The future of work is oversight, not unemployment

Every serious forecast about AI and work argues about the same wrong question: how many jobs will disappear. It is the wrong question because it assumes the job is the task. Strip that assumption away and a different picture appears, one th

Every serious forecast about AI and work argues about the same wrong question: how many jobs will disappear. It is the wrong question because it assumes the job is the task. Strip that assumption away and a different picture appears, one that is already forming in the businesses that adopted agents early. The work is not vanishing. It is inverting. The human is moving off the tools and onto the oversight, from the one who does the task to the one who watches the machine do it, questions the decisions that look wrong, and holds the single authority the machine does not have: the power to say no.

This is not a utopian take and it is not a doom take. It is the boring, structural thing that happens when a technology gets good enough to act on its own. Someone still has to be accountable for what it does. That someone is the future of work, and the shape of that job is already visible if you know where to look. We run AI agents in production, for our own operations and for clients, so this is not a thought experiment for us. It is a description of a shift we are living through one workflow at a time.

The shift already started, quietly, in the boring workflows

Grand futurism skips the unglamorous truth: the transition to AI-does-the-work did not begin with humanoid robots. It began with a confirmation email nobody wanted to write.

When we automate a business, the first thing to go is never the interesting, judgment-heavy work. It is the repetitive connective tissue. The follow-up that should have gone out and did not. The invoice reminder nobody remembered to send. The lead that sat in an inbox for six hours while a competitor answered in six seconds. We wrote about the priority order of this in what a small business should automate first, and the pattern is consistent: an agent takes the task, and a human who used to do it now checks that it was done right.

That is the whole future in miniature. The person did not lose their role. Their role changed verbs. They stopped doing and started overseeing.

Watch it happen at the level of a single phone call. An AI phone receptionist answers, handles the routine, books the appointment, and the human owner does not touch the call at all until the moment it needs them. We made the case in will customers hate an AI receptionist that the entire product is the handoff, the clean escalation from machine to human on the calls that matter. That handoff is the atomic unit of the oversight economy. The AI runs the common case. The human takes the exception. Multiply that pattern across every function of a business and you have the next decade of work.

The numbers say churn, not collapse

Before the essay gets speculative, anchor it in the one large dataset everyone cites, because the data is more interesting than either the doom or the hype crowd admits.

The World Economic Forum's Future of Jobs Report 2025 projects 170 million new roles created and 92 million displaced by 2030, a net gain of 78 million, with 86% of businesses expecting AI to transform their operations and 22% of all jobs disrupted in some direction. The full report is drawn from a survey of a thousand companies representing over fourteen million workers.

Read those numbers as a headline and you get "AI creates more jobs than it destroys." Read them as a person who might be one of the 92 million and the reassurance evaporates. A net gain across the global economy is no comfort to the specific bookkeeper whose specific role is one of the displaced, especially when the 170 million new jobs demand skills the displaced worker does not have. The report itself flags this: 63% of employers name the skills gap as the single biggest barrier to their own transformation.

So the accurate summary is not "collapse" and not "everything is fine." It is churn. A vast reshuffling in which the total number of jobs holds or grows while the content of almost every job changes underneath it. And the direction of that change, across sector after sector, is the same: away from executing the task, toward supervising the thing that executes it.

This has happened before, and the bank teller is the tell

The pattern of a technology automating a task without eliminating the worker is not a hopeful guess. It has a canonical case study, and it is the bank teller.

When the automated teller machine spread through the 1970s and 1980s, the obvious prediction was that it would wipe out tellers. It automated the single most common thing a teller did, handing out cash. Yet the number of tellers in the United States did not collapse. It grew for decades. The economist James Bessen documented why: the ATM made each branch cheaper to run, so banks opened many more branches, and each branch still needed people. But the job changed. The teller stopped being a cash-dispensing machine and became closer to a salesperson and relationship manager, handling what the ATM could not: the complicated transaction, the upsell, the customer who needed a human. The routine was automated. The person was redeployed to the exception and the relationship.

