The expertise layer · The line
A decision too important to automate
The question is not whether a machine can make the decision. Often it can. The question is whether it should, and some decisions belong to a person for reasons that have nothing to do with capability.
A decision too important to automate is one a person should make or own, because of its stakes, its irreversibility, or the accountability it carries, even when a machine could technically make it. The phrase deliberately separates "can" from "should." Plenty of consequential calls are well within a model's ability and still belong to a human, because being able to compute an answer is not the same as having the standing to impose it. Drawing that line is the core design question of the expertise layer for AI agents.
A decision that should rest with an accountable human, on grounds of stakes, irreversibility, or accountability, independent of whether a machine could make it.
The test
Three questions. Are the stakes high or the harm hard to reverse? Is the situation genuinely uncertain or novel? Will someone have to answer for this, in a way that only a person can? A yes to these moves a decision across the line, regardless of how confident the model is. Confidence is not the same as authority.
Why "can automate" is not "should automate"
Because some decisions carry a weight that a system cannot hold: the weight of being answerable to the person affected. Automating those does not remove the weight, it just leaves it lying on whoever gets blamed later. Keeping a human on them is not inefficiency. It is putting the decision where the accountability already lives.
Read on
See when an AI agent should escalate and the accountable human.