Deciding What to Automate — and What to Keep Human

Most conversations about AI in operations start with the wrong question. "What will AI replace?" invites anxiety and produces bad decisions — either rushing to automate everything or freezing and automating nothing. The better question is a design question: how do we arrange the work so that people and machines each do what they are genuinely best at?

Framed that way, AI stops being a threat and becomes a design problem — and design problems have good answers. The goal is not a workplace with fewer people or more software. It is a workplace where the routine runs itself and human attention is freed for the judgment, framing, and relationships that only people bring.

It's a teaming problem, not a replacement problem

The most durable research on service automation reaches a consistent conclusion: the point is complementarity, not substitution. MIT Sloan Management Review describes software that performs repetitive, rules-based service tasks previously done by people, precisely so that those people can focus on more unstructured and interesting work — and, done well, the result is a high-performing human-robot team in which software and employees complement one another .

That reframing matters because it changes what "good" looks like. Success is not the percentage of work automated; it is the quality of the pairing. The same research also makes a point that quietly resolves most of the fear: it is tasks within jobs, not whole jobs, that get automated . Roles are bundles of activities — some repetitive, some deeply human. The work is to separate them, not to judge an entire role as "automatable" or not.

Four lenses for drawing the line

So how do you decide, task by task, which side of the line something belongs on? Four lenses do most of the work. For each, there is an end that leans toward automation and an end that should stay human-led.

None of these are about the sophistication of the technology. They are about the nature of the work — how repetitive it is, how much judgment it needs, how clear the signals are, and how much is at stake.

Automate the routine; protect the judgment

The first two lenses — task shape and rules versus judgment — tend to decide most cases on their own. Work that is repetitive, high-volume, and governed by explicit rules is where software is reliably excellent and where people gain little from doing it by hand. Freeing people from it is a gift, not a loss.

The human core is the opposite: framing the problem, interpreting ambiguity, and weighing competing priorities. This is where MIT Sloan found organizations most often leave value on the table — underestimating how much comes from teaming algorithmic prediction with human expertise and intuition, especially in decision-framing . Algorithms are superb at answering a well-posed question; deciding which question to ask remains stubbornly, valuably human.

Keep humans where the signals are weak and the stakes are high

The other two lenses — signal clarity and stakes — govern the harder cases. When signals are strong and well captured in data, automation can act on them at scale. When signals are weak, contextual, or simply not in any dataset, people read them better. MIT Sloan notes that human input is particularly critical exactly when complex decisions must be made quickly or the signals are weak .

Stakes raise the bar further. Low-stakes, reversible, routine decisions are safe to automate and easy to monitor. But high-stakes decisions, rare edge cases, and genuinely novel situations are where human oversight matters most. The expert consensus is blunt on this: reserve human judgment for edge cases, high-stakes decisions, and novel contexts, while letting automated tools handle the bulk of the volume . The aim is a combined system that extends human judgment at scale — not one that either replaces it or is bottlenecked by it.

Design the oversight in — don't bolt it on

Where a human stays in the loop, when that oversight happens turns out to matter as much as whether it exists. The common instinct is to check outputs after the fact. The stronger practice is to build oversight in at design time: setting thresholds, designing tests, auditing workflows, and deciding up front where AI should not be relied on at all. MIT Sloan's expert panel is explicit that human oversight should be embedded at every stage of an AI solution's design and deployment, not treated as a final checkpoint on the outputs alone .

This is also why human expertise becomes more important as automation grows, not less. In a panel of 31 AI-strategy experts, 84% agreed that responsible AI efforts fail if they do not cultivate human experts who can verify AI solutions . Someone has to own the outcome, understand the system well enough to challenge it, and carry the accountability that software cannot.

A practical diagnostic

How deliberately does your team draw the automate-versus-human line today — and where is the biggest opportunity to sharpen it? The assessment below scores five dimensions and shows where to focus first.

Measure the value — and what you learned

A final discipline separates teams that compound their AI advantage from those that drift: they verify not just what the system produced, but what they concluded from it. It is easy to declare a pilot a success, track a speed gain, and move on — and just as easy to draw the wrong lesson, scale it, and bake a bad assumption into the operating model. MIT Sloan recommends scrutinising what the organization believes it has learned from a deployment, not only whether the outputs were correct .

This is strategic, not merely cautious. In a 2025 MIT SMR–BCG survey, 86% of top management teams said AI is a significant part of their strategic priorities . When AI is that central, the quality of the human judgment around it — what you automate, what you protect, and what you learn — becomes a direct driver of results, not a side concern.

Final thoughts

The anxious version of the AI question — what will it replace? — produces either reckless automation or paralysis. The useful version is a design question, and it has a clear method: decide task by task, match the work to the right doer, and keep a human accountable where judgment and stakes demand it.

Teams that get this right tend to share four habits:

• They decide at the task level — splitting roles into activities rather than judging whole jobs
• They match work to the doer — rules-based execution to machines, framing and judgment to people
• They keep humans where it counts — weak signals, high stakes, edge cases, and novel situations
• They design oversight in — thresholds and guardrails up front, and a person who owns the outcome

Decide deliberately, and AI becomes what it should be: a dependable teammate that frees people to do the work only people can do.

---

Think about one role on your team. If you split it into its individual tasks, which are routine enough to hand over — and which are the human core worth protecting?

Sources

1. 1
   MIT Sloan Management Review. A New Approach to Automating Services. September 2016.Mary C. Lacity and Leslie P. Willcocks. Argues that software robots perform repetitive, rules-based service tasks previously done by humans so that people can focus on more unstructured and interesting work, and that done well this produces high-performing human-robot teams in which software and employees complement one another. Notes the emerging view that tasks within jobs, rather than whole jobs, are what get automated.View sourceBack to citation
2. 2
   MIT Sloan Management Review. Designing AI Systems With Human-Machine Teams. March 2020.Maria Jesus Saenz, Elena Revilla, and Cristina Simón. Reports that many organizations underestimate the value of teaming the predictive capabilities of algorithms with the expertise and intuitions of humans, especially in decision-framing, and that human input is particularly critical when complex decisions must be made quickly or the signals are weak. Frames the greatest potential of AI as the mutual learning between people and machines.View sourceBack to citation
3. 3
   MIT Sloan Management Review. Beyond Verification — What Responsible AI Really Demands of Human Experts. May 2026.Elizabeth M. Renieris, David Kiron, Steven Mills, and Anne Kleppe, with BCG. A panel of 31 AI-strategy experts; 84% agree responsible AI fails without human experts who can verify AI solutions. Recommends embedding human oversight at the design stage rather than as a final output check; emphasising human judgment for edge cases, high-stakes decisions, and novel contexts while automated tools handle the bulk; and verifying what the organization learned, not just what the system produced. Cites a 2025 MIT SMR–BCG survey finding that 86% of top management teams consider AI a significant strategic priority.View sourceBack to citation