Zambo, the verifiable work layer for AI
Enterprise AI agent approval workflow
An approval workflow is only as useful as its record. Here the approver decides on the handover's work record itself, the decision is stored with its reported reviewer details and timestamp, and the approval record exports as PDF through the private client link. This is a record and decision flow, not an enterprise administration suite.
Follow a recorded approval
Open the chain: 5e47dc93-fc76-4fd1-9357-2d0ad0b42c96 has ten steps and the producer-reported Claude label, then 25555933-c043-4e1a-9708-696d117ccb58 continues it with the producer-reported ChatGPT label. The second handover records acceptance at 2026-10-18T11:12:31.042Z, with decision receipt 7a15df28 and release_hint ok_to_release. AI labels are self-reported, not proof of which app ran.
After a decision, the holder of the private client link can download the approval record PDF at /handover/<id>/approval.pdf. The public example does not give you that capability. Approval PDF and JSON are free after a decision; sealing, including producing the dispute pack, requires a pass.
Check the controls that actually exist
- The creator's owner key cannot Accept or Flag its own work. Decisions need the separate client capability. Transferable client links and self-reported reviewer details do not establish identity.
- A pending or flagged record maps to release_hint hold. Only recorded client acceptance maps to ok_to_release. This is a release hint for your process, not custody or an automatic movement of funds.
- Unchecked work stays on the record. The open example names aa725980 and 75253fbc separately from its saved-byte checks.
- Sealing, including dispute pack production, requires a pass. The public page, browser checks and Accept or Flag remain free.
bd7e5325-33fd-4c5e-b029-76b781fa2f51 is an open public handover with five steps. Its work-level not checked list names aa725980 and 75253fbc. Those are work gaps, not failed saved-byte checks. The public page shows steps, verifiable receipts, and what the record proves and does not prove. Its public view is read-only.
Use the client decision path
The client can Accept or Flag through their private link. A recipient agent holding the client capability can use handover_accept; without that capability, picking up work is continuation, not client approval. The HTTP path is POST /api/v2/handover/<id>/respond with the client capability in Authorization: Bearer, not a query-string credential.
For a chain, pick up the preceding handover with handover_accept and use continue_from in the same job session. For recurring work, handover_series creates a living handover link. A client decides Accept or Flag through their separate client link. Public example pages cannot record your decision.
What this record proves, and what it does not
A receipt proves the saved result was not changed since it was saved. It never proves the answer was right and never proves an external event happened.
A handover does not establish who produced the work. AI labels and reviewer details are reported claims, not verified identities. Spot-check important outside outcomes separately.
Questions and answers
What is an enterprise AI agent approval workflow?
It is a review flow that records the work evidence, the approver's reported details, the decision and its time together. Zambo handovers provide that record and a client decision link, not an admin console, organizational roles or identity verification.
How is approval recorded for AI agent work?
The holder of the separate client link records Accept or Flag with a reason for a flag. A recorded acceptance returns release_hint ok_to_release; pending and flagged return hold. Approval PDF and JSON use the client capability after a decision. Sealing, including producing the dispute pack, requires a pass.
Add Zambo to your AI
Connect the free MCP at https://zambo.dev/mcp in Claude, ChatGPT, Gemini, Cursor, Grok, Windsurf, or VS Code. No key is needed. The first receipt arrives with the first call through Zambo. Adding a URL does not automatically install a connector; use your client's MCP settings.
More AI agent answers · Try the demo · Current pricing and tool-specific limits