My AI said the task was done but nothing actually happened. What do I check?
Check the record of the run, not the summary of it. Review the receipt and actual deliverable before accepting a handover.
An AI says "done" for predictable reasons. It finished the part it could see and reported the whole. It wrote the success message from the plan it made at the start, not from the outcomes at the end. Or a middle step failed quietly, a file was never written, an email was never sent, a source never loaded, and the final answer smoothed over the gap. None of this needs malice; it is what a fluent system does with a missing result.
The fix is to stop reading summaries and start reading steps. When the work runs through the Zambo connection, every step lands as a verifiable receipt with a state: the result that was saved, and whether its sources were resolved, not resolved, or not checked. The step that did not happen has no receipt. The step that happened without verification says so. You are no longer arguing with a paragraph; you are reading a record.
For work that already ran somewhere else, you cannot rebuild this record after the fact (page: "Can I create a receipt for AI work after it is already finished?" covers that limit honestly). What you can do is re-run the critical steps through the connection, or check the outputs by hand this once, and then make the record automatic for the next job.
Checklist: the five-minute false-done check
1. List the steps the job was supposed to take, from your instruction, not from the AI's summary. 2. For each step, find its record: a receipt, a file, a sent message, a changed value. No record means treat the step as not done. 3. Mark every step whose sources or result were never checked. 4. Compare the final message against the step records, line by line; every claim needs a step behind it. 5. Re-run or hand-check every unproven step before you ship, pay, or report the job as complete.
Worked example
Sample: a report job claims five sections written from five sources. The step record shows four section receipts; the fifth source is marked not resolved and there is no fifth section receipt. The "done" message was written from the plan. The check takes two minutes and finds the exact missing piece, instead of a vague feeling that something is off. (Sample scenario; no live record is cited.)
Frequently asked questions
Why does an AI say done when nothing happened?
Usually because the final message is generated from the plan and the partial results, not from a check of every step. A step that failed quietly or never ran leaves no obvious hole in a fluent summary, so the summary reports success anyway.
What is the fastest single check?
Pick the step with a visible side effect, a sent email, a written file, a changed record, and verify that effect directly, outside the chat. If the effect is missing, the summary is wrong no matter how confident it sounds.
How do I stop this happening on the next job?
Run the work through a connection that saves a verifiable receipt per step, with source states shown. Then "done" is a claim you can test against the step record in minutes, and unchecked steps are labelled instead of hidden.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
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