The Decision You Cannot Hand Off

A leader sits down with a report an AI system generated overnight. It ranks the workforce by output, attendance, tenure, cost. At the bottom of the list are the names the system recommends cutting first.

This is not one leader. I have watched leaders accept a finished recommendation like this for years. The tool doing the sorting is new. The moment is not.

The list is clean. The reasoning is laid out in plain language. The math behind it is sound, as far as the math goes. And there is a specific kind of relief that comes with a hard call already looking handled. The leader did not have to carry the discomfort of building the list themselves, staring at each name, feeling the weight of what it means for a real person's mortgage or marriage or kid in braces. The system built it. The leader just has to say yes.

That moment, not the technology itself, is what this series is actually about.

A title gives a leader the power to say yes to that list. It does not give them the right to be correct. And it never transferred the weight of what happens to the person whose name is on it. That weight was never the system's to carry in the first place. It was always the leader's, and it still is, whether or not a machine did the sorting.

I wrote in RemAIning Human that artificial intelligence does not multiply moral authority. It multiplies positional authority. The distinction matters more now than it did when I wrote it, because the tools have gotten faster and more confident since then, not slower and more humble. A system can hand a leader a finished answer about a human life in less time than it takes to pour a cup of coffee, and it hands that answer over with the same even, certain tone whether it is right or catastrophically wrong. Nothing in how the answer arrives tells you which one you are holding.

This is not a hypothetical risk. It is already in court. Twenty-six current and former employees sued Meta in 2026, alleging the company used an internal web of AI systems to score and rank employees ahead of a round of layoffs, and that the scoring disproportionately hit people who were on medical, parental, or other protected leave at the time. Meta disputes the claim and says the final calls were made by people, not AI. The lawsuit has not been decided, and I am not telling you who is right. I am telling you what the allegation describes, because it describes the exact failure this piece is about. A system that scored people on activity and output had no way of knowing who was out on leave for reasons that had nothing to do with how well they were doing their job. Whether a human signed off on the list afterward does not change what the list was built on in the first place.

The same blindness shows up at a much smaller scale, and it does not take a lawsuit to find it. A supermarket cashier in Australia with thirty years of experience received a final written warning after an AI monitoring system flagged errors on a sliver of the items she scanned, somewhere near one mistake in every five hundred transactions. The system was accurate about the count. It had no way of knowing what thirty years of reliable work was worth, or whether one error in five hundred is actually a problem worth a warning. It counted. Counting is not the same thing as judging, no matter how clean the output looks when it lands on a screen.

That is the trap in both cases, and it is the trap waiting in your company whether your business ever makes the news or not. Confidence in the output has nothing to do with whether the output accounted for what actually matters. A productivity score does not know that the person at the bottom of the list spent six weeks this year caring for a parent in hospice and still hit most of their targets. A tenure algorithm does not know that the newest hire on the list is outperforming people twice her tenure because nobody adjusted the model for ramp time. The data is not lying. It is just incomplete in a way that looks complete, because it arrived formatted, labeled, and sorted like something a careful person built by hand.

None of this happens because leaders are careless. It happens because the system offers something genuinely hard to turn down: a finished answer that removes the discomfort of having to build the hard call yourself. A leader who has sat across the table from someone they are about to let go knows exactly how much that moment costs. A report that arrives already scored, already ranked, already reasoned out in confident language, offers a way to skip the part that costs something. That offer is not evil. It is just dangerous, because the thing it is removing is the exact thing the job requires.

I have watched leaders accept recommendations like this for thirty-four years, long before anyone called it AI. A spreadsheet ranked by last quarter's numbers can do the same damage a model does now, just slower and with worse formatting. What is different today is the speed and the confidence. A decision that used to take a week of uncomfortable conversations with department heads now takes an afternoon, and the system that produced it never once had to look anyone in the eye. The leader still does. That has not changed, and it is not going to.

So here is the move, and it is not a tool setting or a prompt. Before an AI-assisted decision affecting a person's job, pay, or standing goes forward, the leader reviews it against one question: what does this output not know about this specific person's situation. Not the team. Not the department. This person. The thing happening in their life right now that never made it into any data set, because most of what makes a person's situation what it actually is was never going to show up in a report.

In practice, this is slower than approving the list, and it should be. It means reading every name, not scanning a summary. For a ranking tied to layoffs or restructuring, it means asking whoever is closest to each person on the bottom half of that list, a direct manager, a team lead, someone who actually knows them, one question before any name is finalized: is there anything happening with this person that the numbers would not show. For a performance flag tied to monitoring or output data, it means asking what circumstance produced the number before accepting what the number claims it proves. Sometimes the answer is nothing, and the system's read stands. Sometimes the answer changes everything, and the leader just caught something a model was never in a position to know, because nobody was ever going to tell it.

If the leader cannot answer that question, the decision is not ready to move. Not because the system got it wrong. Because the leader has not yet done the part of the job that was never going to belong to the system. Finding out what the data could not see was always a human responsibility, back when the data came from a spreadsheet and now that it comes from something that sounds a great deal more sure of itself.

This is not a call for slowing everything down out of caution. Some decisions genuinely are as simple as the data makes them look. The discipline is knowing the difference, and that discipline does not come from the tool. It comes from a leader who has decided in advance that speed is not going to be allowed to answer for them on anything that costs another person something real.

What a leader lets speed decide, and what a leader insists on deciding slowly, in full possession of what a model could never know, is the clearest answer available to who that leader actually is. Not what they say in a values statement. What they do the first time a finished answer about a human being lands on their desk looking done.

If you are facing a decision like this in your own company right now, I would like to hear about it. Reach out and let's talk it through.