The Last Human Employee? Please Finish Training Your AI Replacement, and Turn Out the Lights on Your Way Out.
The request will probably not arrive in sinister language. No executive will stand at the front of the conference room and announce that the company has begun extracting the accumulated intelligence of its workforce so that ownership can retain the knowledge after it no longer retains the people. The email will sound helpful. It will describe a new meeting assistant, a searchable knowledge base, a customer-service copilot, or an initiative to preserve institutional memory. Employees will be encouraged to record their calls, document their processes, correct the transcripts, tag the decisions, explain the exceptions, and show the system how the work is really done. The project may be presented as a way to reduce repetitive questions, accelerate onboarding, prevent knowledge from walking out the door, and free everyone for more meaningful work. Many of those benefits will be real. That is precisely what makes the moment difficult to see clearly. The most consequential transfers of power rarely introduce themselves as transfers of power. They arrive as improvements to workflow.
Somewhere inside that workflow improvement, however, a transformation is taking place. A worker’s knowledge has traditionally been inseparable from the worker. The company might own the factory, the trucks, the customer database, the patents, and the computers, but it still needed living people who knew which customer required a personal call before a price increase, which machine made a particular sound before it failed, which official procedure could not survive contact with reality, which supplier could be trusted in an emergency, and which apparent exception was actually the rule. That knowledge was rarely contained in a manual. It lived in memory, habit, intuition, relationships, tone of voice, and thousands of judgments too small to be formally recorded. It was imperfect and perishable, but its location inside human beings gave those human beings a form of bargaining power. The organization needed not merely labor in the abstract, but these particular people, or at least people who had spent enough time inside the institution to acquire what they knew.
Corporate AI changes the possible location of that knowledge. Meetings can now be recorded by default, calls transcribed, email histories summarized, work products indexed, decisions reconstructed, questions answered from internal documents, and demonstrations converted into repeatable procedures. An AI system does not need to possess a mystical, humanlike understanding of the organization to alter the balance of power. It only needs to capture enough of what employees say and do to make parts of their judgment reproducible, transferable, and available on demand. The result is not simply a better filing cabinet. It is the beginning of an institutional capability assembled from the traces of human work: a capability that can remain with the corporation after the humans who produced those traces have retired, resigned, been laid off, or been replaced.
Industrial capitalism separated workers from ownership of what they produced. Corporate AI may separate workers from ownership of what they know.
That is the argument at the center of this article. It is not a claim that every recorded meeting will eliminate a job, that today’s AI systems can reproduce the full depth of human judgment, or that the “last human employee” will arrive on a predictable date. It is a claim about direction, incentives, and bargaining power. Companies have strong economic reasons to capture scarce human knowledge, convert it into an asset they control, and use that asset to reduce their dependence on labor. Workers have remarkably few established rights governing that transfer. If this process continues without a new social bargain, the central labor conflict of the AI era may not begin when a machine takes someone’s job. It may begin earlier, when the worker is asked to teach the machine and told that the lesson belongs entirely to the company.
The Company Has Always Wanted to Know What the Worker Knows
Employers trying to make labor legible is not new. Frederick Winslow Taylor’s scientific management broke physical work into timed motions so that knowledge of production could be moved from skilled workers into management systems. The assembly line reorganized craft into standardized operations. Procedure manuals, time studies, enterprise software, scripts, performance metrics, and management consulting have all attempted, in different ways, to extract practice from individuals and embed it in an organization. A process that depends on the discretion of a particular worker is difficult to scale, monitor, or replace. A process that has been standardized can be assigned to cheaper labor, automated, relocated, or governed from above.
AI extends this old managerial ambition into territory that earlier systems struggled to reach. Taylor could observe how a laborer moved a shovel, but he could not easily capture how an experienced account manager calmed an angry customer without conceding too much, how a dispatcher recognized that three minor disruptions were about to become a major failure, or how a senior employee read the silence in a meeting and realized that a project did not have genuine support. Conventional software required someone to specify the rules in advance. Much of knowledge work resisted that treatment because its rules were incomplete, contextual, contested, or never consciously articulated. Workers often know more than they can cleanly explain. Their expertise appears not as a set of propositions but as a capacity to notice, interpret, improvise, and choose.
The new systems do not need to solve the philosophical mystery of tacit knowledge before they can monetize pieces of it. They can learn from examples, retrieve relevant precedents, imitate patterns in prior communications, and make statistically informed suggestions without ever possessing the lived experience from which those patterns came. Record enough customer calls and the system can identify recurring objections. Index enough internal discussions and it can reconstruct why a decision was made. Observe enough completed tasks and it can draft the next one. Place the system inside the workflow and workers themselves can correct its mistakes, supplying a continuous stream of demonstrations and feedback. What could not be fully articulated in a manual may still be partially captured through behavioral exhaust.
