The loss of bargaining power can arrive long before mass unemployment.
There are two chairs in a salary negotiation. One belongs to the employee. He has prepared a list of projects completed, customers retained, problems solved and perhaps the number he cares about most: what somebody else might pay him. The manager has numbers too. Some are on the screen. Others never leave his head: how difficult this person would be to replace, who else knows what he knows, how long the department could operate without him, what would break if he left on Friday.
This is the part of compensation nobody puts on the payslip.
A worker is not paid only for what he produces. Some part of his salary comes from the inconvenience of losing him. For much of industrial history there was a stubborn physical fact underneath that negotiation: somebody still had to do the work. The factory needed hands, the mine needed miners, the hospital needed nurses, the bank needed clerks and analysts, and a software company eventually discovered that at least one living person had to understand the accumulated ruins of fifteen years of software written by people who had already left.
Employers could underpay workers, move factories, break unions, import labor, outsource departments, replace troublemakers and automate whatever machinery could handle. Somewhere inside the business remained a dependency. Labor could be exploited precisely because labor still possessed something worth extracting.
This is where exploitation and redundancy separate. An exploited worker may have almost no individual power. Twenty people may be waiting outside the gate to replace him. His employer can still be making an awkward admission every morning the gate opens: I need what humans provide.
Redundancy removes that admission. It does not have to remove the human.
Two exit doors
There are two exit doors in every employment relationship, though most discussion of worker power concentrates on one of them. Can the employee quit? Is there another employer nearby? Could she survive three months without a salary? Is her health insurance tied to the job? Does immigration status turn resignation into a legal problem? Does she have children, debt, an elderly parent or a spouse whose work cannot move? Every answer changes the price of saying no.
Economists call these outside options. A 2026 study by Sydnee Caldwell and Emily Oehlsen used randomized pay changes among Uber drivers together with changes in access to a competing platform. When drivers gained access to one additional competitor, their responsiveness to Uber's pay nearly doubled and the employer power implied by the researchers' estimates fell substantially. The work itself had not changed. What changed was the existence of another door.
The employer has an exit too.
What happens if the worker leaves?
Until recently, replacing a departing worker usually meant finding another person: somebody cheaper, younger, offshore, recently graduated or sufficiently frightened by rent to accept the conditions. AI changes that calculation without needing to replace an entire occupation. It only has to absorb enough tasks that the remaining work can be reorganized around fewer people.
Managers are already trained to think this way without any help from AI. Nobody should be indispensable. Document the process. Train a backup. Cross-train the team. Make sure one engineer cannot take the company hostage by resigning on Thursday. As organizational advice, this is sensible. A business that collapses because one accountant gets sick is badly designed.
The same managerial virtue has an economic shadow. Every procedure captured, every backup trained and every dependency removed also reduces some part of the scarcity from which an employee derives bargaining power. AI enters an organization already accustomed to treating replaceability as resilience.
Take an accountant. Call her Sandra because nobody should have to carry a theory of labor economics under their real name.
Sandra has worked at the same company for eleven years. She knows which supplier always enters the wrong tax code, which subsidiary sends numbers three days late, which spreadsheet contains a formula nobody wants to touch and which executive develops a passionate interest in accounting rules near the end of a bad quarter. Some of this knowledge is written in manuals and old emails. Much of it is the residue of eleven years spent cleaning up other people's mistakes.
Then the company buys better software. It cannot do Sandra's job, but it can reconcile transactions, classify invoices, search old documents, draft explanations, prepare reports and answer routine questions from junior employees. Reality still creates cases strange enough to require Sandra.
A year passes. Two accountants leave and only one is replaced. The department continues closing its books every month, people complain about being busy, the software improves, and the missing person's work gradually disappears into a mixture of machines and remaining employees.
Sandra still has her job.
When salary discussions arrive, she can point to eleven years of experience and several recruiters who have contacted her. Her manager may agree that she is worth more money. He also knows something that was not true several years earlier: if Sandra leaves, the remaining accountants, a junior hire and the software can probably keep the department alive. The first few months might be ugly. After that, less ugly.
Sandra has not become useless. She has become less necessary.
Fewer people can be enough
The evidence available in 2026 does not show companies emptying offices because AI has suddenly become competent enough to operate without people.
A U.S. Census Bureau study found that 18 percent of firms were using AI in at least one business function during the November 2025 to January 2026 survey period, rising to 32 percent when weighted by employment. Among adopters, 66 percent reported using AI only to augment tasks, and AI-related employment decreases appeared in just 2 percent of firms.
