What is an AI work evidence packet?
An AI work evidence packet is the complete, checkable record of one job, assembled for a reader who was not there: the handover with its steps and saved results, the verifiable receipt behind each step, the source states showing what each output rested on, and the client's recorded decision. One link carries the whole packet.
The packet exists because evidence scattered is evidence lost. The steps live in one system, the sources in another, the acceptance in a chat thread, and the person asking, a client, an auditor, an arbitrator, receives a tour of your filing instead of an answer. A packet is assembled against that moment: everything the checker needs, in the order checkers read, on one object they can open without your help.
Its contents are specific. The job summary: what was asked, what was delivered. The steps: each stage of the work with its saved result and producer label, gaps labelled rather than hidden. The receipts: one per step, committed at execution (AER-1, IETF Internet-draft-zambo-aer1-21), recomputable at a public verify page. The source states: per step, resolved, not resolved, or not checked. And the decision layer: the client's acceptance or hold, with the decider and the date, recorded on the same object.
Read the rest of this cluster as the packet's applications: the audit offer, the insurer pack, the tax trail, the retention schedule, and the regulator's records room are all this packet, addressed to different readers. Build the packet per job at delivery time, when assembly costs minutes, and every one of those readers becomes a forwarding exercise instead of an excavation.
Anatomy: the evidence packet, piece by piece
1. Cover line: job, client, dates, the one-link address. 2. Summary: the ask and the delivery in three lines. 3. Steps: numbered, each with its saved result and producer label. 4. Receipts: one per step, ID listed, checkable at https://zambo.dev/verify/ . 5. Source states: per step, resolved / not resolved / not checked. 6. Decision: acceptance or hold, decider, date. 7. Export: the packet as files you keep, beside the link. Test: hand the link to someone uninvolved; if they can state what was done, what it rested on, and who accepted it, the packet is complete.
Worked example
A live packet core, public: https://zambo.dev/handover/45434a17-4afc-486a-9646-e689bcee4297 carries the steps, the summary, and the client's recorded acceptance for one job, and its step receipts recompute at the verify page. Read it as a checker would: it is the standard every packet in this cluster is measured against.
Frequently asked questions
How is a packet different from a report?
A report narrates the work in the author's words and asks to be believed. A packet presents the records themselves, each independently checkable, and asks only to be read. Reports are written; packets are assembled, mostly automatically, from records made while the work ran.
Who is the packet for?
Anyone who must judge work they did not watch: the client deciding acceptance, the auditor sampling a project, the insurer reviewing an incident, the arbitrator reading a dispute bundle. The packet's contents stay the same; the cover line names the reader.
Can I assemble a packet for work that was not recorded?
You can assemble the documents that exist, and the packet should label itself as assembled after the fact, with the unrecorded parts named. A packet's authority comes from records made at execution; honesty about which kind you hold is part of the format.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
6. Records, tax, compliance, and retention
- My client is auditing a finished project. What evidence should I give them for the AI work?
- What records should a company keep when employees use AI agents?
- What records would an insurer ask for after an AI agent makes a costly mistake?
- What records do I need for taxes when clients pay me for AI-assisted work?
- How do I prepare AI work records for a regulator or external auditor?
- EU AI Act log-keeping for small teams: what do Articles 12, 19, and 26 actually require?
- How do I run a Q4 audit of everything my AI agents did this year?
- How long should I keep records of AI agent work?
- What records prove I met the EU AI Act Article 50 disclosure duties?
- What AI work records do investors ask for in due diligence?
- I am a bookkeeper. What should I ask clients to send me about their AI work?
- How do I keep a checkable record of AI research done for a legal matter?
- How do I prove what was done for a grant report?
- What AI work records should I pull together for tax season?
- How do I run a mid-year audit of six months of AI agent work?
- What should an internal policy for employee AI work records say?
- What records should I close out before December 31?
- What should I archive when an AI project ends?
- How do I keep a usable record of AI work after changing vendors or cancelling a tool?
- What should I send my accountant from a year of AI work?
- What records should I set up before AI agents touch customer data?
- What should I set up in January so client proof is automatic all year?
- What records should we sort out when an employee who used AI agents leaves?
- What records should an agency keep if a client asks how AI was used six months later?
- How do I build an annual AI work report for clients?
- What records should I keep if a client says AI work infringed their rights?
- How do I organize receipts and AI work records for quarterly estimated taxes?
- What AI work records should I clean up in October before year-end?
- What records should I be able to produce if a client requests all files and receipts behind a deliverable?
- How do I start the year with a clean AI activity record?
- Who owns the record when several employees and agents work on one client account?
- How do I organize AI work records before an annual client audit?
- How do I get my AI work record system back on track after summer?
- How do I answer a client's security questionnaire about our AI work records?
- What records should I keep when an AI agent works in a regulated workflow?
- What should I set up before January if my team adopts AI agents next year?
- What records should I hand my bookkeeper each month for AI-assisted work?
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