The problem with confirmation is it trains the human to always press "yes". You need a system that only activates if something really unusual or wrong.
This is similar to how Github forces you to type the name of a repository when you try to delete it.
You can also configure a threshold such that the "hardcore" prompt kicks in only if the system detects that the user is trying to do something unusual, i.e. sell a stock for much cheaper than it is.
Back in the old days, I feel the broker would have gone to the market makers, said there was an error, and gotten out of the trade.
Which is why I said you would enable it based on a threshold and/or only on unusual threads. These systems are capable of extremely complex decision-making at very high frequency. Surely something like "sell 610,000 shares for 1 yen" would have stood out as strange.
That way, a user doesn't have to know what value to use. You can also set this once in settings and never see it again, except as a report.
In the medical world, many people have been killed by mistakes in prescriptions, despite everything going through computers and multiple humans checking it.
The problem was that the machines were too sensitive to mistakes and trained the humans to ignore warnings. They also didn't distinguish mild warnings from extreme warnings. E.g. a prescription being 2x larger than recommended, or 100x larger than recommended, which are very different situations. A well designed interface might have different shades of red for different levels of severity. To give an indication if something is really wrong.
In a way she is correct, I saw her many times freaking out at a dialog box and calling me to "fix it", and it was just some absolutely useless "warning", but I also saw her many times clicking "ok" on a important message without reading and then wondering why things broke.