Why files plus a summary falls apart
Files show the client the destination and hide the road. The summary describes the road in the words of the one who drove it. Neither can be checked after the fact, and both sides know it, which is why handoff day so often turns into a slow negotiation about what was agreed, what was tried, and which version anyone is even looking at.
The failure is not dishonesty. It is structure. AI work happens as a sequence of steps, calls, sources, and small decisions, and the chat window is the only place that sequence ever existed. When the chat closes, the record closes with it. A real handover moves the sequence itself into the client's hands, in a form they can open and check without your account, your session, or your narration.
The handover checklist: six things, not two
A complete AI work handover contains
- The deliverable, named plainly, the actual output of the job.
- Every step taken, each with its own record, so the client reads what ran instead of reading about what ran.
- The sources checked, attached to the steps that used them, so claims show what they rest on.
- The decisions, including the small ones made inside the work, stated as decisions with an owner.
- The not checked list, the parts nobody verified, said out loud instead of smoothed over.
- The acceptance step, where the client accepts or flags the work and the decision is recorded with its time.
Miss any one of the six and the handover leaks. Skip the sources and the client cannot weigh the claims. Skip the not checked list and the first gap they find costs you the trust the other five items built. Skip the acceptance step and you are back to arguing from memory about whether the work was ever approved.
Hand it off in four moves
- Run the work where it is recorded. Work done through Zambo tools leaves a verifiable receipt for every call as it happens, in Claude, ChatGPT, Gemini, Cursor, or any AI that can connect. The record is made during the job, not reconstructed for the handoff.
- Gather the job into one handover. The handover collects the session's steps into a single page: deliverable, steps with receipts, sources, decisions, and the honest gaps. If another AI continues the job later, it continues from this handover and the chain grows instead of restarting.
- Send one link. The client opens it with no account and no install. They see the whole job and can check any receipt on it themselves.
- Let the client decide on the record. Accept or flag, on the same page. The decision, who made it, and when are saved with the work, and the approval record downloads as a PDF for the client's own files, next to the invoice it releases.
A real handover, accepted today
Two AIs, one job, one chain
This morning a job ran in two hops: research steps in one AI, then a second AI continued from the first handover and finished it. The client link opened the finished handover, walked back through the whole chain to the first step, and the client accepted it on the record at 2026-10-18T11:12:31Z. Open it yourself: the accepted handover, with its first hop at the chain root. For the undecided state, this public handover sits open with five steps and no decision yet, which is the pending state doing its job in public.
Nothing in that record asks the client to trust the person who sent it. Every step can be opened, every receipt checked, and the acceptance is a fact with a timestamp instead of a line in an email thread.
When the client pushes back instead
A flag is not a failure of the handover; it is the handover working. The client marks what they dispute on the same record, the steps underneath stay exactly as recorded, and the argument shrinks to the flagged steps instead of swallowing the whole job. When a disagreement needs the full trail in one file, the dispute pack is that same record, sealed for the conversation. Sealing a job is the paid step: a Day Pass or the Zambo Pass turns it on, and once a job is sealed, its dispute pack stays available to send. The full walkthrough lives at client disputes the AI work: send the pack, not screenshots.
What it costs to hand work over this way
The public handover page, the receipt checks, and the client's accept or flag are free. Handing a client real work today, with a private link, your own branding on the page, and the approval record as PDF and JSON for their files, is what the pass is for: a $0.99 Day Pass covers today's deliveries for 24 hours, and the Zambo Pass at $29 a month makes every handover you send carry the full pack. Start with the free loop on your next real job and watch what changes at handoff: the client stops asking for the story, because they are holding the record.
Questions, answered
How do I hand off AI work to a client?
Hand over five things, not two. The deliverable itself. The steps that produced it, each with its own record. The sources the work rests on. The decisions made along the way and who made them. And an acceptance step where the client accepts or flags the work on the record, with the decision and its time saved. Files plus a chat summary covers only the first, and the summary is the one part the client cannot check. A Zambo handover gathers all five onto one link: the work ran through tools that leave a verifiable receipt for every call, the handover collects the steps, and the client decides on the same page.
What should an AI project handover checklist include?
Six items. One, the deliverable, named plainly. Two, every step taken, with the record or receipt for each step instead of a summary of the steps. Three, the sources checked, so the client can see what each claim rests on. Four, the decisions, including the ones the AI made inside its latitude, stated as decisions. Five, an honest not checked list: the parts nobody verified, said out loud. Six, the acceptance state: pending, accepted, or flagged, with who decided and when. If any of the six is missing, the handover is a story about the work, not the work handed over.
Does my client need an account to receive a handover?
No. The client opens one link and sees the whole job: the steps, the receipts under them, and the chain if more than one AI worked on it. They accept or flag on that page. Nothing to install, no account to create, and the decision they make is recorded with its timestamp, so neither side has to reconstruct later who agreed to what.
What if a second AI needs to continue the job after handoff?
That is what the chain is for. The second AI continues from the first handover instead of starting from a pasted summary, and its work lands as the next link in the same chain. The client link opens the latest handover and walks back through every earlier one, so the job reads as one record from the first step to the last, no matter how many AIs touched it.
What does a handover prove, and what does it not prove?
Each step carries a verifiable receipt, and a verifiable receipt proves the saved result was not changed since it was saved. It does not prove the answer was right or the work is good. What the handover gives the client is the checked record itself: what ran, what came back, what was decided, and what was never checked, so their judgment about quality rests on evidence instead of on your summary of it.