Those numbers make the current moment look much less dramatic than the advertising around AI.
The same research contains a more interesting distinction. Worker-level AI use was not associated with lower employment once deeper organizational deployment was considered, but broader adoption across company functions and operational investment was associated with employment decreases. An employee using a tool and a company rebuilding its operation around the tool are different events.
"Augmentation" can hide that distinction. A hammer augments a carpenter, but an industrial system full of machines can still reduce the number of carpenters required to produce the same output.
The reduction does not need to happen on installation day. Somebody retires and is not replaced. Another employee resigns and the vacancy stays open for six months. Management discovers that the delay causes less pain than expected, so the position disappears from next year's budget. The organization becomes thinner through decisions that make sense one at a time.
Nobody was fired because of AI. Everyone who remained employed can say, truthfully, that AI did not take his job. The company can make the same claim without lying.
There are still fewer chairs. And I cna’t help but wonder: is this just a pattern that repeats in the history of employer-employee relationship every time a major scientific or economic shift happened?
The people who never get fired
Some workers affected by AI may never lose a job because they never get one.
Erik Brynjolfsson, Bharat Chandar and Ruyu Chen have been tracking payroll records for millions of U.S. workers. In their August 2026 revision they report no widespread economy-wide job displacement associated with AI. They also find that workers aged 22 to 25 in highly AI-exposed occupations had fallen substantially behind similarly aged workers in less-exposed occupations. By June 2026 the measured gap was about 19 percent relative to the path of the less-exposed group, with most of the difference appearing through reduced hiring rather than increased separations. The authors describe these as early descriptive patterns, not causal estimates.
A separate Census Bureau study using employer-employee administrative records finds a related pattern. Lee Tucker documents a sharp decline in hiring of workers aged 22 to 24 in the most AI-exposed industry-state groups after ChatGPT's release, with weaker hiring accounting for most of the later relative employment decline. The paper also finds that monetary-policy shocks explain part of the weakness, not all of it.
These numbers will change.
They should.
We are measuring a moving technology inside a moving labor market. Definitions of AI exposure will change. Adoption will change. Interest rates will change. Companies will discover that some systems work and others merely generate expensive enthusiasm.
The mechanism matters more.
Nobody needs to fire the apprentice if the apprentice is simply never hired.
The missing apprentice
Junior workers are cheap partly because they are not yet senior workers.
They review documents, prepare first drafts, run basic analyses, sit quietly in meetings, fix small errors and ask questions that appear obvious to anybody who has already spent ten years doing the work. A surprising amount of this looks automatable because it is codified.
It is also how people learn.
The Stanford analysis reports a split between codified knowledge, which can be written in manuals, textbooks and procedures, and tacit knowledge accumulated through practice, mentorship and repeated encounters with situations that refuse to follow the manual. Employment weakness among younger workers appears more heavily in occupations built around codified knowledge, with experienced workers holding up better where tacit knowledge carries more value.
A firm can look at a junior position and reasonably decide that software already performs much of the work. Eliminating the position saves money today.
Five years later the same firm may need somebody with five years of experience.
It can hire that person from somewhere else.
Every firm cannot do this at the same time.
The labor market does not keep senior accountants, lawyers, engineers and managers on shelves waiting for somebody to submit a requisition. Somebody has to pay for the years in which beginners are slower, annoying and occasionally expensive. Remove too much of the bottom rung and companies may discover that they also damaged the mechanism that produced the people whose tacit knowledge remains hardest to automate.
This complicates the simple story about bargaining power. AI could make junior employees easier to replace and experienced workers more scarce at the same time. Some senior workers could become more valuable because fewer replacements are coming behind them. Firms might respond by paying more, inventing new forms of apprenticeship, using AI to accelerate training or pushing automation deeper into the tacit work that keeps experienced people valuable.
AI may concentrate necessity into a smaller number of humans.
When the machine learns from the best worker
The same contradiction appears inside the company.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,172 customer-support agents after the introduction of a generative AI assistant. In the final published study, access to the system increased productivity by about 15 percent on average, with the largest gains among less experienced and lower-skilled workers. Highly experienced workers gained much less. The researchers also found evidence that the system helped spread successful practices through the workforce.
From the company's side this is excellent. New workers become productive faster. Customers get answers sooner. Supervisors spend less time fixing basic mistakes.
Look at the same system from the chair of the experienced employee.
Part of what made her valuable was accumulated failure. She had heard the strange request, made the embarrassing mistake, discovered which official procedure failed and learned which tiny warning meant a small customer problem was about to become expensive. The company once had to keep paying rent on that knowledge because much of it lived inside her.
