Guide
The usual answer is both, structured properly
Updated
The framing as a choice is mostly false. What decides whether the combination saves money is how the work is split and who checks what.
Split by consequence
Route the high volume, low consequence work to the tool and keep judgement calls with the person. That is the split that survives contact with reality.
Write the routing rule down. An informal understanding degrades into everything going to the person, or worse, everything going to the tool.
Checking is the cost
A tool whose output must be checked line by line has not saved the work, it has moved it and added a step. The saving comes from checking by exception: sampling, thresholds, or confidence-based escalation.
Design the checking regime before deployment, because retrofitting it after a bad output has reached a customer is done under pressure and badly.
Watch the drift
Tools degrade quietly as inputs change and the world moves. Unlike an employee having a bad month, nothing announces it.
Keep the baseline you measured at the start and re-measure periodically. That is the whole of monitoring for most small deployments and it is usually skipped.
Do not shed the capability entirely
If you remove all human capacity for a task, you cannot check the tool, you cannot handle the exceptions, and you cannot go back.
Keeping enough capability to verify and to fall back is a cost, and it is the cost of not being locked in.