What changes, and how?
Warning sign: the fix was a policy, a template or a training session, and the work itself hasn't changed.
What I think
The change that matters is from checking each piece of output to checking the way it was produced. "Is this output correct?" becomes "Can I stand behind how this was made?" You can answer the second question in a way you can't answer the first. You can inspect a process, but you can only ever sample what it produces.
In practice, that means testing the process on purpose, watching how it fails, and changing it so the same failure can't happen again. The answer isn't to review harder. It's to change what produced the problem.
Every piece of AI-assisted work involves four decisions:
Is it worth doing this way?
How good does it need to be?
What will it cost in money?
What will it cost in time, once you count the checking?
You can't skip one. Not deciding is still a decision, just one made without thinking.
Those decisions tell you how much checking the work deserves. A rough internal note needs a glance. A report going to a regulator needs checking back to source. Choosing the lighter touch is a judgement about risk, and it's yours to make.
Here's the part that surprises people, and it's my own reasoning rather than an established finding. Past a certain point, more checks make quality worse. When there are more checks than people have attention for, gaps open where everyone assumes someone else is covering. A good process removes old checks as well as adding new ones. Every new check should come with a note of when it can be retired.
Two things about the checks themselves. A check can stop working and keep reporting that all is well, which is worse than having no check, because it gives you confidence you haven't earned. And a fixed rule, such as a figure that must match or a format that must be followed, is sturdier than asking an AI to check. It runs the same way every time and doesn't share the model's blind spots.
Finally, the unglamorous one that does most of the work: write decisions down. For any decision, ask "where is it written down?" If the answer is "in the chat" or "the model remembers", it isn't written down.
What we don't know yet
Where the tipping point is, at which extra checks start doing harm. I'm confident the effect is real, but I can't give you a number. As of September 2026 I haven't measured it myself, so it's reasoning, and the book marks it that way too.
Whether a risk-based level of checking survives inside a large organisation. When one person sets the level, it's a considered judgement with their name on it. The critical evaluation often falls away when it becomes a box on a form.
What you can do
Before the work starts, answer the four decisions in a line each. Is it worth doing this way? How good does it need to be? What will it cost in money? What will it cost in time, once you count the checking?
Set how much checking it needs by what failure would cost, and write that down too.
Keep decisions where the work lives, not in a chat thread or your memory.