By Drew Soule
Picture the job. A benefits coordinator opens her queue at eight in the morning and it is already forty cases deep. A leave question that is really three questions. An address change someone submitted twice. A manager asking whether a write up can go in a file, with two people copied who did not need to be. A note from an employee whose insurance card never arrived, written in the tone of someone who has called the carrier four times. She works the list. She finds the right tab, the right policy, the right saved reply, and she closes what she can before the next batch lands. By noon the queue is forty deep again. This is what HR is for most of the people who actually do it.
Most companies still run the function this way, more or less the way they ran it twenty years ago. The titles have grown longer. The tools have multiplied. The work itself has barely moved. It is transactional, it is manual, and it remains the default in a year when the technology to remake every piece of it already sits unused on the same desks.
That is the part worth sitting with. Not that the tools exist. That they exist, and too few of the companies buying them are preparing the people who will have to work alongside them.
Start with what the word actually means, because it gets thrown around until it means nothing. A chatbot answers a question. An agent pursues a goal. Give an agent an objective and a set of tools and it will take a series of steps on its own, deciding what to do next based on what it finds, until the job is finished or it hits something it cannot resolve. The difference is the difference between a calculator and an accountant. One waits to be asked. The other works the problem.
Drop that into the coordinator's queue and the morning changes shape. The agent takes in each case, reads it, classifies it, pulls the governing policy and the employee record, drafts the response, completes the routine ones from end to end, and surfaces only the handful that carry real ambiguity, with its reasoning attached, for a person to decide. The forty cases become six that need a human. And the six are the ones that always mattered. The leave request where an employee mentions, almost in passing, a condition that quietly triggers a legal obligation the form never asked about. The address change that is really someone fleeing a situation at home. The write up that is the third one this quarter for the same protected complaint. The agent can route those. It should never be the thing that decides them. This is not a forecast. Teams are running versions of it now. The capability is here. The preparation is not.
So watch what companies actually do with it. Many buy the capability and cut the team in the same breath. Then they stand in front of the organization and call it efficiency.
Sometimes a cut is honest. A company over hired in a boom and the math no longer works. Runway is short and payroll is the largest line on the page. Two teams do identical work after an acquisition. Those situations are real, and pretending they are not is its own kind of dishonesty. A leader who faces one of them and makes the hard call is not the problem.
But too often that is not what happens, and everyone in the building can feel the difference. When the tool arrives and the people disappear together, with no plan for what the survivors will now be able to do that they could not do before, that is not a strategy. It is a confession the company does not hear itself making. It admits it never understood what those people were for. It wanted hands. The instant a machine outran the hands, the people became line items. That is not a story about powerful technology. It is a story about a function that was built to process and never asked to think, run by leaders who never thought to ask.
It helps to understand why the queue exists at all, because the answer is not flattering and it explains the panic. HR began as personnel administration, a back office that filed the paperwork and kept the company out of trouble. Somewhere along the way it split. One track became the business partner who was supposed to sit with leaders and shape the organization. The other became the service center that absorbed the volume so the partners could be freed for the strategic work. The freeing mostly did not happen, because the volume work still needed bodies, and the bodies were cheaper than the systems that might have replaced them. So the queue endured. Most of that manual work survived not because it created value but because automating it well used to be expensive, brittle, and rarely worth the trouble. The company tolerated the cost and called it a service. It was never a strategy anyone chose.
AI removes the excuse, and in removing it, exposes something uncomfortable. It reveals very quickly which companies built a function that thinks and which built a processing center with an HR sign on the door. If the only thing your people function ever optimized was throughput, you are in real trouble now, because throughput just got cheap. If your people were always doing the part the machine still cannot do, you are about to pull away from everyone who was not.
Nearly every argument about AI and work runs on one buried assumption: the machine does the job the person did, one out, one in. Call it substitution. It is the logic under every alarmist headline and every spreadsheet that treats a workforce as a stack of swappable parts. There is a gentler version, augmentation, where the tool sits beside the worker and makes the same job faster. Most of what gets announced as an AI strategy is really substitution wearing augmentation's clothes. Buy the tool, promise it will empower the team, and quietly plan for the team to be smaller by spring.
