A model answers a narrow statistical question. A system has users, failure modes, logs, permissions, and a Tuesday when the input looks nothing like the sample. Confusing the two is how intelligent systems become demonstrations.
Skyneski's interest is the system. The question has to be stated. The data the model is allowed to see has to be stated. The estimate has to stay an estimate when it reaches a person.
Where automation belongs
Some work should be automated because the rule is stable and the cost of repetition is real. Some work should not, because the cost of a quiet error is higher than the cost of a review. That distinction is a design decision. It is not a property of the model.
Predictive systems are legitimate engineering. They are not oracles. Publishing a performance number without a method, a dataset, and a limit is a marketing act. We do not do that on this site.
Oversight is a component
If a person is accountable for the outcome, the interface has to show enough for that person to disagree. A score with no inputs is not decision support. It is a request for trust. Applied AI, done properly, is comfortable being questioned.