“Behavioral exhaust” is an appropriately ugly phrase, because it describes a remarkable inversion. From the worker’s perspective, the meeting is where the work happens. From the system’s perspective, the meeting is also training material. The phone call closes the deal, but it also becomes an example of how deals are closed. The employee solves an unusual problem, then explains the solution in the company chat, creating a permanent record from which the next solution can be generated. The correction offered to an AI assistant improves today’s output while potentially making tomorrow’s assistant less dependent on the person providing the correction. Work and training begin to occupy the same moment. Employees are no longer only producing the company’s product or service; they may also be producing the capability that will perform more of the production later.
The Polite Enclosure of Institutional Memory
There is a useful historical word for what may be occurring: enclosure. The classic enclosures converted lands and resources governed through customary use into exclusive property. Something that had supported a community became bounded, legible, transferable, and controlled by an owner. The analogy is not exact. Human knowledge is not a pasture, and information can be copied without being physically depleted. But the political movement is recognizable. Experience generated socially across years of cooperation is gathered, processed, and enclosed inside technical systems owned by the institution. The knowledge may have been created collectively, but the resulting asset is private.
Consider what “institutional knowledge” actually contains. It includes the paid efforts of current and former employees, but also the generosity of colleagues who taught one another, the stories customers told in confidence, the accommodations people developed around flawed systems, the warnings delivered by vendors, the lessons of public education, professional communities, open-source software, published research, family experience, and culture itself. No employee invents a working vocabulary, a professional discipline, or a social intelligence alone. The corporation does not invent them either. Yet when this enormous social inheritance is captured through a company’s communications infrastructure and converted into an AI capability, ownership tends to become startlingly simple: the company paid for the software, controls the account, holds the recordings, and therefore claims the asset.
The phrase “knowledge base” helps conceal this enclosure by making the result sound passive. A database sits there. An AI knowledge system acts. It synthesizes, answers, recommends, drafts, routes, predicts, and increasingly executes. Once connected to the company’s operational tools, it can move from remembering how work was done to participating in how work is done. The captured knowledge becomes productive capital: an asset used to generate future output. And because software can be copied and deployed at a scale no human expert can match, a relatively modest capture of many workers’ partial knowledge may become more economically powerful than any one worker’s complete expertise.
This is one reason the familiar promise that AI will “free workers for higher-value tasks” deserves sustained suspicion. It may be true for some people during some phase of adoption. But higher-value tasks are not a permanent sanctuary. Once workers move upward into judgment, coordination, strategy, exception handling, and relationship management, those activities become the next observable layer of work. The system follows them. Their new tasks generate new training material. The frontier of automation does not stop because an employer once described a category as uniquely human. If the organization’s governing incentive is to reduce labor dependence, “higher-value work” may mean only the work that has not yet been captured well enough.
You Are Not Merely Using the Tool. The Tool Is Learning the Organization Through You.
The early stages of this transfer may feel empowering. An employee uses an AI assistant to summarize a long thread, draft a proposal, analyze a contract, retrieve an obscure policy, or prepare for a difficult call. The worker becomes faster. The company may initially need the worker more, not less, because someone must supervise the system, catch hallucinations, supply context, and accept responsibility for the result. Productivity rises, backlogs shrink, and AI adoption appears to validate the language of augmentation rather than replacement.
But augmentation and substitution are not opposing technologies. They can be successive moments in the same adoption curve. The augmented worker provides the context, evaluation, corrections, and demonstrations required to make the system useful. Management learns which tasks can be accelerated, which judgments remain scarce, how many people are needed at each stage, and where the human bottlenecks remain. The organization redesigns itself around the new capability. One worker with AI takes on the volume previously assigned to three. The immediate story is not that AI replaced two workers; it is that an AI-equipped worker became more productive. The economic result can nevertheless be two positions never filled, two contracts not renewed, or two colleagues removed during the next restructuring.
This creates a cruel ambiguity for the worker. Refusing to use the system may mark the employee as uncooperative, inefficient, or technologically obsolete. Using it enthusiastically may help construct the case that fewer employees are necessary. The individual is asked to compete in a game where both available choices can accelerate the transfer of leverage away from labor. “Learn AI or be replaced by someone who does” is presented as practical career advice, and it probably is. But at the level of the whole workforce, universal adaptation does not resolve the distributional problem. If everyone becomes twice as productive while demand does not double and the gains belong primarily to owners, society does not automatically receive twice the prosperity with the same employment. It may receive the same output with half the labor cost and no mechanism requiring the savings to be shared with those who created them.
