Step 1. Ask for receipts, not summaries
The single most powerful sentence in any AI work dispute is: "Please send the execution receipts for each task." A summary describes what someone says happened. A receipt is a checkable record of a specific tool execution: a stable ID, a timestamp, what ran, and a SHA-256 commitment over the exact result.
If they can produce receipts, you are most of the way to resolved. If they cannot, you have learned something important before spending another dollar: there is no evidence trail, only testimony.
Step 2. Know what a valid receipt looks like
Real proof
- One record per action. Each tool the agent called gets its own receipt with its own ID and timestamp.
- Checkable by you. You can open a public verifier, paste the receipt, and confirm the fingerprint matches. No account on their system required.
- Bound to the moment. Timestamps and sequence positions tie each receipt to when the work actually ran.
- Open format. The record follows a published spec (AER-1, an IETF Internet-Draft), so verification does not depend on the vendor staying in business.
Theater
- A chat transcript. The agent saying "I completed the task" is the claim, not the proof.
- Screenshots of a dashboard. Controlled by the party being evaluated. Recomputes nothing.
- A PDF summary written after the fact. A memo about the work, not evidence of it.
- "Our system confirms completion." The vendor attesting to its own work is the dispute restated as the answer.
Step 3. Send this message
Copy, paste, adjust the names. Calm, specific, impossible to misunderstand:
Hi [name], Before I approve this invoice, I need execution evidence for the agent work, not a summary. For each task billed, please send: 1. The execution receipt for every tool call (ID, timestamp, tool name, and result fingerprint per call) 2. A link where I can independently verify at least one receipt without an account on your system If your stack issues AER-1 receipts, the receipt IDs alone are enough. If it does not, please say so plainly and we will agree on an evidence standard before the next milestone. Thanks, [you]
Notice what this does: it makes "trust me" expensive and proof cheap. Serious operators will respect it. Anyone who stalls, gets vague, or sends another summary has told you everything.
Step 4. Verify what comes back
When receipts arrive, check them yourself. Open the public verifier at zambo.dev/verify, paste a receipt ID, and confirm the fingerprint matches. Spot-check several across the billed period, not just the first one.
What you are confirming: the saved result was not changed after the fact. What you are not confirming: that the agent's work was good. A receipt proves integrity of the record, not correctness of the model. Price the work on quality; settle the dispute on evidence.
Step 5. Set the standard going forward
Disputes are expensive; standards are cheap. Put one line in every future statement of work:
Each billed agent task ships with its execution receipts (AER-1 format or equivalent checkable record), delivered with the invoice.
From then on there is nothing to argue about. The receipts arrive with the bill, or the bill waits.
Run it yourself in 30 seconds
The fastest way to understand what you are asking for: hold a receipt yourself. Type any task, get a real verifiable receipt, no signup:
The 30-second demo. Then send the template above with confidence.