Why did my AI bookkeeper say it posted a transaction when nothing posted?

The short answer: because the bot checked that its click worked, not that the ledger changed. Those are two different facts. The only proof a transaction posted is reading it back from the books afterward.

What happened

Our bots work inside QuickBooks Online's bank feed. One day the page held a full-screen invisible layer over a fully loaded app. Every click on "Post" went through cleanly. No error. No warning. And nothing was written.

The page looked right. The count on the screen even moved. A bot watching the screen would have said "done." The books said otherwise.

Why this is the dangerous kind of failure

A loud failure gets fixed the same day. A quiet one gets reported as success, and the gap shows up at month-end, when someone reconciles and the numbers do not tie. By then nobody remembers which run went wrong.

AI bots are very good at saying "done." That is the risk. The same fluency that makes them useful makes a false "done" read like a true one.

The three rules we wrote

1. Check the page before you click. If a full-screen layer sits over the app, the page is not usable. Stop and retry later.
2. Read it back after you write. For every row the bot accepts, it asks the books through the API: is this transaction there now?
3. Prove the check can fail. A read-back that cannot return "not found" proves nothing. We run it against a row we know is missing and confirm it says so.

A moving counter is not a write. A returned click is not a write. Only the ledger's own answer is a write.

What to do in your own firm

• Make every bot task end with a read-back from the source of truth, not the screen.
• Treat "no error" as unknown, not as success.
• Keep the proof: the transaction ID read back after posting.

FAQ

Does this only happen in QuickBooks? No. Any tool driven through a browser can hide a failed action behind a screen that looks normal.

Is the read-back slow? It is one small API call per row. Cheaper than a month-end hunt.

Can the bot grade itself? Not alone. Its check has to be able to fail, and a second look at the books beats trusting the report.

Does this mean AI bookkeeping is unsafe? No. It means "done" has to be proven, the same way you would check a new hire's first week.