Bolt human approval onto every step of an AI workflow and you'll have a queue, not a workflow. The team will route around it within a quarter.
Skip human approval everywhere and you'll have a great demo until the first time the model gets something embarrassing wrong, and then you'll have a moratorium.
The work is in deciding which decisions actually need a human, and being honest about it.
The two questions
For every decision the AI workflow makes, we ask two things:
- What's the cost of getting this wrong? Reputational, financial, regulatory, customer-relationship. If it's high, the decision needs a human at the gate. If it's low, the decision doesn't.
- Can a human catch the error after the fact? If yes, post-hoc audit beats pre-hoc approval, you keep the speed and you still catch the misses. If no, the human has to be in the gate.
What this looks like in practice
Drafting an internal summary? No human in the loop. Operator reads the output, edits if needed. The cost of a bad draft is a five-minute rewrite.
Sending an external email to a customer? Human in the loop, every time. The cost of a bad email is a customer-trust event you can't unsend.
Classifying low-stakes inbound tickets? No human in the loop. Audit a sample weekly, retrain on the misses.
Triggering a refund or a billing change? Human in the loop, full stop. The cost of automating a billing mistake at scale is the kind of incident that ends the engagement.
The pattern: the bigger the blast radius of "wrong," the closer the human sits to the decision. The smaller the blast radius, the further away, but never absent.
What "without the drag" really means
It does not mean fewer humans. It means fewer pointless approval steps. The workflows that fail aren't the ones with too much oversight; they're the ones with oversight in the wrong places.
Approval steps don't slow workflows. Approval steps in the wrong places do.
If the team is bypassing your human-in-the-loop step to get the work done, the step is wrong. Either the cost of error is lower than you thought, or the audit can replace the gate.
That's a design decision. Not a default.