Both of those miss the only move that matters, which is elevation. Elevation means the human stops performing the task and starts directing the system that performs it. The work does not get faster. It changes kind. The practitioner who answered fifty questions a day now designs the agent that answers five thousand, sets its guardrails, audits its judgment, intercepts what it gets wrong, and reinvests the recovered hours in the problems that were always too consequential to rush. Same person, larger job. You do not end up with a smaller team doing the same small work. You end up with a leaner core of humans doing far larger and far more valuable work. That is the opening hiding in plain sight, and too few are funding it.
Here is where most pieces like this one start lying, and I would rather not. Elevation is not a happy ending for everyone. A function that runs on agents supervised by skilled people is, in plain fact, a smaller function. The arithmetic that frightens the workforce is not wrong. If one elevated practitioner can do what five used to do, four of those five have to land somewhere, and "go learn to design agentic workflows" is not a path open to every person or wanted by every person. Pretending the lean function employs as many people as the old one is a fairy tale, and people can smell a fairy tale from across the building.
So the honest version comes with obligations, and the obligations are exactly the part leaders skip. If the work is going to change under people, they are owed real reskilling, with real time and real budget, not a single training session and a slogan. They are owed honest conversations early, while there is still room to move, instead of a surprise meeting once the decision is already made. They are owed redeployment where it genuinely exists. And where it does not, they are owed a separation handled with dignity, with runway and real help landing somewhere new, rather than a calendar invite and a walk to the door. The difference between the company that elevates its people and the company that hollows them out is not whether anyone leaves. People leave in both. The difference is whether leadership treated the people as a cost to minimize or a responsibility to honor on the way through a hard transition. One builds a function that talent fights to join. The other builds a reputation that arrives before you do in every future hire.
It is worth being plain about what that does and does not condemn, because the argument is easy to misread as a refusal to ever reduce a team. It is not. Three things look similar from a distance and are not the same at all. The first is the blunt cut, where the tool arrives, the headcount disappears, no one redesigns the work, and the company pockets the savings and calls the subtraction a strategy. The second is deliberate redesign, where the work is mapped, the people are trained, the roles and the pay are rebuilt around the new leverage, and the function is then staffed honestly for what it has become, which may well be smaller. The third is the necessary separation, where some roles genuinely end, but with notice, with transition support, and with the dignity owed to people who trusted the place. The objection is not to the second or the third. It is to the first wearing the costume of the other two. The argument is not that no one should ever reduce a team. The argument is that you do not get to call a reduction a strategy when you skipped the redesign.
And elevation is hard. It is slow. It costs money before it saves any. It asks leaders to learn something genuinely new and then to teach it, which is harder than learning. Plenty of organizations will try it and fail, because they will buy the tools and skip the teaching, or teach without redesigning the work, or declare victory at the press release and wonder a year later why nothing changed. None of that is an argument against the path. It is an argument for taking it seriously instead of sloganizing it.
I said the function gets smaller, and it does. But the story of the people inside it is not only a story of subtraction, and the word we keep reaching for to describe the work at the bottom of it deserves a harder look. Remedial is a condescending name for jobs that are often grinding, relentless, and badly paid. The benefits coordinator clearing her queue. The shared services rep answering the same question for the hundredth time before noon. The onboarding clerk chasing a form someone forgot to sign. These roles get treated as disposable precisely because the work is bounded and the pay is low, and the two facts are taken as proof of each other. Agentic AI is the first thing in a long while that offers a real way to break that loop, not by freezing the jobs in place, but by turning them into something worth far more than they pay today.
Start with how pay actually works, because the romance about hard work obscures it. Wages do not track how difficult a job is or how many hours it eats. They track leverage and scarcity. A role pays little when the work is bounded, quickly learned, and produces a modest amount of value in an hour, which is another way of saying the person is replaceable and what they make in that hour is small. A role pays well when one person's hour generates a large amount of value and few people can generate it. That is the whole equation. It is why a job can leave you exhausted and still leave you broke. The remedial roles sit at the bottom of it not because the people in them are less able but because the work, as currently designed, hands them no leverage at all. Each hour buys one hour of bounded output. There is a ceiling on what that can ever be worth, and it sits low.
