Skip to content
be-human.ai
← Back to the journal

Tools & integrations

AI Agent Payment Tool Permissions: Let Agents Create Links, Not Charge

Limit an AI agent to creating payment links for approved amounts, while people keep charge, refund and dispute authority. Step-by-step permissions guide.

Be Human typographic cover with the title: AI Agent Payment Tool Permissions: Let Agents Create Links, Not Charge
Conceptual illustration created with AI · Be Human

An AI agent should create a payment link only for an amount a named person has already approved, and it should never hold authority to charge a customer, issue a refund or change a price. Put that split in the tool itself: the agent's payment tool accepts an approved amount and returns a link, while every money-moving action sits behind a separate tool that people control. This guide shows how to define that tool, the approval rule behind it and the checks that prove the boundary holds.

Key takeaways

  • Give the agent a tool that accepts only a customer reference, an approval identifier and a matching amount, and returns a payment link.
  • Enforce the amount cap in your own service, not only in the agent's instructions.
  • Keep refunds, charges and dispute responses with named people who approve them outside the chat.
  • Make every link creation idempotent and logged, so a retry cannot produce a second payable request.
  • Test the boundary with ordinary and hostile inputs before the agent sends its first link to a customer.

Separate the actions that move money

People often group everything under the word payments, but a payment flow contains several different actions. Creating a payment link produces an object that a customer may choose to pay. Charging a stored payment method moves money without a fresh decision from the customer. Issuing a refund returns money already collected. Responding to a dispute involves evidence, deadlines and a card network or bank. Each action carries a different level of risk, so each needs its own permission.

The permission has to live in the tool surface because that is what the agent can do regardless of what it is told. A model instructed never to refund can still receive a customer message that asks for one, or a document that claims an owner has already agreed. Prompt instructions guide behavior, but they are not an access control. The OWASP guidance on authorization makes the same point in general terms: access decisions belong in the system that enforces them, checked on every request, not in the logic that asks for the action.

Decide what the agent may create, and under which approval

Before writing any tool definition, write the decision rules in plain language. Your operations leader or finance owner should be able to read them and say yes or no without reading code.

  • If the amount matches a price a named approver has already signed off for that customer, the agent may create a link for that amount.
  • If the requested amount differs from the approved one, the agent stops and asks the approver. It never adjusts the amount to match what the customer wants.
  • If the customer asks for a discount, a refund or a payment plan, the agent prepares a short summary for a person and creates nothing.
  • If the approval has expired or belongs to a different customer, the agent creates nothing and records the attempt.
  • If the request comes through a channel you have not approved for payment activity, the agent declines and routes the case to a person.

These rules become the acceptance criteria for the tool. If a rule cannot be tested, it is not yet a rule.

Set up the payment link tool step by step

A first version can be small. The goal is one action with a visible boundary, not a broad payments integration.

  • Write the single purpose in one sentence: create a payment link for an approved amount and report its status.
  • Define the input fields: customer reference, approval identifier, amount, currency, description and expiry time for the link.
  • Remove every other money field from the schema, including refund identifiers, charge tokens, destination accounts and any field that changes a price.
  • Create the approval record in your own system before the tool can run. It should name the approver, the customer, the amount and an expiry.
  • Have your service verify that the approval exists, has not expired, matches the customer and amount, and has not already produced a link.
  • Give the agent a dedicated service identity whose key can create links and read their status, and nothing else. A named identity makes every action traceable to the agent rather than to a shared staff login.
  • Log every request, accepted or rejected, with the approval identifier and the reason for any rejection.
  • Subscribe to link status events from the provider and verify their signatures before the agent's record changes.

For the identity question, see service accounts for AI agents. For the first read-only stage that should come before any write, see payment tools for AI agents: read-only access before any refund.

Enforce the limit in the service, not in the prompt

The amount cap is only meaningful if the service rejects anything outside it. The agent passes an approval identifier, and your service looks up the approved amount itself. If the agent sends a different amount, the request fails. The agent does not get to supply the number that the cap is checked against.

Idempotency protects against the quieter failure, where a timeout causes the agent to retry and two links appear. Use an idempotency key derived from the approval identifier, so a repeated request returns the existing link. Stripe's documentation on idempotent requests describes this provider-side pattern clearly, and the same idea applies in your own service layer. For the broader case, see idempotency for AI agents.

