Give it few weeks, and they have collected meteics that the AI did the job just as good as the human within that timeframe. A few weeks down the line, the tables have turned, the humans are in charge of notifying the AI if it makes errors. A few weeks later, metrics will show that humans hasn't really notified anything, and humans are removed from their position, leaving the AI fully in charge.

Within a year, a critical halucination or a job which AI lacks experience to handle will happen and a very preventable accident will occur. The owners will shrug it off as something that would definitely happen with humans in charge, and that the reduced costs outweigh the additional risk.