Now some of it lives somewhere else.
She may still be the best employee. She may become more productive with the same machine.
Her scarcity can fall anyway.
Much of the history of skilled employment trained us to expect productivity and bargaining power to move together. Learn something difficult, become more productive, become harder to replace, get paid more. AI can interfere with that sequence by making expertise easier to distribute.
A worker can produce more and become less scarce at the same time.
The human in the loop
This is why I have become suspicious of the reassuring phrase human in the loop.
The machine drafts and the human approves. The machine recommends and the doctor signs. AI prepares the analysis and the accountant verifies it. An automated agent deals with the customer until the customer becomes sufficiently angry, confused or unusual to require a person.
A human remains.
That tells us very little about the human's bargaining power.
People may remain because regulators require signatures, liability needs somebody's name, customers prefer another human during an emergency, physical objects stubbornly refuse to become software or machines fail just often enough that somebody must clean up afterward.
Some humans inside these systems could become extraordinarily valuable. A small number of people supervising vast automated processes may sit at the last point before an expensive failure, making their judgment scarce enough to command salaries far above today's.
Others may remain employed under almost opposite conditions. If the remaining human tasks can be performed by many people and the company requires very few of them, employment survives without much leverage. A warehouse can still employ people after automation. A newspaper can still employ journalists after one journalist produces what several once did. A hospital can still employ doctors after machines absorb documentation, monitoring and pieces of diagnosis.
The interesting number is not how many humans remain employed.
It is how many humans the organization cannot afford to lose.
We barely measure that number. Workers notice it anyway. They see how long the vacant desk remains empty, whether the colleague who resigns gets replaced, whether a manager still becomes nervous when somebody important threatens to leave.
The machine has two doors too
There is a serious problem with everything I have argued so far.
Workers get machines too.
An accountant who leaves with access to similar AI capabilities may be able to serve dozens of small clients herself. A programmer may need two partners instead of twenty employees to build a company. A designer may need less of an agency. A consultant may need fewer analysts. A local business can gain access to research, translation, software and marketing work that previously required outside specialists.
AI can reduce the firm's dependence on the worker and the worker's dependence on the firm at the same time.
Daron Acemoglu, David Autor and Simon Johnson distinguish technologies that automate existing tasks from technologies that augment workers, spread expertise or create new human tasks. New task creation matters because it can increase demand for human capability rather than merely making an existing capability cheaper.
This sends the argument toward ownership.
Can the employee take her AI tools when she leaves? Can she take the customer relationship? Does she own the reputation accumulated through years of work, or does it belong to a platform? Can she afford the same models outside the corporation? Who owns the proprietary data that makes the internal system useful? Does the machine remember what the worker taught it after she walks through the door?
An employee equipped with a portable machine may gain an exit. An employee who spends three years teaching her employer's machine how she works may be building the memory that makes her own departure cheaper.
The machine is the same. The door is not.
This is where the ownership question from the previous essay leaks into this one. If powerful AI becomes cheap, portable and available to individuals and small firms, workers may gain outside options strong enough to offset some of the employer advantage described here. If the strongest systems depend on proprietary data, expensive compute, privileged distribution and deep integration into corporate infrastructure, the gains may accumulate somewhere else.
The future of bargaining power may depend less on whether AI becomes intelligent enough to replace us than on who gets to leave the building with it.
The evidence can still say no
There is a more basic objection.
We do not yet have good evidence that generative AI has reduced bargaining power across the economy.
Anders Humlum and Emilie Vestergaard linked large surveys of AI adoption in Denmark with administrative labor records. By their March 2026 revision, they found widespread employer initiatives, reported productivity gains and substantial changes in tasks, yet essentially no effect on earnings or recorded hours two years after ChatGPT's launch. Their estimates are precise enough to rule out effects larger than about 2 percent in their setting.
Work changed.
Pay and hours barely moved.
The Census evidence also says current employment reductions among AI adopters are rare. The Stanford payroll study finds a troubling pattern among younger workers, but its authors explicitly stop short of claiming AI caused all of it.
These findings are not debris to clear from the path of the argument. They set its present boundary.
We are early. Firms are still learning where these systems work, workers are still learning when to trust them, and reorganizing a company usually takes longer than buying software. The labor market is being moved at the same time by interest rates, demographics, immigration, remote work and sector-specific booms and collapses.
