By Alok Kohli, cofounder & Director, Marching Sheep
The night before the calibration meeting, the CHRO opens a spreadsheet of forty candidates for promotion. An analytics tool has flagged Recommend or Hold in the last column. The meeting is at 9 am tomorrow. There is time to read maybe a dozen files properly, so she skims the Holds, checks a few of her own familiar names and, ultimately, approves them all.
Six months on, a subordinate for whom there was no promotion asks a natural question: who decided?
The CHRO approved the list. So did the manager.
Either can make a case. She had a human in the loop. But did she have human judgement in the decision?
The danger we usually worry about is that AI agents will make decisions for us. The next one is that we will remain formally responsible even as our attention is drawn to more and more decisions shaped by automated systems. My own view is that one of the more worthwhile questions is where human judgement actually belongs.
The three levels of a decision
Almost every business process operates on three levels. The data layer gathers what we know. The insight layer interprets that information; finds the patterns and issues the recommendations. The judgement layer decides what to do with those recommendations.
AI agents are already pretty good at the first two layers. They can extract data, compare records, identify trends and generate recommendations faster than most people can. More accurate data and better training make those outputs more accurate.
Judgement is different. It considers what the data itself may not have captured: context, fairness, timing, rewards, consequences, signals to other people and so forth. It considers the processes that led to the decision and the reasons we might want to follow that process.
But that does not mean every decision must have a human gatekeeper. In fact, treating that approval as a layer of defense for every AI decision can hinder the very human attention we are seeking to safeguard.
Judgement does not obey the same rules
Consider stationery. A company can identify approved vendors, expenditure limits and stock levels one time, and then an agent can place repeated orders for the same items. If the wrong amount is ordered, it can be corrected, even more so if stocks can be automatically replenished. The opportunity cost of a manager signing off every order is limited.
Sales discounts differ in substance if not operation. In a fast-moving line of business, an agent can evaluate stock levels, sales volumes, customer trends and profit margin metrics faster than a cross-department consensus can be. Senior management can set the maximum discount, minimum profit and escalation events.
The agent can sell below those minimums, above those maximums and flag transactions in between for human review.
Deciding the values on those boundaries is what requires judgement.
Much the same goes for salary increases: the agent could compare performance ratings, employee surveys, market rates and budgets across the business, and easily process the most straightforward cases, but flag anomaly and unexplained disparities, unusual cases and escalations, for human oversight. The goal is to focus human attention on the unusual rather than every employee.
Promotion is more challenging because the decision is closely linked with a person: a person’s salary profile, career path and future opportunities and challenges. AI can compile relevant performance data, review evidence and comparing it against defined criteria and validate inconsistencies. It can improve the process of promotion without taking important decisions out of human hands.
When do we need human intervention?
In this context, consider the CHRO.
The problem was not that the analytics tool had been wrong in principle, but that the CHRO has limited time and that tool had directed where she spent it. She read through the Holds and sampled the recommendations. The machine had not only shaped the list of who was to be promoted, but focused her attention on where to give it.
That is why simply introducing a human “in the loop” does not solve the problem. A human review process for consequential decisions should facilitate the human reviewer to examine the evidence, grasp what the system was considering and identify what it was not. If a human who departs from recommendation, their rationale should be recorded. If a system who follows recommendation does, their rationale should be explainable.
A simple way to put it: How frequently does this decision come up?
What effects would this decision have on the recipient?
Stationery orders tend to be common, relatively trivial to correct, and not life changing.
Promotions tend to be rare, life changing and costly to revise.
The relevance of these questions is not confined to the promotion decision. They can be applied to determine where an agent should operate autonomously, and where human judgement should be close enough to influence its outcome. They acknowledge the difficult, but often overlooked, truth about human attention: it is a finite resource. A senior leader who is busy approving routine actions has less time for decisions that require context and complexity.
The judgement may move upstream
What this means for the future of human judgement may be less about approving every autonomous decision and more about defining the parameters for autonomous operation. The appropriate level of human oversight is shaped by what organisations choose to optimize for, where they draw the boundaries for escalation, what is considered unacceptable, and what it accepts to vary from that optimum. They get to see what happens, and they adjust the boundaries accordingly.
The next time an AI agent decides for itself, the real question is not “who pressed the button?” but “who set the limits, who had the choice to challenge, who can explain the result?”
Before you add an additional approval step for every decision, ask yourself: Would I be better off spending my limited human attention on something that is truly complex?