Agentic AI raises that ceiling further than anything since the spreadsheet. The person who learns to direct a system that does the work of ten is no longer trading an hour for an hour of output. She is producing the work of a team, and the value she creates in an hour climbs to match. The ceiling that kept the role cheap is simply gone. What can take its place is a role whose pay tracks the much larger value the person now creates, on one condition, that she holds the skill that makes the leverage possible in the first place. That is the conversion at the center of this. The same person, from the same starting line, now holds a completely different position in the only equation that has ever set the price of work.
None of that is a hopeful guess. It is the pattern technology has followed nearly every time it was paired with investment in people instead of used as a reason to be rid of them. The most cited example is the bank teller, and it gets cited because it embarrassed the forecasts. When the cash machine arrived in the 1970s, the obvious prediction was that it would erase the teller. For decades the opposite happened. Because the machines made each branch cheaper to operate, banks opened far more branches, and the total number of tellers actually rose. The job transformed underneath the people doing it. Counting and dispensing cash, the bounded part, went to the machine. The human part climbed toward what the machine could not touch, knowing the customer, untangling the messy problem, selling the product, holding the relationship that kept the account from leaving. The work moved up the value chain, and the better version of it paid better. The same arc ran through bookkeeping as the ledger went digital and the people who stayed became analysts and advisors rather than clerks, and through drafting as the drawing board gave way to design software and the ones who learned it designed instead of traced.
The analogy is useful, but it can be pushed further than it holds, and the honest version marks the seam before a skeptic does. The teller story turned on a fact specific to banking. Cheaper branches led banks to open more of them, more branches needed more people, and so the total count of tellers grew even as the work changed under them. HR has no equivalent engine. Nothing about clearing a benefits queue more cheaply makes a company want more benefits queues. So the lesson to take from the teller is the narrow one, not the triumphant one. Automation removing the bounded task does not by itself create more seats. It creates a better seat only where an organization deliberately builds higher value work above the part the machine took and keeps a person in it. Where no one builds that work, the task leaves and takes the job with it.
The telephone switchboard operator is the proof of that downside. That job mostly disappeared, and the people who held it were not handed a better one on the way out. Plenty of typing pools and clerical lines went the same way. Technology does not lift a role on its own. It lifts the role when someone invests in moving the people above the automated part, and it eliminates the role when no one bothers. The difference between the teller and the operator was never the sophistication of the machine. It was whether the work sitting above the automated piece was valuable enough to be worth keeping a human for, and whether anyone actually built the path to carry them there. In HR, the work above the automated piece is not marginal. It is the most valuable thing the function does. The path is buildable. The only open question, the same one this essay keeps arriving at, is whether anyone builds it.
Put it in the concrete terms of the function this whole piece is about. A benefits or HR coordinator, depending on the market, tends to earn somewhere in the high five figures, often less than that. The work is high volume and bounded, and the pay reflects the low leverage rather than the difficulty, because the difficulty is real and the pay still does not follow it. Now move the bounded work to agents that person designs and supervises. She is no longer processing cases. She is running the system that processes them, governing its judgment, auditing it for the quiet errors that turn into filings, designing the workflows that keep the queue from refilling. That is people operations systems work, and the roles that look like it, the ones that own the design and the governance instead of the queue, clear six figures and keep rising as the scope grows. The distance between the two is not a bump. It is the difference between a job people escape the moment they can and a career people deliberately build. And the person best positioned to make that jump is not some outside hire with a sharper title. It is the coordinator, because she already carries the one input the entire thing depends on.
That is the promise, and it is fair to be suspicious of it, because the obvious corporate move is to take the operator's work and keep paying the coordinator's wage. Hand her the agents, call it growth, and let the title sit where it was while the responsibility triples. Role expansion without pay expansion is the oldest trick in the function, and naming it cheerfully as a development opportunity does not make it anything else. So the pay does not follow the value on its own, and anyone who promises that it will is selling the same fairy tale from the other direction. Pay follows the value through specific machinery, and the machinery has names. The work has to be re leveled, lifted out of the coordinator band and placed in a higher one whose salary range reflects design, governance, and the accountability that rides with them. The job architecture has to recognize the role as a different kind of role and not a busier version of the old one. And none of that happens out of fairness either. It happens when the skill is scarce enough that the person can leave and the company knows it, because the only thing that has ever reliably moved someone into a higher band is the credible possibility that they will take the skill somewhere that pays for it. The protection against role expansion without pay expansion is not the employer's conscience. It is the employee's mobility. Which is exactly why the worker who builds the scarce skill early holds the leverage, and the one who waits until everyone has it holds none.
