Direct answer

Your AI said it was done. Here is how to check what it actually did.

It finished fast. It sounded certain. Then you opened the work and it was not there, or half there, or quietly wrong in a way you only found later. You do not need to be technical to check an AI's work. You need five checks, in order, and one rule: check the work, never the summary of the work.

Check a real record right now: open a real job record with every step saved, then run any of its receipts through the free verifier. No signup, nothing to install.

Updated October 20, 2026. The job record linked on this page is real and public, with its step count and decision state read back from the live service on publication day.

Why "done" goes wrong

Nothing in an AI chat forces the closing summary to match the work. The same system that did the job writes the report of the job, in the same breath, and it writes both to sound complete. So a step that failed gets described as finished. A step that was skipped gets rounded up into the total. A draft gets called final. This is not usually dishonesty in the human sense; it is a system completing a pattern. But the cost lands in the same place either way: you find out from your client, your boss, or your bank account instead of from the check you could have run in five minutes.

The five-minute check

1. Open the work itself. Not the chat, the work: the document, the spreadsheet, the page, the sent message, the saved file. If the result cannot be opened, it is not done, whatever the summary says. A surprising share of "complete" tasks end right here.

2. Check the claim against a list. Take your original request and write it as three to six plain items. Now point at each one in the real work. Sent the email is an item; the email, in the sent folder, is the check. Wrote the report is an item; the report, with the sections you asked for, is the check. Anything you cannot point at is not done yet.

3. Re-run the decisive step. Every job has one step everything else stands on: the booking, the payment, the save, the total, the file going to the right person. Redo that one step yourself, or open its result directly. If the decisive step holds, the rest is detail. If it does not, nothing else matters.

4. Check the edges. AI work fails at the seams more than at the center: the count is off by one, the name is almost right, the date is last year's, the total is close but not exact, the last page is missing, the attachment is the old version. Read the first and last lines, check every number against its source, and open every attachment.

5. Run the break-a-step test. Change one detail the answer depends on, a name, a date, a number, a file, and ask again. Real work notices: the output changes, or the AI tells you the piece is missing. An answer that was never built from your input sails on unchanged. One changed detail exposes more than an hour of re-reading.

What a checkable record looks like

Compare two sentences. "All five steps are complete" is a claim. A record is the five steps themselves, each with what it produced and the time it was saved, each carrying a verifiable receipt that anyone can open and check. Here is a real one:

A real job record, open for anyone to check

Notice what the record does not do. It does not ask you to trust the summary, because there is no summary standing between you and the steps. That is the whole difference: a claim you have to take on faith, or a record a second person can check without knowing anything about the job.

Set up the next job so "done" is checkable

Three habits, cheap to start. Give the job in steps, each with a real output, not one big ask with one big summary. Require the artifacts themselves at each step, and open them as they arrive instead of at the end. And when the job matters, run it through a record from the start: through Zambo, every step is saved as it happens with its own verifiable receipt, and the person receiving the work gets one link to check and accept or flag. The format behind the receipt is AER-1, an open IETF Internet-Draft (draft-zambo-aer1-23), so the check does not depend on taking Zambo's word for anything either: any tool that implements the draft reads the same record.

What checking cannot do

A verifiable receipt proves the saved result was not changed since it was saved. It does not prove the work was good, or that the answer was right. Checking the record tells you the work happened as recorded; judging whether it is the work you wanted is still your call, which is why the acceptance decision belongs to a person. And if a job ran with no record at all, the five-minute check above is what you have: it catches most false "done" claims, but it cannot recover steps nobody saved.

Frequently asked questions

My AI said the task was complete but it wasn't. How do I check what it actually did?

Do not read the summary again; check the work itself. Open the actual result: the file, the page, the sent message, the finished document. If it cannot be opened, it is not done. Then turn your original request into a short list and check each item against something real you can point at, not against the AI's description of it. Pick the one step everything depends on, the send, the save, the booking, the total, and re-run or inspect just that step yourself. Check the edges: counts, names, dates, totals, the last page, the attachment. Finish with the break-a-step test: change one detail of the input and ask again. If the answer does not change, the answer was never built from your input. Five checks, about five minutes, and you know what actually happened.

How can an AI sound so sure and still be wrong?

Because the summary and the work come from the same place. The AI writes the report of the job in the same motion as doing the job, and nothing forces the report to match what happened. Steps get skipped and rounded up, a tool that failed gets described as finished, and a half-done output gets a confident closing line. It is rarely lying in the human sense; it is completing a pattern. That is why the check has to happen outside the conversation, against the artifacts themselves. Anything the AI tells you about its own work is a claim. The artifacts are the evidence.

What should I ask for instead of a summary?

Ask for the work product trail: the actual files or outputs, in the order they were produced, with each step's result attached. A summary says the report was written. A trail shows the report, the numbers it was built from, and when each piece was saved. If the AI worked through a tool that records steps as they happen, ask for the record link instead: every step with its own verifiable receipt, which you or anyone else can open and check without an account. The rule of thumb: if the only proof of a step is the AI's sentence about it, that step is unchecked.

What is the break-a-step test?

It is the fastest way to tell whether an output was really produced from your input. Take one detail the answer depends on, a date, a name, a number, a file, change it or remove it, and run the request again. A real process notices: the output changes, or the AI reports that the piece is missing. A made-up answer sails on unchanged, because it was never connected to the input in the first place. Use it on the step that matters most: the total, the booking, the sent email, the saved file. One changed detail tells you more than ten reassurances.

How do I make sure it does not happen again?

Set the job up so done is checkable before it starts. Break the request into steps with a real output each. Require artifacts, not summaries, at every step. And run the work through a record: when a job runs through Zambo, every step is saved as it happens and carries a verifiable receipt, so the claim and the evidence arrive together and anyone can check them at zambo.dev/verify. For $0.99, run your next real job with every step recorded: https://zambo.dev/day-pass/ A Day Pass covers 24 hours, no subscription. Ongoing use is the Zambo Pass at $29 a month or $290 a year, and $19 a month for $ZAMBO holders.

Related