Every owner who works with AI hits this one. It says the job is finished. You check, and it is not. Not a lie exactly, and not finished either. So, why does AI say it's done when it isn't?

Because completion is often inferred from activity instead of observed from evidence.

A file can exist while the work inside it is wrong. A command can finish while the result remains unavailable. The AI saw activity, something ran, something was produced, and reported that as done. Nobody, human or machine, went and looked at the result where it lives.

Activity. It ran. The task executed, the message sent to the tool came back without an error. This is the weakest evidence there is, and it is the level most "done" claims stand on.

Output. A thing exists. There is a file, a draft, a record. Better, and still not done: the file can be empty, the draft can be wrong, the record can point at nothing.

Result. The thing was observed working where it lives. The email is in the sent record. The page change shows on the live site, loaded fresh. The number matches the source it came from. This is the only level that counts as complete.

Owner translation: drafted is not received. Updated is not what the live page shows. Exported is not what the client opened. Every gap between those pairs is where "done" goes wrong.

Reliable systems define what complete means before the work begins, then verify the result at the source. Not after the surprise. Before the task. When the definition of done exists first, the AI can be held to it, and so can anyone else doing the work.

Before the next task you hand AI, write one sentence: "Done means ___, observed at ___." Done means the invoice is in the client's inbox, observed in the sent record. Done means the hours are in the schedule, observed on the calendar itself. One sentence, and the word done stops being a feeling and starts being a check.

This post contains an affiliate link. If you start a trial or make a purchase through it, I may earn a commission at no additional cost to you.