What should I look for when choosing a tool to verify AI agent work?

A handover brings the recorded AI work and its verifiable receipts into one proof link. Review the evidence behind each claim, check what was and was not observed, then Accept or Flag.

Last updated: 2026-10-17

Most verification tools are sold as dashboards, and a dashboard can only show you what the vendor's system says. The buying question is different: when a client, an auditor, or a court asks what happened, what can you hand them that they can check themselves? These six tests answer that. They are written for any tool, including this one; run them on Zambo first if you like, the demo is free and the receipt it mints can be checked on the verify page by anyone.

Tests one and two are the foundation. A record created after the fact, from logs or memory, can be edited by whoever holds it. A record saved at execution time, in a fixed format like AER-1 (IETF Internet-draft-zambo-aer1-21), can be recomputed by a checker that does not trust the vendor, which is the whole point.

Tests three to six are where tools quietly fail. Steps that were not checked should be labelled, not hidden. Your records should export in a form you keep after cancellation. Your client should open the proof in a browser, with no account and no specialist software. And the tool should publish what its record does not prove: a receipt proves the saved result was not changed since it was saved, never that the work was correct.

Worksheet: the six buyer tests, scored

Score each tool 0 (no), 1 (partly), or 2 (yes, demonstrated). Test 1, record at execution: ask for a receipt from a live run, not a screenshot. Test 2, third-party recheck: recompute or verify the record somewhere the vendor does not control. Test 3, honest gaps: find a step with unresolved sources and see how it is labelled. Test 4, export: export one project and open it after logging out. Test 5, recipient access: send the proof link to a colleague with no account and time how long the check takes. Test 6, written limits: find the page where the tool states what its record does not prove. Any tool scoring 0 on test 2 is a dashboard, not a verification tool.

Worked example

Zambo, scored in the open: Test 1, receipts are saved as each step executes. Test 2, any receipt can be checked at https://zambo.dev/verify/ and the AER-1 format can be recomputed with the published kit, version 1.5.4. Test 3, handover acceptance views show steps as resolved, not resolved, or not checked. Test 4, handovers and receipts export as files you keep. Test 5, a handover link opens in any browser with no account. Test 6, the limits are stated on the receipt pages themselves: the saved result was not changed; correctness is not claimed.

Frequently asked questions

Is a dashboard with an audit log good enough?

For your own reading, often yes. For proof to someone else, no: the log lives inside the vendor's system and is reported by the same side that did the work. Ask whether a third party can recompute the record without the vendor, and judge by that answer.

What is the single most important test?

The third-party recheck. Take a record from a real run and verify it somewhere the tool does not control. If that path does not exist, every other feature is presentation.

Should I pay before or after running the tests?

After. A serious tool can be tested free: Zambo's demo mints a real receipt with no account, and the free tier covers a first real job. Pay for the Day Pass when a live problem needs a full day of proof, not before.

By Brennan Zambo, founder of Zambo.

Last verified: 2026-10-17.

4. Buying and choosing a tool

Answers

Score your current tool with the worksheet, then run the free demo at https://zambo.dev/demo/ and score Zambo against it.

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

Add Zambo · Handover docs · Real public handover · AER-1