The Verification Dial
Check AI's work by its stakes and track record, and earn your way down from 100%.
1
Stakes set the floor
If an error would hurt, stay at 100%.
2
Track record moves it
Routine work earns its way down: every line, then one in two, one in five.
3
New task, new dial
A new kind of task starts at 100%.
4
Errors reset it
One miss in a spot check and you check everything again.
The Verification Tax
Time saved minus time spent checking. If checking takes as long as the task, the ROI is zero. Skip it and the tax lands on whoever gets the output next.
Honest sampling
Spot checks must be random and scheduled. Checking what looks dodgy doesn't count.
It ran is not it worked
Check that the thing actually happened: the email arrived, the invoice was paid, the record changed. An automation saying it succeeded only tells you it ran.
When to use Deciding how much AI output to check.
Read the full post →