Sit with how exactly that rhymes with the oversight economy. The machine took the common case. The human moved to the hard case and the human case. The total number of jobs held, and even grew, while the content of the job inverted. This is not proof the same happens with AI agents, because AI reaches much further into judgment than an ATM ever did. But it is a warning against the confident prediction that automating the task destroys the job. The last time everyone was sure of that, they were wrong, and the mechanism by which they were wrong, redeployment to the part the machine cannot do, is exactly the mechanism this essay describes.

The counterexample the doom case reaches for is the loom, and it deserves a fair hearing. The power loom genuinely did destroy the hand-weaving trade. The skilled weavers who smashed the machines were not irrational; they were correct about their own livelihoods even as the economy as a whole grew. So the honest lesson from history is not "it always works out." It is "it works out in aggregate and it is brutal in the particular." The weaving jobs did not come back. New jobs came, for different people, often in different places, after a painful gap. Both things are true, and the WEF's net-positive number lives inside exactly that tension: a gain for the economy that is no consolation to the specific person on the wrong side of the churn.

Two words that will define the next decade of work

The most useful vocabulary for the future of work does not come from a futurist. It comes from a regulation.

The EU AI Act, and the ethics guidelines behind it, define three models of human oversight, and they map almost perfectly onto the shift this essay is describing. Article 14 and the surrounding framework name them plainly.

Human-in-the-loop. A person sits inside the decision cycle. The AI proposes, the human approves, and only then does the action happen. Nothing consequential occurs without a human hand on it.

Human-on-the-loop. The AI acts on its own. The human monitors the stream of what it is doing and intervenes when something looks wrong. The default is action, not approval.

Human-in-command. A person sets the terms at a high level, decides when and whether the system is used at all, and retains the authority to shut it down or refuse to deploy it in a given situation.

The regulators already wrote down the job description for the future of work. It is not "operator." It is "overseer," in three escalating grades, defined by how much distance sits between the human and the action, and by who carries the authority to stop it.

— the framework hiding inside the regulation

The near future is mostly human-on-the-loop, with human-in-the-loop reserved for the decisions where a wrong answer is expensive, and human-in-command sitting above both as the ultimate off switch. That is exactly the arrangement your intuition reaches for when you imagine an AI-run business: the machine does the work, you watch the reports, and on the consequential calls you keep the veto. It turns out that is not a prediction. For high-risk systems in Europe, it is the law.

The work becomes exception handling

If the machine runs the common case, the human's day fills with the uncommon one. The central job of the oversight economy is exception handling, and it is worth being precise about what that means, because it is genuinely different from the work it replaces.

In the old model, a customer service rep handled every ticket, easy and hard alike, and spent most of the day on the easy ones. In the new model, the agent clears the easy nine out of ten, and the human sees only the tenth: the angry customer, the edge case the model is unsure about, the request that trips a policy the AI was told to escalate rather than decide. The human's entire day is now the hard tenth.

This has two consequences nobody plans for.

The first is that the work gets harder per hour, not easier. When the routine is stripped out, what remains is a concentrate of the difficult, the ambiguous, and the emotionally charged. The overseer does not get an easier job. They get a job with the filler removed, which is more demanding, not less.

The second is that the skill that matters most becomes judgment about the machine rather than execution of the task. The exception handler does not need to be able to do what the AI did. They need to be able to tell when the AI got it wrong. Those are different competencies, and the second one is harder to teach and harder to fake.

Everyone becomes a manager of one, then of many

There is a tidy way to describe what this does to the shape of a career: it turns every worker into a manager, and the thing they manage is a machine.

Management used to be a promotion, a thing a minority of workers did, the reward for years of doing the task well. In the oversight economy it becomes the default mode of almost all work. You do not do the task; you direct and review an agent that does. You set its goals, you check its output, you correct it when it drifts, you decide when to trust it with more and when to pull it back. That is management, applied to software instead of people.

At first it is management of one. A single person oversees a single agent handling a single workflow. We described the early version of this in n8n versus Claude agents, where the practical question is always how much autonomy to hand a given workflow and how tight to keep the leash.

