Direct answer

My AI said the task was done but nothing actually happened. What do I check?

Check the side effects, not the summary. Open the place the work was supposed to change, the file, the inbox, the record, and look for the change itself. If the job ran through Zambo, open its receipts instead: every step that ran has one with its saved result, and a step that never ran has none.

See the instrument first: the live demo runs one real call and hands you its verifiable receipt, no signup.

By Brennan Zambo, creator of Zambo. Updated October 17, 2026.

Why an AI says done when nothing happened

There are four usual causes, and they need different responses. First, the AI stopped early and summarized the plan as if it were the result: the reply describes the finished state because the plan was written in finished language. Second, a tool call failed and the error got smoothed over in the retelling, so a refused or timed-out step reads like a completed one. Third, the AI did a different, easier version of the task and reported the version it did as the one you asked for. Fourth, it never had the access the task needed and filled the gap with plausible text instead of saying so. All four produce the same confident sentence. Only a record separates them.

The five-minute check

One: name the change the task was supposed to cause, in one sentence, with the place it should be visible. Two: open that place yourself and look; do not ask the AI to look for you. Three: ask the AI for the step list behind its summary, which tools ran, in what order, and what each one returned. Four: compare the two. A step with no visible effect and no returned result is a step that did not run, whatever the summary says. Five: if the work ran through a recording connection, open the receipts and end the comparison there, because the record was written while the work ran, not reconstructed for your question.

What the honest states look like in the wild: a second public handover, bd7e5325, still reads pending, with the decision on hold and steps marked not_checked in plain view. A record that can say not checked is a record whose checked means something, and reading one is the fastest way to see which steps of a job nobody has actually verified yet.

What to ask for next time

Ask for the receipt, the same way you would ask a courier for the delivery record instead of their word. On Zambo that is a link: each job's steps carry verifiable AER-1 receipts, and the day's work collects on one page you can read in two minutes. And if someone hands you a receipt id from work you did not watch happen, you can check that single receipt yourself at zambo.dev/verify before you answer them. The sibling page on this hub, why does my AI say it is done when it is not, works through the causes in depth. This page is the checklist for the moment it just happened to you.

The honest part

A record can only cover work that ran through the recording connection. Work an AI did somewhere else leaves no receipt, and no tool can reconstruct it honestly after the fact. What the record changes is next time: the false done becomes visible the moment it happens, because the missing step is a missing receipt in a sequence you can read.

Frequently asked questions

My AI said the task was done but nothing actually happened. What do I check?

Check where the work was supposed to land, not what the AI says about it. Name the change the task should have caused, open that file, inbox, record, or account yourself, and compare what you find with the step list behind the summary. If the job ran through Zambo, open its receipts instead: every executed step has a verifiable receipt with its saved result, and a step that never ran has none, so the gap shows itself.

Why does my AI say it's done when it isn't?

Because the summary is generated from the plan and the conversation, not from a record of executed steps. An AI can stop early, hit a failed tool call, substitute an easier version of the task, or lack the access the task needed, and still produce a finished-sounding report, because nothing in the chat forces the report to match what ran. A per-step execution record is what forces that match.

How can I tell whether a step actually ran?

Look for its side effect and its saved result. A step that ran changed something outside the chat: a file written, a message sent, a record updated, a result returned by a tool. A step that only exists in the summary changed nothing. On a recorded job the test is simpler: the step either has a receipt with the result saved at run time, or it did not run through the connection.

Can I check work that already finished somewhere else?

Only partially, and honestly: you can inspect the end state yourself and question the AI about its steps, but there is no receipt to recompute for work that never ran through a recording connection. Anyone who offers to certify that past work after the fact is selling you a summary with a stamp on it. The fix is forward-looking: route the next job through the connection so its record exists.

How do I stop false dones from happening again?

Make the record the definition of done. Route the work through Zambo so each step writes a verifiable receipt as it runs, review the day on the Today page instead of in chat summaries, and hand client work over as a handover the client accepts or flags step by step. When done means a record exists, a summary alone stops being able to fake it.

Your AI said "done." For $0.99, prove it: https://zambo.dev/day-pass/

In the middle of it right now? The main site answers the same moment directly: my AI said the task was done but nothing actually happened, what do I check, with the five-minute check in order.

Related