There’s a version of this conversation everyone’s been having. It goes: AI will take jobs. Or it won’t. Or it’ll create new ones. Or we’ll be fine. Or we won’t.
I’m genuinely tired of the jobs debate. Not because it’s unimportant. Because it’s the wrong argument.
The thing AI is actually changing is more uncomfortable than job loss. It’s this: for a lot of people, the thing they’ve been paid to do, the core activity they call their job, doesn’t require human judgment. Never did. AI is just making that visible.
That’s the shift. Not headcount. Not automation. Not even productivity.
The shift is from doing work to owning decisions. And those are not the same thing at all.
The comfortable middle is already gone
Before getting into what changes, it helps to understand what’s already happened. Because the “AI is coming” framing is weirdly premature. The middle of the labor market has been hollowing out for two decades.
Economists David Autor, Lawrence Katz, and Alan Krueger documented the early version of this pattern in the late 1990s, showing how computerization started eliminating routine cognitive work while high-skill and low-skill roles stayed relatively intact.
David Autor extended that work significantly, and by the early 2010s the evidence for what he called “job polarization” was fairly well established: the roles most at risk weren’t the lowest-paid but the mid-wage, routine-heavy ones. Data entry clerks. Bookkeepers. Quality control auditors whose entire job could be described in ten rules.
That pattern didn’t stop. It accelerated.
What AI does is push the boundary upward. The “routine” category now includes things that used to sound complicated: drafting legal summaries, writing first-pass code, analyzing competitive positioning, building financial models, responding to customer escalations based on policy guidelines. These tasks look professional. They’re cognitively demanding in a surface sense. They require training, familiarity with jargon, sometimes years of practice to do quickly.
But they share something with the bookkeeper’s job: given enough examples and a clear success criteria, they can be done without anyone actually making a call.
That’s the tell. Not whether the task is hard. Whether it requires a call.
A junior analyst who spends sixty percent of their week pulling data, formatting it, and summarizing it for a director to interpret is not making calls. The director is making calls. The analyst is producing inputs. AI produces inputs faster, cheaper, and without burnout.
That’s not the uncomfortable part, though. It’s not that the analyst might lose their job. It’s that the analyst’s job was always the director’s job, outsourced downward for cost and convenience, and AI just made that obvious.
MIT economist David Autor noted in a 2022 NBER working paper that middle-skill job displacement has been ongoing since roughly the 1980s, and that what distinguishes the current wave is the speed of capability expansion into tasks previously considered non-automatable. His concern wasn’t apocalypse. It was the compression of the timeline for adaptation.
That compression is what makes this moment different. Not the destination. The speed. And the fact that this time, the capability expansion is hitting roles that have traditionally felt quite safe: analyst, associate, coordinator, specialist. The cushion between “I do knowledge work” and “my role is economically exposed” is thinner than it has ever been before.
What “knowledge work” actually meant
This is happening in nearly every org I know. A team gets built around specialization. You need someone who knows the analytics platform. You need someone who knows the legal compliance requirements. You need someone who knows the presentation layer of the client relationship. And you need someone who can write.
Each of those people becomes the organization’s memory for their slice. The system works. It’s inefficient, but it works. The knowledge lives in people, so the people have to stay.
There’s a mild irony in how much of what we called “knowledge work” was really “organizational memory work.” The knowledge wasn’t insight. It wasn’t judgment. It was familiarity. Knowing where things lived. Knowing which form to use. Knowing that Lucy in legal always takes three days to turn around a contract so you have to plan for that. Knowing the PowerPoint template is actually version 4.2 from February, not the one on the shared drive.
This is not a small amount of what knowledge workers do. In my experience it’s a large amount. Maybe most of it for mid-level roles.
Management consulting, interestingly, built an entire industry around this problem. When consultants come in and spend two weeks just “understanding the org,” what they’re doing is acquiring organizational memory fast enough to be useful. The firms that got very good at this, McKinsey, Bain, BCG, developed structured approaches: process maps, knowledge repositories, interview frameworks. They were not solving intellectual problems. They were solving organizational amnesia problems efficiently.
The consultants who survived long careers were never the ones who could acquire organizational knowledge fastest. They were the ones who could sit with a CFO who’d just gotten contradictory data from three departments and help her figure out what to actually do. Not what the data said. What to do.
AI dissolves the organizational memory problem. That capability, the thing most of those mid-level roles were actually protecting, evaporates.
What remains is the judgment layer. And we have not, as a society, trained people for that in any systematic way.