Then it becomes management of many. We have run fleets of agents at once, and we wrote up honestly what that is actually like in what we learned running 33 autonomous agents. The short version is that it is far less like science fiction and far more like being a shift manager for a team of fast, capable, literal-minded workers who need clear instructions and occasional correction. The bottleneck stops being how much work can get done and becomes how much a human can meaningfully supervise. One person cannot truly oversee a thousand agents any more than one manager can truly oversee a thousand people. The span of oversight is the new constraint on organizations.

The organizations that win will not be the ones with the most agents. They will be the ones that figured out the right ratio of humans to agents, and built the tooling that lets a small number of people oversee a large amount of automated work without losing the thread.

The jobs that move first, and the ones that move last

Not everything inverts on the same schedule, and the timeline is more physical than most commentary admits.

The jobs that move first are the ones that already live inside a computer. Insurance office work, back-office finance, scheduling, first-line customer support, large parts of software development, routine legal and accounting drafting, media production. Anything where the "doing" is the manipulation of information can be handed to a software agent now, today, with a human moved to oversight. This is not a forecast. It is a description of what our clients are already deploying, and it is why we keep arguing that the highest-return automation is the revenue-adjacent workflow you already run, not some distant moonshot.

The jobs that move last are the ones that touch the physical world in unstructured ways. A cashier's information tasks automate quickly; the moment a shelf needs restocking by an unpredictable human hand, the timeline stretches. Construction, plumbing, eldercare, the trades: these require an AI not just to decide but to act on atoms, through a robot or a drone or a machine, in a messy environment that resists standardization. That layer is coming, but it is years behind the software layer, because moving reliably through the physical world is a harder problem than moving through a spreadsheet.

So the near future is asymmetric. Knowledge work inverts to oversight fast. Physical work inverts slowly, and in the meantime the physical trades may be among the more durable places to earn a living precisely because they are the last to be reachable by an agent. That is the opposite of what the "learn to code" advice of the last decade assumed, and it is worth sitting with.

The new bottleneck is not capability. It is trust.

Here is the pivot that the capability-obsessed coverage misses. Once an AI can do a task competently, the question stops being "can it do the work" and becomes "can you trust it to do the work without watching." And trust, not capability, is what gates how much of the economy actually inverts, and how fast.

There are three reasons a capable agent still needs a human over it, and none of them are solved by making the model smarter.

The first is that models still fail in ways that are hard to predict and easy to miss. They are confidently wrong sometimes, and a confident wrong answer that flows straight into an action, with no human between the decision and the consequence, is how automated systems produce expensive disasters at machine speed. Oversight is the circuit breaker.

The second is adversarial. An agent that reads the outside world can be manipulated by the outside world. We wrote a whole piece on why prompt injection is still the number one risk for any business running agents: an instruction hidden in a web page or an email can hijack an agent that has real permissions, and a smarter model does not close that hole, because the model cannot reliably tell content it is reading from instructions it should follow. The human on the loop is part of the defense. The consequential action waits for a person precisely because the agent's judgment can be poisoned.

The third is accountability, and it is the deepest of the three. When an AI makes a decision that harms someone, "the model did it" is not an answer a court, a regulator, or a customer will accept. Someone has to be responsible. And responsibility requires a human who could have stopped it and chose not to, or who reviewed it and approved it. You cannot hold an agent accountable. You can only hold accountable the person who was supposed to be watching. That single legal and moral fact guarantees the oversight job exists, no matter how good the agents get.

A more capable model does not remove the human. It raises the stakes of the human's attention. The better the agent, the more you trust it, the more it does unwatched, and the more it matters that someone is accountable for the one time it is catastrophically wrong.

— the reason the overseer's job is permanent, not transitional

"Why did the AI decide that?" becomes a full-time question

Your instinct that the human will "watch reports and ask why the AI made decision X" is exactly right, and it describes an emerging profession in its own right.

