What evidence should I keep when an AI agent handles money or payments?
Keep four records for any agent that touches money: the authorization (who allowed which actions, up to what amount), the limit at the time, the record of what the agent did, and the refusals, the calls a limit stopped. Review the receipt and actual deliverable before accepting a handover.
Money is where "the agent said it did it" stops being acceptable. The evidence standard is the one a bank, an insurer, or your own future audit would apply: an action is explained by its authorization, bounded by its limit, and proven by its record. Keep those three aligned per agent, per account, and review them together.
The authorization record names the human decision: which agent may move money, through which tools, for which purposes, up to what ceiling, from what date. The limit record is the machine side of the same decision: the spend limit actually in force, set before the work, so a reviewer can compare what was allowed with what the human authorized. Zambo's spend_limit_set and spend_limit_status exist for exactly this comparison on its own surface, and the limit's refusals are first-class records, not errors to clear.
The per-call record is the work record itself: each step saved as a verifiable receipt at execution, so the sequence an auditor reads is the sequence that ran. Keep the three records in one place per agent, export them on your retention schedule, and rehearse the question an incident will ask: show me the call, the limit it ran under, and the human who authorized both. If any leg of that answer needs a chat scroll, the evidence is not ready.
Checklist: the money-agent evidence set
Per agent: 1. Authorization note: human name, actions allowed, ceiling, start date. 2. Limit record: the cap in force, where it is set, screenshot or export with date. 3. Call records: receipts for money-touching steps, indexed by date. 4. Refusal log: every limit refusal, reviewed weekly. 5. Reconciliation: weekly, calls against account statements, mismatches flagged same day. 6. Revocation drill: how this agent's money access is cut, tested quarterly. File the set with the agent's other records; review the whole set monthly.
Worked example
Sample: a shopping agent authorized to buy supplies up to a set ceiling hits its limit mid-run; the refusal is recorded, the run pauses, and the weekly reconciliation shows three purchases, each with a receipt, each under the authorized ceiling, plus one refusal that proves the ceiling was real. That file is what "evidence" means when money moved. (Sample scenario; no live record is cited.)
Frequently asked questions
Does Zambo process or hold the payments an agent makes?
No. Zambo records the work and enforces spend limits on calls to its own tools. Payment itself happens wherever you authorized it (your provider, your wallet, your accounts), and those providers keep their own transaction records, which belong in the same evidence set.
Why keep refusals? They are non-events.
A refusal is the control working, on the record. In any review after a loss, the refusals are what separate "limits existed on paper" from "limits were live and enforced," which is the distinction reviewers price.
What ceiling should I authorize?
The smallest amount that lets the agent do its real job without asking, reviewed monthly against the call records. Ceilings set from imagination drift; ceilings set from last month's receipts fit the work.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
Try a handover
A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.
try it: say 'use zambo to hand this over to <client>' in your ai