The seven pieces of evidence for one agent payment
When an agent handles money, the receipt for the payment alone proves almost nothing. Anyone reviewing the spend later, you, a finance lead, a client, an auditor, needs to see why the payment was allowed to happen at all. Keep these seven pieces together for every payment an agent makes or triggers:
- The instruction. What the agent was told to do, in the words it was given, saved before the run starts. Most money disputes are really instruction disputes.
- The spend cap in force. The limit on the connection at the moment of the call, from spend_limit_set, and the running total from spend_limit_status. A payment inside its cap is a different fact from a payment nobody had bounded.
- The approval state. Whether a person approved this spend, approved the run it belonged to, or never saw it. The approval record names the approver and the time, or shows plainly that no approval step existed.
- The payment reference. The reference the payment system returned: the transaction id, the invoice number, the order id. This is the join key between the work record and the money record.
- The execution receipt. One receipt for the call that made or triggered the payment: what was called, what went in, what came back, and when.
- The exception record. Every rejected or refused call around the payment, including the call that hit the spend limit. Attempts are evidence: they show the fences working, or show what the agent tried before it succeeded.
- The reconciliation note. One line, written while it is fresh, tying the payment reference to the delivered work it paid for and to the handover the receiver accepted.
The setup for the cap itself is the agent spending limit guide, and the incident version of this problem is stopping an agent from overspending.
Why the receipt and the payment record need each other
A bank or wallet record proves money moved. It does not prove which instruction sent it, which cap allowed it, or which work it paid for. A Zambo receipt proves the saved record of the call was not changed since it was saved. Neither record replaces the other. Together, joined by the payment reference, they answer the only questions that matter after agent spend: what was it told, what bounded it, who approved it, what executed, and what did the business get. Recall the definition: an agent receipt is a record of one ai agent tool call that anyone can check: it keeps the exact saved bytes and their sha-256 digest, and it proves the result was not changed since it was saved, not that it was right. Anyone can recheck the receipt side of that chain at zambo.dev/verify with no account.
The routine that keeps the evidence complete
- Set the cap before the run, per run. A cap chosen while everyone is calm is the cap that holds at 2 a.m. The spend tracking guide shows the running totals.
- Gate the big calls behind a person. Routine spend under the cap can flow. Spend above a threshold you choose waits for a recorded approval, the pattern in the human approval record.
- Reconcile weekly, in minutes. Match the week's payment references against the week's receipts and handovers. A mismatch found this week is a question. A mismatch found at year end is an investigation. The Q4 audit guide shows where this evidence lands at year end.
- Keep the exceptions. Never clean up the refused calls. The refusal record is the evidence the cap acted: its receipt proves that saved record was not changed since it was saved.
Zambo records the work and the spend state around it. Zambo does not hold your funds and does not move the money.
Set it up before the agent touches money
- Route the agent through the one endpoint. Connect the AI to https://zambo.dev/mcp so its calls pass the cap and mint receipts. Start at zambo.dev/add: 20 free calls per day, no install.
- Write the instruction down as a record, not as a chat message that scrolls away. The instruction is exhibit one.
- Set the spend limit and the approval threshold before the first money call, and note both in the workflow.
- Save the payment reference into the handover when the work is delivered, so the money trail and the work trail close together.
Put the evidence trail in place before the next payment run: add Zambo to your AI. Start with 20 free calls per day; the Day Pass is $0.99 USDC on Base for 24 hours of full use, and the Zambo Pass is $29 for 30 days, one payment.
Frequently asked questions
What evidence should I keep when an AI agent handles money for a business?
Keep seven pieces together for every payment: the instruction the agent was given, the spend cap in force at the time, the approval state, the payment reference the payment system returned, the execution receipt for the call, the exception record of any refused calls around it, and a short reconciliation note tying the payment to the delivered work.
Does a payment receipt prove the agent was right to spend the money?
No. The payment record proves money moved, and the execution receipt proves the saved record of the call was not changed since it was saved. Whether the spend was wise is a judgment you make from the full set: the instruction, the cap, and the approval state are what let you judge it fairly.
What proves an agent stayed inside its spending limit?
The cap recorded by spend_limit_set, the running total from spend_limit_status, and the receipts for the calls themselves, including any call the limit refused. A refused call still mints a receipt, so the record shows the cap working, not just the spend that landed.
Who should be able to check this evidence?
Anyone you hand it to. Each execution receipt can be rechecked at zambo.dev/verify with no account, so a finance lead, a client, or an auditor can verify the work side of the trail without access to your systems. The payment side stays in your payment records, joined by the payment reference.
What is an agent receipt?
An agent receipt is a record of one ai agent tool call that anyone can check: it keeps the exact saved bytes and their sha-256 digest, and it proves the result was not changed since it was saved, not that it was right.