Perhaps bargaining power will not fall in the way I am describing. AI may create enough new tasks, businesses and independent work that employees gain outside options faster than employers gain substitutes. Senior skills may become scarcer. Productivity gains may be shared through higher wages because firms still compete aggressively for people. The current weakness in entry-level hiring may turn out to be a temporary adjustment.
The evidence is allowed to say no.
This is not an announcement that humans have already lost their economic leverage. It is an argument about where to look before unemployment makes the change impossible to ignore.
Exploitation had an inconvenient limit
There is something perversely comforting about the old conflict between capital and labor. The factory owner wanted cheaper labor and workers wanted more of the value they produced. They could despise each other, deceive each other, organize, lock each other out and occasionally send police or bricks across the factory gate, but the conflict itself exposed their dependency. During a strike machines stopped, ships waited and orders accumulated at the same time wages stopped before rent and groceries did. Workers usually suffered sooner, sometimes much more severely, but damage could travel in both directions because production still required them.
The strike contained one sentence: You need us.
Automation does not have to make that sentence false. Making it less true is enough.
If ten humans can perform what once required twenty, the remaining ten may still be necessary. Some may become extremely well paid because their tacit knowledge sits exactly where machines continue to fail. The missing ten still change the negotiation. A resignation can be absorbed more easily, a vacancy can stay vacant longer, a strike can become less expensive to survive.
In another profession the missing junior pipeline may do the opposite and make experienced workers harder to replace. There will probably be no single experience called "AI and labor." The same technology can weaken one worker, enrich another and prevent a third from entering the profession at all.
Usefulness is the wrong threshold.
I do not expect humans to become economically useless, and nothing in the evidence requires that claim. A person can remain useful, productive and employed after becoming easier to replace. A profession can survive after its entrance narrows. A company can employ thousands of people and depend critically on only a fraction of them. An economy can report respectable unemployment while the price of saying no changes inside private negotiations that never appear in government statistics.
The obsession with the day AI "takes all the jobs" may be looking too far ahead at a day that never comes. Economic change is less considerate. It arrives as one vacancy that does not open, one resignation that is not backfilled, one senior employee whose methods become part of the software, one junior employee who never receives the boring assignment from which expertise used to grow, one manager who discovers that losing somebody hurts slightly less than it did last year.
Each decision makes sense by itself.
The employee enters the manager's office and asks for more money. The manager listens. Nobody threatens him with an AI replacement. He calculates what happens if the employee walks out, and the employee performs the same calculation from the other chair.
They reach an agreement.
Two people leave the room still employed.
The worker still has his chair.
What disappeared was the empty one he could threaten to leave behind.
The ideal resilient company and the ideal powerful worker want opposite things. The company wants nobody indispensable. The worker's bargaining power begins precisely where he is difficult to replace.
Sources
1. Sydnee Caldwell and Emily Oehlsen, “Outside Options and Labor Supply: Evidence from the Gig Economy,” NBER Working Paper 35626, August 2026. Used for the relationship between competing work opportunities and employer wage-setting power.
https://www.nber.org/papers/w35626
2. Kathryn Bonney, Cory L. Breaux, Emin Dinlersoz, Lucia S. Foster, John C. Haltiwanger and Aditya A. Pande, “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks,” U.S. Census Bureau CES Working Paper 26-25, April 2026. Used for firm AI adoption, augmentation, employment reductions and the distinction between employee-level use and deeper organizational deployment.
https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-25.html
3. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab, revised August 12, 2026. Used for the early-career employment gap, reduced-hiring result, codified-versus-tacit-knowledge distinction and causal cautions.
https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/
4. Lee C. Tucker, “You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators,” U.S. Census Bureau CES Working Paper 26-27, April 2026. Used as separate administrative-data evidence on early-career hiring and for alternative explanations including monetary policy.
https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-27.html
5. Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, “Generative AI at Work,” The Quarterly Journal of Economics, Vol. 140, No. 2, May 2025, pp. 889–942. Used for the productivity effects among customer-support agents and the larger gains among less experienced workers.
https://academic.oup.com/qje/article/140/2/889/7990658
6. Anders Humlum and Emilie Vestergaard, “Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI,” NBER Working Paper 33777, revised March 2026. Used for Danish evidence of significant task changes alongside near-zero measured effects on earnings and recorded hours during the first two years.
https://www.nber.org/papers/w33777
7. Daron Acemoglu, David Autor and Simon Johnson, “Building Pro-Worker Artificial Intelligence,” NBER Working Paper 34854, February 2026. Used for the distinction among automation, worker augmentation, expertise-leveling and creation of new human tasks.
https://www.nber.org/papers/w34854