The difference between doing and deciding
A few weeks ago, I found a National Bureau of Economic Research paper written by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond that I’ve thought about more than almost anything else in this space. They studied the deployment of an AI assistant in a large software customer service operation and found that the workers who improved most were not the most experienced.
They were mid-experience workers who could now perform like experts because the AI was handling the retrieval and pattern-matching work. The experienced workers improved less, percentage-wise, because they already knew the patterns.
The finding I keep returning to is a follow-on implication: if AI lifts average performance toward expert performance in retrieval-heavy tasks, what distinguishes the expert now?
Not faster retrieval. Not more accurate pattern matching. Not a better memory for the organizational manual.
What distinguishes the expert is knowing when the manual is wrong.
That sounds simple. It is not simple. Knowing when a rule should be broken, when a client relationship is about to go somewhere no process was designed for, when a decision looks right on the numbers but feels wrong for reasons that require articulating, when you should escalate versus absorb a risk yourself, when to trust a good outcome that arrived via a bad process.
This is not knowledge. It’s judgment. And judgment comes from having been wrong in consequential ways enough times to have a calibrated sense of where you’re likely to be wrong again.
I’m not fully confident that judgment can be taught or that organizations will reward it properly even when they need it. When I was working in an agency, I’ve seen companies say they want people who “challenge the status quo” and then freeze out the first person who actually does. The gap between organizational rhetoric about judgment and organizational tolerance for the actual exercise of judgment is large. I don’t know how to close that gap. I’m not sure this piece closes it.
What I’m confident about is that it’s the right problem. The roles that will hold value are not the roles where humans beat AI at retrieval or speed. They’re the roles where the human is accountable for a call, where the output requires owning a position, where being wrong has a name and a face attached to it.
Responsibility is becoming the scarce resource. That’s new. Or at least it’s happening faster than it ever has before.
Teams built for specialization will break
Most organizations are still structured as if information is expensive to move.
The departmental model, marketing over here, analytics over there, product somewhere else, legal on a different floor or a different continent, exists because historically you couldn’t give everyone access to everything. So you hired specialists. Put them in groups. Built handoffs between the groups. The handoffs created friction. The friction created coordination roles. The coordination roles created more friction. You end up with an org chart that’s a map of information bottlenecks, not a description of how work actually flows.
AI removes a lot of the reason information is expensive to move.
A generalist with good tools can now perform work that would have required three specialists six years ago. Not in every domain. Not for the hardest problems. But for a meaningful percentage of the problems organizations actually face on a given week.
The structural response to this is not complicated in concept: smaller teams, flatter reporting, roles organized around outcomes rather than functions. You stop asking “what’s your job title” and start asking “what problem are you solving.”
In practice it’s very disruptive. Entire departments whose value proposition was managing the complexity of their own existence become hard to justify. Middle management layers that existed to translate between specialists and executives get squeezed from both sides. The executive can now consume a synthesized briefing from AI tools that would have taken three analysts two days to produce. The specialist can now self-deploy in ways that used to require a manager’s bandwidth.
At this moment, I’m uncertain how fast this actually happens at scale. Organizations resist structural change with more force than they resist almost any other kind of change. The people who benefit from the current structure, and there are many of them, do not make transitions easy.
Regulation may slow adoption in some sectors. Some industries, healthcare, financial services, legal, are subject to oversight frameworks that add friction to every deployment of automated tooling.
The 2-to-4 year horizon is probably right for the early-adopter companies, the ones already organized for speed. For large traditional enterprises? Maybe 5 years, maybe a decade, maybe longer. Who knows. Structural change at a Fortune 500 is not a product sprint.
But the direction is clear. And people who have built their careers entirely inside a specialization, who’ve never been asked to own an outcome from beginning to end, who’ve always been insulated by role definition from having to make a call that wasn’t covered in the job description, a responsibility that wasn’t in their job description, are in a really tough spot.
Not because AI is taking their job. Because the organizational logic that required that job in the first place is being dismantled.
This is not going to feel like progress
I want to end here without the standard move, which is to pivot to “but here’s how you adapt.” I’ll say a few things that are actually useful. But not before saying something honest.
This is going to be hard for a lot of people. Not in a temporary disruption before everything gets better way. In a the path you were on is gone and the new one requires something you haven’t been asked to develop way.
Maybe there won't be a “jobs apocalypse” due to AI, but there will be job chaos, as Gartner predicted.
