AI produces more than you can check

AI produces text that looks finished. Fluent, no typos, confident in tone. That is exactly why mistakes slip through: last year's price, a promise you cannot keep, a client name lifted from a different quote. When a colleague spent an hour on something, you read it with a critical eye. When it appeared in ten seconds, you skim it and hit send.

We recently saw a draft email to a client that mentioned a discount nobody had ever agreed to. It was written so matter-of-factly that no one questioned it. Nothing went wrong, but only because someone happened to read it one last time. That is not a process. That is luck.

Not everything needs checking. An internal meeting summary can go straight through. Anything leaving the building with an amount, a commitment, a name or legal wording gets read by a person, ideally someone who knows the subject. Write that line down, or it will drift a little further with every busy week.

Then make checking as light as you can. Pull numbers from your own systems instead of the model's memory. Have the AI say where a claim comes from. And give whoever reviews it a fixed list of four questions: is the amount right, is the name right, is the date right, can we actually deliver on this promise?

Teams that adopt AI and book the time saved straight away tend to forget the review. Checking becomes one more thing on top, which makes it the first thing to go. Build it in as a fixed step. Ten minutes of reading on something that used to take two hours is still a big win. Zero minutes is a gamble with your name on it.

This is how we work ourselves: AI writes the draft, a person always sends it. Not because the models are bad, they are surprisingly good. But responsibility does not move over to the software. It is your name at the bottom of that email, not the model's.