The coordinator is not a special case. The same conversion sits inside most of the bounded roles in the function. The recruiting coordinator who spends her day scheduling interviews and chasing feedback holds a working map of how hiring actually moves and where it stalls, and the person who turns that knowledge into the design and oversight of an agentic hiring operation is doing talent operations, which pays on a different scale entirely. The payroll clerk who knows every edge case that breaks a pay run holds exactly what you need to govern an automated one without quietly shorting someone's check, and compensation and payroll operations is not a clerk's wage. In each case the bounded task is the part that leaves, the judgment is the part that stays, and the pay follows the judgment once the person is the one running the system rather than the one buried beneath it.
That is what makes this shift different from the ones before it, and far more promising for exactly the people the old economy left at the bottom. Climbing the value chain used to demand something most remedial workers had no way to get. You could not turn a teller into a financial analyst over a weekend, because the work above demanded years of formal schooling the job below never offered. The skill agentic AI rewards does not work that way. The interface is plain language, not code. What makes a person good at directing these systems is not a credential. It is judgment about the work itself, knowing what a good outcome looks like, knowing the places it tends to break, knowing which case is genuinely routine and which one is a person's livelihood wearing routine as a disguise. The coordinator who has worked ten thousand cases has precisely that judgment. Her hard won sense of where the work goes wrong is the scarce input, and the system cannot generate it on its own.
It is worth being careful with the word easy, because plain language is not the same thing, and the gap between them is where most reskilling quietly fails. What got lower is the cost of access, not the demand of mastery. The interface no longer requires code, which means no computer science degree stands in the doorway and a coordinator can begin. It does not mean the work is simple. Designing an agentic workflow that survives real cases still asks for process thinking, for the discipline to test a system before trusting it, for documentation and data sense, and for a comfort with ambiguity that not everyone has and not everyone wants. The barrier that fell was the one at the entrance. The barrier of being good at it once you are inside is still standing, and it is real. Mistaking the low entrance for low difficulty is precisely how a company ends up with a training day, a slogan, and a roomful of people who were told they were operators and handed none of the time it takes to become one.
And being closest to the work, for all that it is the scarce advantage, is still not the same as being ready for the seat. Knowing where the queue breaks does not by itself mean a person can design the system that replaces it, audit it for the failures she cannot see from inside her own habits, hold the risk when it touches a protected category, or walk into a room and convince a nervous executive to fund any of it. Those are different muscles, and some of them are learned slowly. The coordinator is the best raw material for the operator, not the finished operator, and the honest version of the promise says exactly that. The skills she lacks are teachable, and teaching them is cheaper and faster than most leaders assume. The lived knowledge she already carries is the part that cannot be bought at any speed, which is why the distance she has to cover to become the operator is far shorter than the distance from teller to analyst ever was, and why she sets out with an advantage no outsider holds. The people who know the bounded work most intimately are frequently the best candidates to run the system that absorbs it, not the worst. A company that cuts them and hires fresh is discarding its cheapest source of the exact talent it will later spend a fortune hunting for.
It helps to be specific about what becoming that operator actually involves, because left vague it sounds like magic, and it is not magic. It is a progression, and it begins with the queue the person already has. First you use the tools on your own work, not to erase yourself but to see, from the inside, what they do well and where they are confidently and dangerously wrong. Then you learn to give a system a clear objective and the boundaries it has to stay inside, which is a real skill and a teachable one. Then you learn to read an output and feel the error in it before it ships, which is mostly your existing domain judgment pointed at a new target. Then you learn the specific failure modes of your specific work, the places these systems reliably break in benefits or hiring or pay, so you know exactly where a human has to sit in the loop. And then you design the workflows other people run, which is the point at which you have stopped being the labor and become the architect. None of those steps requires a degree. Every one of them rewards the person who already knows the work.
Now the hard part, because none of this arrives by gravity, and anyone who has watched the last fifty years has earned every ounce of their doubt. Productivity has climbed for decades while pay for most people sat still. The gains from squeezing more out of each worker did not reach the workers. They went up. So it is fair to ask why the surplus from agentic leverage would land any differently, and the honest answer is that it will not, unless something makes it. The surplus that comes from one person doing the work of ten can be captured three ways. The company can keep all of it by cutting the other nine and pocketing the difference, which is the default and the thing these pages have argued against from the start. The worker can capture a share by holding a skill scarce enough that the company cannot easily replace her, which is the entire mechanism by which any wage has ever risen. Or the two split it. What tips the result toward the worker is scarcity, and right now the orchestration skill is scarce, because so few people have bothered to build it. Scarcity is leverage, and leverage is how a person captures a piece of the value she creates instead of watching it sail past her to a shareholder.
This is also where the two arguments that have run side by side through these pages finally close into one, because until now they have looked as though they might pull apart. One says a company owes its people reskilling, dignity, an honest transition. That is a claim about what is right. The other says wages move only with leverage and scarcity, and markets do not pay anyone for being owed something. The reconciliation is that the moral path and the market mechanism turn out to be the same act seen from two sides. The reskilling a company owes its people is the very thing that makes their skill scarce and their footing strong. Treat them well and you manufacture the leverage that lets them hold a share of what they create. Skip it, and the market does what it always does with the unprotected, which is to pocket the surplus and move on. The moral case does not deliver the raise by itself. It delivers it only when it is built deliberately enough to make the worker hard to replace.
That window is the part to take seriously. The premium will be largest for the people who move while the skill is still rare, and it will compress as the skill spreads, the way every premium eventually does. That is not a reason to wait. It is the reason to move now.
It would be dishonest to pretend this erases the headcount math. One operator running a fleet of agents does the work that several coordinators used to, so even with upskilling there are fewer of these seats than there were of the old ones. But two things sit beside that. The first is that the people who fill the seats are paid like operators rather than clerks, so a function can shrink its headcount without shrinking its payroll, which is a very different thing from gutting it. The second is that clearing the volume work opens room for work that did not exist before. Someone has to govern the AI itself, audit it for bias, own the privacy questions, design the employee experience now that the queue is no longer eating the day, turn the data these systems generate into something a business can actually act on. Those are real jobs, they are skilled jobs, and they did not fit on the org chart of the service center because the service center never had the room. The function does not only redistribute toward fewer and better paid roles. It also expands into work that was always needed and never staffed.
There is one more reason this matters more than any story about efficiency, and it runs straight through the lens I see this entire field through. The remedial roles are held, disproportionately, by the people the traditional ladder served worst. The ones without the degree, without the network, without the early access, and in plenty of cases without a body or a circumstance that fit the path the ladder quietly assumed. A route upward that rewards domain judgment over credentials, that runs on language rather than years of specialized school, and that can be done from anywhere, is a wider door than most of what came before it. That door will not open on its own. Left alone, the gains will pool with the people who already hold the most, and the gap will grow, because that is what gains do when no one designs against them. But the raw material of an actual equalizer is sitting right here, for the first time in a long while, and it would be a particular kind of waste to have it and let it rot. The chance to take work the world treats as throwaway, and the people who do it, and turn both into something leveraged and respected and well paid, does not come around often.
And all of it rests on one load bearing fact. The wage a worker can capture here is exactly the judgment the machine cannot replace, which means the opportunity and the operator are the same person. Strip the human out and you lose both the safeguard and the raise.
There is also a darker possibility that the optimists wave away, and naming it is not a betrayal of the argument. It is the argument. Agentic AI in HR carries risks that are specific and serious. The first is bias at scale. A biased human recruiter makes a handful of bad calls a week. A biased screening agent makes the same bad call ten thousand times before lunch, consistently, defensibly, with a clean audit trail that hides the harm inside its own confidence. Automating a flawed process does not remove the flaw. It industrializes it. The second is surveillance. An agent given access to every message, ticket, and record so it can read sentiment is one configuration change away from a monitoring apparatus no one agreed to. The third is the quiet cruelty of handing people's worst days to a machine. The cases that reach HR are often the hardest moments of someone's working life, a harassment complaint, a denied accommodation, a job ending, and routing those to something fluent but not present is its own kind of harm. The fourth is the gap between what vendors sell and what the tools can do, and the legal exposure that opens when an automated decision touches a leave, a complaint, a disability, or a union question, because that is a lawsuit with a paper trail already written.
Every one of those risks points to the same conclusion, which is why the cost cutters have it precisely backwards. The dangers are not a reason to keep AI out of HR. They are the reason you cannot run it without skilled, accountable humans in the seat. The risk is the case for the operator.
Because agents are quick, certain, and wrong in ways that accumulate. They are wrong politely. They are wrong at volume. In most jobs a wrong answer is a typo you fix in the next meeting. In HR a wrong answer has a name attached to it. A leave misclassified so an employee loses a protection the law actually gave them. An accommodation handled as an inconvenience instead of a right, which is at once a moral failure and a filing. A reduction where the selection criteria look neutral on the page and produce a pattern that is anything but. A pay decision that bakes last year's inequity into next year's structure because the model learned from the inequity and called it signal. A performance flag a manager trusts because the system sounded sure, attached to a person whose context the system never had.
So the person who can tell when the machine is wrong does not lose value as the machines improve. That person becomes the most valuable one in the building. And here the claim has to be precise, because the loose version of it is easy to knock down. A model can already approximate a great deal of what looks like judgment. It can draft the policy, triage the case, surface the pattern, recommend the call. What it cannot do is own the judgment, stand behind the decision when it lands on a person and be answerable for it. You can automate the drafting of a policy. You cannot automate the instinct that knows how it will land on the person who has to live under it, or the one that catches something quietly off in an output that looks completely clean, and you certainly cannot automate the willingness to be the one accountable when it is wrong. Keep the engine and fire the driver and you have not saved money. You have removed the one thing standing between a fast system and an expensive human mistake. A fleet of agents with no skilled operator is not a high performing team. It is a liability on a short clock.
But a human in the seat is not the same as oversight, and it is worth saying so plainly, because rubber stamping wears the same job title as governance and produces none of the protection. Oversight that means anything is built out of specific habits, and the difference between a function that catches its machine and one that merely signs off on it lives entirely in whether those habits exist. There have to be escalation rules that say, in advance, which cases an agent is forbidden to close on its own and must hand to a person. There has to be sampling, a steady audit of the decisions the system made unwatched, because the errors that matter are the ones no one was looking at when they happened. There has to be bias testing run on the outputs as a matter of routine rather than after a complaint forces it. There have to be privacy boundaries the system cannot cross regardless of how useful crossing them would be. There have to be clear decision rights, a documented line showing which calls belong to the machine, which belong to the operator, and which belong above her. There has to be a path for an employee to contest a decision and reach a human who can actually reverse it. And when something does go wrong, there has to be a name attached to the ownership of it, because accountability that belongs to everyone belongs to no one. None of that is exotic. It is the ordinary machinery of a function that takes its own power seriously. Its absence is how you get a system that is fast, confident, unaccountable, and quietly doing harm with a signature at the bottom of every page.
There is a deeper reason the human in the seat matters, and it cuts against the grain of how these systems are built. A model learns from the record of what has already been written down. But some of the most important knowledge in this work was never written down, because it lives in the gap between what a policy says and how it actually lands on a person. The training data is mostly the policy. It is almost never the lived consequence of the policy, because no one files the consequence.
I do not write that from the outside. I have spent my career moving through organizations whose defaults were never built with someone like me in mind, navigating buildings and processes and unspoken assumptions that quietly took for granted a body and a path that are not mine. That vantage taught me something the org chart never shows: the distance between a policy that complies and a policy that includes, and how completely invisible that distance is to anyone who has never had to cross it. The interactive process that sits at the center of accommodation is the literal opposite of automation. It is individual. It is contextual. It is a conversation that depends entirely on hearing what the form did not think to ask. An agent can route that case and draft that letter. It cannot do the listening, and worse, it cannot know what it is failing to hear, because the thing it is missing is the part no one bothered to record.
This is why a diverse and experienced human core is not a luxury bolted onto an automated function. It is the failsafe. The people who have lived on the receiving end of rigid systems are exactly the people you want auditing those systems, because they can see the failure coming before it has a victim. Build a lean function out of narrow people and what you have built is a fast machine with no one inside it who can feel when it is about to do harm.
The obvious objection is that all of this is temporary. The models keep improving. The gap between what the machine can do and what the human must do is real today, the argument runs, but it is closing, and soon enough the system will be good enough to run with no one in the seat. Wait it out and the operator becomes a cost like any other. I think that argument is wrong in a way that matters, and not for sentimental reasons. Two things do not close as the models get better. The first is accountability. Someone has to own the decision when it lands on a person's livelihood, and ownership is not a capability you can train into a model. A company cannot put an algorithm on a witness stand. It cannot ask a system to look an employee in the eye and explain why. The law and the basic decency of the thing both require a human who is answerable, and the better the machine gets at acting, the more that answerable human matters, because the actions are faster now and they reach further. The second is that the relational, judgment heavy core of this work is not a gap waiting to be filled. It is the work. Trust is not a task you automate once and tick off. It is something carried by a person the other person actually believes. A machine can write the words of a hard conversation. It cannot be the one trusted to have it. A better model does not shrink that. It raises the value of the people who can do the part no model will ever own.
So why do leaders reach for the cut anyway, even the ones who would fail every part of this argument if you pressed them on it? Because headcount is the lever they know how to pull. A great many of the people making these calls could not describe an agentic workflow on the spot. They cannot tell you where a human belongs in the loop, or why, or what breaks when no one is there. So they feel the ground move beneath them, they reach for the one instrument every pressured executive has always had within reach, and they cut. Then they call it transformation.
But it would be too easy, and not quite honest, to leave it at a failure of imagination. Some leaders genuinely lack the imagination to see another path. Many more can see it perfectly well and are trapped inside incentives that punish the slow, expensive, correct version and reward the fast, visible cut, which is why this is meant as charity and not only as an insult. Boards reward the AI narrative. Public markets price in the efficiency story before a single workflow has been redesigned. There are consultants whose entire deliverable is a smaller org chart with a confident font. A leader who announces a leaner team built around AI gets applause this quarter. A leader who says we are going to spend a year and real money teaching our people to do a harder job gets a room full of questions. The incentives all point at the cut. None of that makes the cut right. It makes the leader who resists it rare, and worth following.
But there is a line, and it stops being subtle the moment you look for it. A leader making a hard, necessary decision owns it, explains the reasoning, and treats the people leaving like people. A leader laundering avoidance hides behind the technology, calls a failure of imagination a strategy, and lets "the AI made us more efficient" do the talking so they never have to say the rest out loud. The first is a difficult call. The second is a leader offloading their own discomfort onto the careers of people who trusted them to know more.
So what does the other path actually look like, the one that does not fit cleanly on a slide? It starts where the cutters never start, which is with the work and not the tool. Before you buy anything, you map what the function actually does and you sort it honestly. Some of it is genuinely judgment light: the address change, the routine verification, the policy lookup, the first pass triage. That work can and should move to agents, and the people doing it today should be the first ones taught to run them, not the first ones shown the door. Some of it is judgment dense and relational and stays firmly with humans: the investigation, the interactive process, the conversation no one wants to have, the design of the program itself. Most of the value in HR always lived in that second pile. The quiet tragedy of the service center model is that it buried your best people in the first.
And the sorting is harder than it sounds, because in HR the category of routine is a trap. Routine is not defined by how simple the request looks on arrival. A leave question can look like a two minute reply and turn out to sit on top of the ADA, an unasked pregnancy accommodation, a state statute, a retaliation risk, or a union provision, none of which the form announced. The only safe definition is the demanding one. A case is routine when the system can confidently rule out every protected, legal, relational, and equity implication, and not one second before. Anything that cannot clear that bar is not routine no matter how trivial it appears, and it belongs with a person. Designing agents that respect that line, that escalate on the faint signal rather than the obvious one, is most of the actual work, and it is exactly the work that requires someone who has been burned by the case that looked simple and was not.
Then you build fluency as a real competency, not as a slogan painted on the wall. Fluency here is not knowing the vocabulary. It is the ability to hand an agent a clear objective, to design the guardrails it runs inside, to read an output and sense when it is confidently wrong, to know the particular failure modes of the particular work, and to hold onto the judgment the system cannot have. That is a skill. It can be taught. It takes time and repetition and a leader willing to spend both. I have spent fifteen years building and rebuilding people functions, across a table from a twelve hundred person bargaining unit in a manufacturing plant, inside a hypergrowth product organization, in the run up to a public offering, and once from nothing in sixty days. The pattern holds across every single one of them. The leverage was never in the headcount. It was in the design, and in the judgment of the people running it. The build that started from zero did not work because it was large. It worked because it was designed well, and the right design pulled attrition down and cost out at the same time, which is the thing the cutters never believe is possible until they watch someone do it.
And you redeploy, on purpose, the hours the agents hand back. That recovered time does not get harvested as savings and walked up to the board. It gets reinvested in the work that was always starved: actually knowing the managers, getting ahead of the problems instead of mopping up after them, designing the programs that stop the queue from forming in the first place. Measured correctly, the lean function does not merely cost less. It produces more of the only things HR was ever supposed to produce, which are trust, fairness, and an organization that people do not quietly plan to leave. I have watched a small team built this way absorb the load of one three times its size, and not by grinding the survivors into dust but by handing the volume to systems and aiming the people at the parts that needed a person. The team got smaller and stronger in the same motion. That version is real. It is simply harder to build than a layoff, and it does not photograph as well.
And this is the part that should keep the cutters awake. The humane path and the smart path are the same path. Elevating your people is not generosity you indulge instead of winning. It is the mechanism of winning. The company that spends the year teaching its people to run AI well will, in three or four years, have a people function that is faster, sharper, less expensive, and more humane than the one that fired its way to efficiency, and the distance between them will not be a rounding error. It will be a moat. You can rehire bodies in a quarter. What you cannot rehire in a quarter is the people who knew where the bodies were buried, what the system gets wrong, and how to keep a fast machine from doing harm. Fire that and you are buying it back later at a premium, if it is for sale at all.
If you are the coordinator, or the generalist, or the partner reading this between cases, none of it is a reason to wait for a leader to save you, because most of them will not. The same shift a short sighted company will use as an excuse to remove you is the one a clear eyed person can use to become impossible to remove. The leverage is available to you directly, and it does not require anyone's permission. Learn to direct these systems before you are told to. Get your hands on the tools and build something small, a workflow, a draft process, a way to clear part of your own queue, so you understand from the inside where they are strong and where they fail quietly. Make yourself the person who can run the agents and catch them, who can hold the relational, judgment heavy work that does not automate, and who can tell a nervous executive exactly where a human has to stay in the loop and why. That person does not get cut. That person gets handed the function, and over time gets paid like someone who runs it rather than someone who feeds it.
And yet none of that lets the people in charge off the hook, and it would be a quiet betrayal to end as though it did. Individual preparation is real, and it matters, but it is not a substitute for institutional responsibility, and it cannot be allowed to become the alibi's last move, the one where the burden of a transition the company chose gets transferred onto the workers least able to carry it. A worker can prepare for the future. A company still has to stop using that future as an excuse to abandon the people who built the present. The hard truth of this moment is that the people most exposed are often the ones with the least time to prepare. The opening, for anyone who can find the hours, is that the skill is learnable and the bar is still low, because so few people have bothered to clear it. The ground is moving either way. You can be moved by it, or you can learn to stand on it.
Go back to the coordinator and her forty deep queue. There are two futures in front of her. In one, the tool arrives, the queue empties, and so does her desk, and a leader somewhere puts the savings in a deck and calls it a win. In the other, the tool arrives, the queue empties, and she is the one who learns to run the system that emptied it. She designs its guardrails. She catches what it gets wrong. She spends her recovered mornings on the work that actually needed a human, which was most of it. Same technology. Same arrival. The only variable is whether anyone in charge had the imagination to see her as something more than a pair of hands.
The travesty was never that AI showed up. It is that so many leaders were handed a chance to build something rare out of the people they already had, and used it to make them fewer. The tools are here. The talent is here. What is missing is partly the imagination to put them together, and mostly the nerve to lead people through the change instead of around them, against every incentive that rewards the cut and questions the build.
Lean was supposed to be a discipline. Too many leaders are using it as an alibi. The ones who refuse to will own whatever comes next…
Top comments (0)