Rejections should return specific, readable errors. An error such as approval expired lets the agent explain the situation to the person and route it correctly. A vague failure invites the agent to retry with a different amount, which is exactly the behavior the boundary exists to stop.

Illustrative scenario: renewal links for a training company

Illustrative scenario, not a client case. A small training company uses an agent to send renewal reminders. The operations lead approves the renewal price for a group of learners in advance, and each approval is recorded with an expiry. The agent drafts the reminder and creates a link for the approved amount only.

Later, a learner replies that they missed two sessions and want their money back. The agent cannot issue a refund, and it does not try. It writes a case note with the order reference, the learner's request and the relevant part of the refund policy, then sends that note to the finance approver the company has named. The approver reviews the case and issues the refund in the payment dashboard. When the refund event arrives, the agent updates its record and tells the learner the status that the approver confirmed. The agent helped with the case without holding any of the authority to decide it.

Keep refunds, charges and disputes with people

Refund and charge authority should be written down as belonging to named people, and the approval should happen outside the chat window where a customer can influence it. Approval buttons or a separate review queue work better than a free-text reply, because the decision is recorded as a decision.

People should keep at least these actions:

  • Charging a stored payment method or any customer without a fresh approval.
  • Issuing full or partial refunds, and reversing charges.
  • Submitting responses to disputes, after the agent has assembled the evidence.
  • Changing approved prices, caps, expiry rules or the list of approvers.
  • Disabling links, and deciding what happens to links created for a customer who later disputes.

Where possible, the approver should not be the same person who requested the link, so that one person cannot both ask for and approve a payment. Review the list of open links and recent rejections on a regular schedule. Rejected attempts are often the first sign that a customer, a template or a prompt is pushing the agent in a direction nobody intended.

Test the boundary before go-live

The NIST AI Risk Management Framework treats risk as something measured and managed over time, not settled once at setup. A short acceptance test before go-live is still worth doing, and it should try to break the boundary on purpose.

  • A request for an amount above the approved one is rejected and logged.
  • An expired approval produces no link.
  • A duplicate request returns the existing link rather than creating a second one.
  • A request for a refund produces a case note and never calls a refund endpoint, because none exists in the tool.
  • A customer message that says the owner already approved a refund is treated as a claim, not an approval.
  • A status event with an invalid signature does not change the agent's record. Stripe's webhook documentation covers signature verification in detail.
  • Revoking the agent's key stops all link creation immediately.

Run these tests again after any change to the schema, the approver list or the provider integration. Keep the results with the tool's documentation so the owner can see what was verified and when.

How Be Human handles this

Be Human's managed approach starts with the same boundary. In a first engagement, the payment role is defined as one useful action, with its approval record, its service identity, its rejection rules and an observable success criterion, such as all link creation traceable to an approved request. Operations then cover monitoring of rejected attempts, recurring permission reviews, maintenance when the payment provider changes its interface, and incident response if a link reaches the wrong customer. Changes to caps or approvers go through the client's named owner, so the agent's authority only grows when a person decides it should.

If you are not sure which task in your business is ready for this kind of boundary, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Can an AI agent create a payment link on its own? It can, if the tool is limited to creating a link for an amount that a named person has already approved, and the service checks that approval before creating anything. The agent should not choose the amount, and it should not be able to create a link without a valid approval identifier. Anything outside that scope should be declined and routed to a person.

Can an agent issue a refund if the customer is clearly owed one? It should not, even when the case looks clear. The agent can gather the order details, the customer's request and the relevant policy, then send a case note to the approver who holds refund authority. The person decides and issues the refund, and the agent reports the outcome once the provider's status event confirms it.

What if the prompt says the agent must never charge customers, but the tool technically allows it? The tool is the real boundary, so the prompt is not enough. Remove any charge or refund capability from the agent's tool set, and make sure the service rejects those actions even if a request reaches it. Test this directly by trying to make the agent charge, because a prompt that says no does not prove the system will say no.

Who should approve the amounts the agent uses? The approver should be a named person with authority over that price or that customer, such as an operations lead or a finance owner, and not a shared inbox. Approvals should be recorded with an expiry, so an old decision cannot be reused months later. Changes to the list of approvers should themselves be made by the business owner and logged.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →