The most consequential decision in an automated workflow is where a person has to sign, rather than which model you use.
Get it wrong in one direction and the workflow is slower than the manual process it replaced. Get it wrong in the other and you have an incident with your name on it.
The wrong way to decide
Most teams place review points by comfort. Whatever makes the nervous stakeholder in the room feel better gets a human.
This produces workflows with a reviewer at every step, which is just the old process with extra software. It also concentrates review attention on the wrong things, because a person asked to approve forty low stakes items an hour stops reading them by item eight. Blanket review produces rubber stamping rather than scrutiny.
The axis that matters
Cost of error, split into two questions:
What does a wrong output cost? The cost at the tail matters rather than the average cost. A misfiled internal note costs a few minutes. A payment to a changed bank account costs the payment.
How quickly is a wrong output caught? Some errors surface immediately downstream. Some sit undetected for a quarter.
Cheap and quickly caught, run it unattended. Expensive or slow to surface, keep a human, permanently, regardless of how good the model gets.
The third question people forget
Is the error reversible?
A wrong classification can be reclassified. A sent email cannot be unsent, and a paid invoice is a recovery exercise involving lawyers. Irreversibility is worth more weight than accuracy in the placement decision, because it changes what a failure costs rather than how often one happens.
Confidence thresholds are not the same thing
A confidence threshold routes uncertain cases to a human. That is useful and it is not the same as a review point.
Thresholds handle the cases the system knows are hard. Review points handle the cases the system is confidently wrong about, which are the dangerous ones. You need both, and a threshold is not a substitute for the second.
A practical test
For each step, finish this sentence: “If this is wrong and nobody notices for a month, the consequence is ______.”
If you can finish it without discomfort, the step can run unattended. If the sentence makes you want to add a caveat, that is your review point.
We use exactly this in scoping workflow automation builds, and it settles arguments faster than any amount of discussion about model accuracy.