But workers most at risk are not the people who think they’re most at risk. The junior analyst who’s worried about AI taking their entry-level job is right to be concerned about the market.
But the mid-level manager who assumes their judgment is safe because they’ve been “doing strategy” for ten years should be equally worried.
Especially if the strategy work mostly involved collecting inputs from reports, formatting them into a deck, and presenting a consensus recommendation. That’s not strategy. That’s a well-dressed coordination function.
The people who will come through this in a strong position share a few traits, and none of them are credentials. They can hold ambiguity without freezing. They take positions even when the data is imperfect. They’ve been wrong in ways that were costly and they know it and they adjusted because of it, not despite it. They’re accountable for outcomes in a way that means someone calls them when something fails.
Junior roles are almost certainly shrinking in the near term, the entry-level slowdown among workers in their early twenties is showing up in hiring data pretty clearly. But whether those roles disappear or just mutate into something nobody has named yet, I genuinely don’t know.
The historical record says new role categories emerge. What those categories require and whether the people being frozen out now can access them later is a different and harder question. I keep reading optimistic takes about it and I keep not being fully convinced either way.
Recent McKinsey Global Surveys on AI adoption have highlighted that most large organizations still feel their workforce is not fully prepared for AI-augmented work roles, with many continuing to focus on initial experimentation and training rather than deep integration.
Which is a polite way of saying that eighty-five percent of companies know they have a problem and most of them are still doing workshops about it.
The honest advice is not complicated to say. Be the person who makes the call. Get into positions where you’re accountable for outcomes, not just contributory to them. Learn to be wrong in front of people and recover from it fast, because that’s what building judgment actually looks like.
But I should acknowledge that not everyone has access to those positions.
Not everyone has been given the runway to develop judgment because the orgs they’ve been in didn’t trust junior people with anything real.
And not everyone has the safety net to take career risks. The people who’ll thrive under this shift are disproportionately the people who already had structural advantages.
The shift is real. The direction is clear. But I keep coming back to something that the frameworks and the forecasts don’t really address: the people most exposed to this transition are often the ones who did exactly what they were told. They built expertise inside systems that rewarded specialization. They were reliable. They showed up.
The fact that the system is now changing its mind about what it values is not a personal failure on their part. It’s a structural one.
I don't have a satisfying answer for those people. What I'd offer instead: the discomfort of not knowing where you stand is more honest than the confidence of someone who thinks they do. Start there. It's not nothing.
The three-question test for whether your role survives the next five years
Most of the “is your job AI-proof” frameworks I’ve seen are reassurance dressed up as analysis. They ask things like “do you work with people?” or “do you use creativity?” and then confirm that yes, you’re probably fine. That’s not useful. Plenty of roles involving people and creativity are going to be restructured significantly.
Here’s a more useful frame. Three questions. Answer them about your actual day last Thursday, not about your job description.
Question 1: Who bears the consequence if your output is wrong?
Not who gets annoyed. Who actually faces a material consequence: lost revenue, legal exposure, a damaged relationship, a failed project?
If the answer is “someone above me,” your role is input-production. Those roles compress. If the answer is “I do, directly,” you’re in accountability territory. Those roles hold and in some cases grow stronger.
This is not about seniority. There are VPs who produce inputs for board members who make calls. There are account managers three years out of college who own client relationships that cost real money to lose. The org chart doesn’t tell you which one you are. Your inbox does.
Think about the last time something you produced was wrong. What happened? If the answer is that someone above you got a call from someone above them and you only heard about it later, you’re an input. If the answer is that your phone rang, that’s different.
Question 2: What percentage of your week required a judgment no one gave you a rule for?
Be precise. Not “I had to think.” Thinking is not the same as deciding under ambiguity. The test is: was there a moment where a sensible person in your position could have reasonably gone either way, and you chose, and you owned it?
If that percentage is under twenty, your role is heavily rule-based. Not necessarily replaceable by AI today, but structurally fragile. The path toward durability is increasing that percentage, which usually means asking for, and getting, problems that don’t have scripts.
This is harder than it sounds. Most organizations actively resist giving ambiguous problems to people below a certain level. The reasoning is usually framed as protection, “we don’t want to set them up to fail,” but the effect is that mid-level people accumulate years of experience operating inside guardrails without ever developing the capacity to operate outside them. Then the guardrails come down and they don’t know what to do.
If you’re in that position, the practical move is not to wait for permission. Find the edges of your role, the places where the process doesn’t quite cover what’s actually happening, and make a call. Tell your manager what you did and why. Do this consistently enough and you start building a record of judgment that’s visible. That record is what survives restructuring.