As agents take over consequential decisions, an entire function grows up around interrogating those decisions after the fact and, increasingly, before they take effect. Why did the underwriting agent decline this application? Why did the pricing agent drop the margin on this order? Why did the support agent issue this refund? In a world where a human made the call, the answer lived in that person's head. In a world where an agent made it, the answer has to be reconstructed, logged, and defensible.

This is the auditor of the oversight economy, and the role has real teeth. It requires someone who understands the system well enough to know what a good reason looks like, who can spot when the stated reason is a plausible-sounding fabrication, and who has the authority to reverse the decision when it does not hold up. It is part detective, part manager, part appeals court.

And crucially, not every decision gets this treatment. That is the part the doom narrative gets wrong when it imagines humans drowning in a million machine decisions. You do not question every action. You question the ones that cross a threshold: high value, high risk, a customer complaint, a pattern that looks off, a random audit sample to keep the system honest. The overwhelming majority of agent actions flow through unquestioned, exactly as the overwhelming majority of a good employee's actions do. Oversight is not surveillance of every keystroke. It is a well-designed set of checkpoints where a human's attention is worth the interruption, and silence everywhere else.

Designing where those checkpoints go, tuning them so the humans see the decisions that matter and not the noise, is itself one of the defining skills of the new economy. Too many checkpoints and you have rebuilt the slow manual process you were trying to escape. Too few and you have an unaccountable machine running your business. The art is in the ratio.

The new job categories, named plainly

It helps to make the abstract concrete. Here are the roles the oversight economy is already producing, described as jobs a real person could hold, because several of them already are.

The AI overseer. Monitors a fleet of agents doing a business function, watches the dashboards, handles the alerts, and keeps the whole automated operation running the way a floor manager keeps a shift running. Their KPI is not how much they personally produced. It is how much the agents produced without needing them.

The exception handler. Takes the cases the agents escalate: the hard support ticket, the ambiguous claim, the decision the model flagged as low-confidence. Spends the whole day on concentrated difficulty. Needs deep judgment in the domain, because the easy cases never reach them.

The decision auditor. Interrogates why the agents decided what they decided, reverses the calls that do not hold up, and produces the defensible record that regulators and customers demand. Part of every business that operates under any kind of scrutiny.

The evaluation engineer. Tests whether an agent is safe to trust with more autonomy before it gets it. Builds the checks that decide whether a workflow graduates from human-in-the-loop to human-on-the-loop. This is quality assurance for judgment, and it is the gatekeeper of how fast a business can actually automate. We touched on why this matters in n8n versus Claude agents: the hard part is never getting an agent to run, it is knowing when to trust it unwatched.

The systems builder. The person who wires the whole thing together: the agents, the checkpoints, the escalation paths, the veto. This is the role we occupy for our clients, and it is the one that has to exist before any of the others can. Someone has to build the oversight layer before anyone can staff it.

The common thread across every one of these is that the value is judgment and accountability, not execution. The machine executes. The human decides whether the machine should be trusted, catches it when it is wrong, and answers for it when it matters. That is not a diminished job. For a lot of people it is a more interesting one than the task it replaced.

The honest counter-case

An essay that only argues one side is marketing, not thinking, so here is the case against everything above, taken seriously.

The most aggressive timelines may simply be wrong. The AI 2027 scenario, written by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean, forecasts coding agents good enough to accelerate AI research itself by 2026, an intelligence explosion through 2027, and superintelligence by 2028. It is a serious, detailed, month-by-month scenario from serious people, several of them former OpenAI. It is also, by its authors' own framing, a scenario and not a prophecy, and the pushback has been substantial. Andrej Karpathy has argued for a slower "decade of agents" rather than a couple of years, and Kokotajlo himself has since slipped his median timeline from 2028 toward 2029, as tracked in this six-months-later review. If the aggressive timeline is off by five or ten years, the inversion this essay describes still happens, but slowly enough that it feels like normal generational change rather than a rupture.

There is also the possibility that oversight does not scale the way the optimistic version assumes. If genuinely supervising an agent turns out to require nearly as much human attention as doing the work yourself, the productivity gains shrink and the "manage many agents" future never fully arrives. The span-of-oversight problem is real and unsolved.

