the process, with a case ID
- 02Match a matrix rowPurchase type, seller, work served
- 03Check the limitsPer request, per case, per period
- 04Pay or escalateWithin limits: pay. Over: owner approves
- 05Record against the caseCost, seller, agent version, result
Each month the budget owner reviews spend per resolved case, not the total bill.
Five design rules before the first payment
- Put the limits in the platform, not the prompt. AWS says the controls are set "before agents transact" and that "Private Marketplace policies still apply." Use them. An instruction like "do not spend more than you need" in a prompt is a wish, not a control.
- Tie every purchase to a case. The payment record should carry the case or process ID that triggered it. AWS says transactions appear in CloudTrail and Cost Explorer; those tell you what was bought and when. Only your case ID tells you why, and whether the case was resolved.
- Budget per case, not only per day. A daily cap stops a runaway loop late in the day. A per-case cap stops it on the first case. An agent that retries a paid call ten times on one stuck request is a design bug that a daily limit hides until the bill arrives.
- Design the over-limit path as a real step. When the agent hits a limit, the case should not just fail. It should go to a named queue with the context attached, so the budget owner can approve, refuse or change the rule. If nobody owns that queue, the limit turns into a silent stop in the middle of a customer's request.
- Start narrow and widen with evidence. Begin with one purchase type, one seller and a low limit. Widen a row only after a month of records shows the spend produced resolved cases at a cost you accept.
The metric to watch is cost per resolved case
Once an agent can pay, its cost is no longer only tokens. It is tokens plus every request it buys from a third party. Measure both against the same unit: the case the agent helped resolve. A data purchase that turns a stuck case into a resolved one is cheap at almost any reasonable price. The same purchase on a case that ends up escalated anyway is pure cost.
This is why the case record matters more than the payment feature. Spend can only be judged against work when every piece of work has a record with a stage and an outcome. Without that, the monthly review becomes a conversation about the total on the bill, and the only lever left is to cut the limit for everything.
A first step for this week
Pick one agent that is close to needing paid inputs: a lookup, a dataset, an external API. Write its authority matrix with one row. Name the process step, the seller, the limits and the budget owner. Then ask the budget owner one question: if this agent hit its limit at three in the afternoon, who would find out, and what would happen to the case? If the answer is "nobody, and the case would stop", design that path before you switch on the first payment.