Question 3: Does your organization know something is wrong faster because you’re paying attention?
This one is underrated. Early signal detection, the ability to notice that a client relationship is shifting before the client says it, that a team is quietly losing cohesion before it shows up in output, that a product assumption is breaking down in ways the data doesn’t yet capture, is something AI tools are not good at. Not yet.
The people who are wired for this, who absorb the ambient information in a system and synthesize it into an early warning, have a capability that is genuinely rare and currently very difficult to replicate. If you’re one of them, the shift probably helps you. The noise that used to obscure your signal decreases when AI is handling more of the routine.
If you’re not, it’s a skill worth deliberately building. Which means: put yourself in situations where you have to give an opinion before the evidence is conclusive, and see how often you’re right. Keep a running note. Check back in a month. Over time you build a calibration record that shows you where your intuition is reliable and where it’s noise.
What the structural changes actually look like early
I’ve worked with a handful of companies that are further along on this than most. None of them are doing what you’d expect.
They’re not eliminating departments wholesale. They’re shrinking the ratio of production roles to decision roles. In one company I know well, a 35-person marketing org went to 18 people over 18 months, but the ratio of people who could speak authoritatively to a client or a board about strategy went up.
The people who left were the people whose job was to move information from one format to another. The people who stayed were the ones who could say “here’s what I think this means and here’s what I’d do about it” and be willing to be wrong in front of executives.
Another pattern: the job titles are getting weirder. “Growth lead” instead of “content manager plus SEO manager plus email manager.” “Revenue operations” absorbing what used to be sales operations, marketing ops, and CRM administration across three separate headcounts. The title compression is a structural signal. Problem-based roles collapsing the function-based ones.
The org design change that I didn’t see coming, and that I think is underappreciated, is what’s happening to middle management specifically. The traditional middle manager role had three core functions: translating strategy downward, aggregating performance information upward, and being the first human escalation point when something went sideways.
AI tools are absorbing the first two functions faster than almost anything else. Strategy translation can be automated. Performance aggregation is genuinely easier with AI synthesis tools than it was with humans.
What remains is escalation handling. The moments when something has gone sideways in a way that involves a person, a relationship, a judgment call under pressure. That’s still human. But it’s maybe a third of what the job used to be.
The middle managers who are coming through this are the ones who were always actually good at the human part and tolerated the administrative part. The ones who built the middle management layer as a way to process administrative load without developing the harder skills are going to find the job changing faster than they expected.
There’s also something happening with performance visibility that I think is underappreciated. When AI handles the volume work, what’s left is more visible, not less. In a world where everyone on a team is producing roughly similar output volume via similar tools, the differentiator becomes the quality of the calls you make, the judgment you bring to the edge cases, the way you handle the situations that fall outside the template.
In a purely execution-based role, a moderately capable person can hide inside the volume. In a judgment-heavy role, there’s no place to hide. If you’re good, that’s fine. If you’ve been coasting, it’s not.
One implication of this that doesn’t get discussed enough are performance reviews that are about to get more accurate, and not in a way that will feel good to everyone. The people who have been doing solid but unremarkable execution work for years, reliably hitting their numbers in a role where the numbers were mostly a function of showing up, are going to find the new visibility uncomfortable. The people who were actually exceptional but whose signal was drowned out by the noise of a high-volume operation are going to find the opposite.
None of this is comfortable to read if you’re mid-career and have spent the last decade building expertise in a function that AI is now absorbing. I’m not going to pretend it is. The adaptation path is real, but it requires things from people that many organizations haven’t given them the opportunity to develop. That’s a structural problem that individual advice doesn’t fully solve.
What I’d say to the person reading this who’s trying to figure out where they stand: start with the three questions above. Be honest. If the answers are uncomfortable, that’s more useful than reassurance, because at least it tells you what to work on.
The companies that end up in the strongest position through this shift are not necessarily the ones that move fastest on AI tooling. They’re the ones that figure out early what they’re actually paying people for. The ones that confuse “we have fewer seats” with “we have better judgment” are going to learn the difference expensively.
And for individuals: the goal isn’t to become indispensable through scarcity of knowledge. That worked before. It doesn’t work now. The goal is to become the kind of person an organization cannot afford to have wrong. That’s a different thing entirely, and it takes longer to build, but it’s also harder to automate away.
Thank you for reading!
Jan






