"Could this be automated?" is almost always answerable with yes, given enough engineering effort. It's the wrong question to build a workflow around. The better question is whether it should be — and that's a judgment call about the business, not a technical one about the model.
Capability and judgment are different questions
We can build a system that decides which customers get a retention discount and how large it should be. Whether we should build that system depends on things a model can't tell you: how much your team values the relationship data a human notices that a transcript doesn't capture, how a customer would feel finding out the decision was fully automated, and what it costs you culturally if your account team stops being the ones who make that call.
Confusing "can we" with "should we" is how teams end up automating the parts of the job that gave people a reason to care about the work, while leaving the genuinely repetitive parts — the ones nobody wanted to do anyway — untouched.
Three signals that a decision should stay human
The clearest signal is when the decision carries weight beyond the immediate transaction — pricing exceptions for a strategic account, how to handle a long-standing customer's complaint, anything where the relationship matters more than the specific outcome this one time.
The second is when getting it wrong is expensive in a way that's hard to reverse, and the volume is low enough that automating it doesn't save meaningful time anyway. If a decision happens twice a month, the case for automating it has to be about consistency or compliance, not efficiency, because the time saved is trivial.
The third is more subtle: when making the decision is part of how someone on your team develops judgment for their role. New account managers learn what a reasonable exception looks like by making some unreasonable ones early and getting coached on it. Automate that decision away too early and you remove the training ground along with the task.
What this looks like in a real workflow
In a lead qualification flow, for instance, we'll use sales automation and CRM automation for the parts that are genuinely mechanical — enriching a new lead's data, scoring it against firmographic rules, routing it to the right rep based on territory. We won't automate the final judgment call on whether to override the score for a lead that looks weak on paper but has a signal a human noticed in the actual conversation. That override is exactly the kind of judgment that's cheap to leave human and expensive to get wrong by delegating it to a rule.
The same logic applies in HR and recruitment — automate the screening and scheduling, keep the actual hiring decision with the person who has to live with it. The goal isn't to automate less. It's to automate the parts of the workflow that were never really judgment calls to begin with, so the humans on your team spend their time on the smaller number of decisions that actually need a person behind them.