x

Before You Give an AI Agent a Wallet, Decide What It Can Buy

By Kristi Shock · November 11, 2026

Before an AI agent can spend money, someone needs to define its authority. I’d want that decision documented before I saw the payment demo.

What can it buy? From whom? For what purpose? Who owns the exception when something goes wrong?

Those are business questions, and the answers need to carry through into the actual controls.

Human judgment, built into the system

Purpose

What can it buy?

Seller

Who can receive payment?

Budget

What is the total limit?

Owner

Who handles exceptions?

A spending instruction needs enforcement

Telling an agent to “be careful” gives the model a preference. A payment system needs enforceable rules outside the model’s own judgment.

OWASP identifies excessive permissions and autonomy as sources of agent risk. Its recommendations include limiting available actions, enforcing authorization downstream and requiring human approval for high-impact actions. OWASP’s guidance on excessive agency

For payments, I’d turn that principle into a short operating agreement between the business owner and the builder.

Write the purchase policy in plain language

Here’s an illustrative policy for a research assistant. The amounts are examples for discussion, not recommended limits for every business.

Purpose: Purchase public business research inputs for one weekly brief. No personal data purchases, subscriptions, trading or transfers to individuals.

Approved sellers: Only the providers and payment destinations reviewed for this workflow. A new seller goes to a person first.

Budget: Up to $0.10 for an approved result, $2 for the brief and $10 for the week. Fees count toward the budget. Parallel requests must share the same remaining allowance.

Exceptions: A price increase, changed recipient, unknown service or unclear usage rights stops the purchase for review. The agent cannot rewrite its own policy.

Ownership: One named person reviews the first live runs and owns the decision to pause or expand the test.

That’s specific enough for a builder to implement and for a business owner to challenge.

Make the approval useful

If the system needs my approval, I want to see the proposed purchase, the reason, the maximum total cost and what will be sent to the seller.

“Approve agent action?” gives me very little to work with.

“Approve up to $0.08 to buy this public dataset result from this approved provider for Friday’s brief?” gives me a decision I can actually make.

I’d also want the person reviewing it to see the remaining budget and whether the agent already tried this purchase. Otherwise, a timeout can turn into two approvals for the same thing.

A tiny payment can still expose sensitive information through the request attached to it. The data-sharing rules deserve their own review, even when the purchase stays below the spending limit.

Plan for an unclear outcome

Suppose the payment settles but the service response never arrives.

I’d want the system to check its recorded payment and request status before trying again. The receipt should connect the approved task, seller, amount, settlement reference and delivery outcome. Store only the information needed for reconciliation, with appropriate access and retention limits.

A refund process also needs an owner. Circle’s disclosures explain that onchain USDC transfers are irreversible; a business should not assume it can recall a payment the way it might reverse an internal database entry. USDC transaction risks

The team needs a way to stop new spending too. A dashboard that reports a problem after the wallet is empty is too late.

Test the uncomfortable cases

Before increasing a budget, I’d ask for a demonstration of these situations:

  • A seller changes its price or destination.
  • Two requests arrive together with almost no budget left.
  • A service times out after payment.
  • A tool response tells the agent to ignore its purchase rules.
  • The person who normally approves exceptions is unavailable.

The last example is easy to overlook. The default should be clear before anyone goes on vacation.

My approach would be to start with simulated payments, review the exceptions, and move to a small real-money pilot only after the controls behave as intended.

I like building systems where people can spend less time supervising routine work. That takes more clarity about the decisions we’re delegating, especially when those decisions move money.

Independent operational commentary. This is a planning framework, not a security audit or legal, tax or investment advice.

Related strategy: Arc Ecosystem Adoption: a 90-day independent strategy concept. Not commissioned by or affiliated with Circle.

KHightower survey daily checkins blog post