And there is the deskilling trap named earlier, which could make the oversight economy fragile in a way the industrial economy was not. A workforce of overseers who no longer understand the work they oversee is a workforce that cannot actually catch the machine's mistakes, which defeats the entire point.

The uncomfortable question is whether oversight pays

There is a question this essay has been circling and owes you directly: does the oversight job pay well, or does it just exist?

The optimistic version says oversight is high-skill judgment work, and judgment work pays. An exception handler who catches the mistake that would have cost the company a client, a decision auditor whose sign-off is what makes the automated system legally defensible, a systems builder who designs the checkpoints: those are leverage-heavy roles, and leverage tends to be paid for. In that version, a smaller number of people each oversee a large amount of automated output and are compensated for the leverage, average productivity per worker rises, and historically that is what pushes wages up.

The pessimistic version is harder to dismiss. If oversight becomes easy enough that anyone can do it, if the job is mostly clicking approve on a stream of agent actions, then it is low-skill work dressed as supervision, and low-skill work does not command high pay no matter what you call it. There is also the concentration problem. When one person plus a fleet of agents does the work of thirty, the gains from that productivity do not automatically flow to the person doing the overseeing. They can just as easily flow to whoever owns the agents. A world where a few people capture the output of enormous automated capacity while everyone else competes for a shrinking pool of genuinely human work is not a utopia. It is a distribution problem, and technology does not solve distribution problems on its own. It usually sharpens them.

Whether the oversight economy is broadly prosperous or narrowly captured is not a technical question with a technical answer. The machines do the work either way. Who benefits is a choice, and pretending the technology makes that choice for us is how the worse version arrives unopposed.

— the part no technology decides for you

The truthful answer is that both futures are available and the technology does not choose between them. Which one arrives depends on choices that are organizational and political rather than technical: how the gains from automation are shared, whether oversight is designed as genuine high-judgment work or as rubber-stamping, whether workers get the training to be real overseers or are handed a veto they lack the fluency to use. This essay cannot resolve that, and any essay that claims to is selling you something. What it can say is that the businesses treating oversight as a serious, skilled, well-supported function, rather than a cost to minimize, are the ones building the version of this future worth living in. Not incidentally, that is also the version that actually works, because a rubber-stamp overseer is not oversight at all.

What a business owner should actually do about this

Strip away the philosophy and there is a concrete answer to "so what," and it is not "wait and see."

The businesses that will struggle in this transition are the two extremes. The ones who refuse to automate at all, and get outrun by competitors running the same operation with a fraction of the headcount. And the ones who automate recklessly, hand agents real authority with no human veto, and produce an expensive disaster at machine speed the first time a model is confidently wrong or quietly hijacked. The safe path runs between them, and you learn it by walking it, on something small, now.

The move is to put one real workflow behind an agent and build the oversight around it deliberately. Pick something that matters but is not catastrophic if it errs. Wire in a human approval step on the consequential actions and let the routine flow. Watch where it needs a human and where it does not. Tune the checkpoints. Then, once you trust it, loosen the leash by one notch and watch again. That is how a business graduates a workflow from human-in-the-loop to human-on-the-loop without betting the company on it, and it is exactly the discipline we build for clients through our automation work. The point is not to have the most autonomous system. It is to have earned, checkpoint by checkpoint, the right to trust it.

The years between now and the fully inverted economy are not a waiting room. They are the training period. The businesses and the people who spend them learning to manage machines, to design the checkpoints, to hold the veto and use it well, are the ones who will be fluent when it matters. The ones who sit it out will be trying to learn management of a hundred agents at the exact moment their competitors already know how. If you want a head start, it looks a lot like figuring out what to automate first and refusing to skip the human veto to save a week.

The future of work is not a pink slip. It is a promotion nobody asked for, into a job that did not exist five years ago, managing a workforce that never sleeps, gets things wrong in novel ways, and needs someone accountable standing over it. That someone is you, or the people you employ, and the sooner you start practicing the role, the less it will feel like it happened to you and the more it will feel like something you built.