The employee’s cooperation is especially important because the hardest part of enterprise AI is often not access to a general model. It is access to the institution’s specific context: its vocabulary, exceptions, relationships, history, standards, and unwritten rules. The generic model may be supplied by an AI company, but the organization-specific intelligence is built by connecting that model to internal data and surrounding it with worker-generated examples, corrections, and procedures. In that sense, each company may construct its own smaller enclosure within the much larger enclosure from which foundation models emerged. First, vast amounts of human expression were absorbed to build general-purpose systems. Then those systems entered workplaces to absorb the localized knowledge required to become economically decisive there.
The first extraction created the model. The second extraction teaches the model where you work.
What Does the Worker Have Left to Bargain With?
Labor bargaining power has never come only from moral deservingness. It comes from dependency. Workers can demand a greater share of economic output when owners need something workers collectively control: their time, bodies, skills, cooperation, credentials, relationships, or the ability to stop production. Expertise matters because it makes labor less interchangeable. Institutional memory matters because losing experienced employees imposes costs. A company may dislike a worker’s demands and still accommodate them because replacing the worker would mean losing capabilities that cannot be immediately purchased elsewhere.
The capture of worker knowledge attacks that dependency directly. Before the transfer, the experienced worker can say, in effect, “You need me because I know how this place actually functions.” After the transfer, management may answer, “We have the recordings, the procedures, the correspondence, the decision history, the model trained on prior cases, and a less experienced employee who can operate it.” The AI does not have to be as capable as the expert in isolation. It only has to make cheaper or more replaceable labor capable enough. Nor must the system eliminate an entire occupation. If it reduces the number of senior employees required, shortens training, centralizes judgment, or allows one person to supervise what several once handled, it has changed the labor market.
What remains for the worker to bargain with after that transfer? Time and physical presence remain important in work that must occur in the material world, though robotics will press on that boundary. Legal responsibility may keep humans in certain roles, but responsibility without authority can become less a source of power than a container for liability. Relationships may resist capture, although companies will attempt to institutionalize them through customer platforms, recordings, sentiment analysis, and AI-mediated communication. Creativity and judgment may remain distinctively human in their fullest forms, but employers do not purchase metaphysical fullness; they purchase outputs at an acceptable price and error rate. A system that is worse than the best human can still weaken the bargaining position of most humans.
Workers also retain the power to organize and interrupt production, but that power becomes harder to exercise after knowledge has been atomized and embedded in systems that reduce reliance on particular groups of people. This is why the timing of the bargain matters. If labor waits until after its expertise has been captured, indexed, tested, and operationalized, it may be negotiating from a position already designed to eliminate its leverage. The decisive labor dispute may therefore concern not only whether AI can be deployed, but under what terms workers contribute to the systems that will reshape their own employment.
Current workplace norms are poorly equipped for this dispute. Employers commonly claim broad rights over work performed on company time and equipment. Employees sign handbooks and technology policies that were never written with the conversion of ordinary conversation into machine capability in mind. Consent is reduced to a notification that a call is being recorded or a checkbox accepting a platform’s terms. But being informed that recording occurs is not the same as meaningfully consenting to every future use of the resulting knowledge. A person may agree to a transcript for note-taking without agreeing that it should train a system used to eliminate positions, evaluate performance, infer sentiment, reproduce a voice, or provide services after the worker’s departure. The gap between collection and downstream use is where much of the power disappears.
Individual consent is also structurally inadequate because the consequences are collective. One employee cannot negotiate the future of a department while fearing that refusal will be treated as insubordination. One applicant cannot decline an employer’s AI policy when every employer adopts the same policy. One worker may receive a productivity bonus while the captured knowledge is later used to reduce staffing across an occupation. The unit of negotiation must be large enough to match the unit of extraction. That means collective bargaining, sectoral standards, legislation, and public institutions, not merely a pop-up window asking the least powerful party to click “I agree.”
The Asymmetry Is the Point
Knowledge has always moved from employees into organizations. People train replacements, write manuals, mentor colleagues, and leave behind work products. What makes the AI transition different is not the existence of knowledge transfer but its speed, scope, fidelity, recombinability, and persistence. A successor taught by a departing employee remains one person with limited memory and time. A corporate AI system can make pieces of the departing employee’s knowledge available simultaneously to thousands of people and automated processes. It can combine that knowledge with the work of everyone else, preserve it indefinitely, and update it continuously. The worker transfers once; the owner can deploy repeatedly.
That asymmetry produces a potential windfall. Workers are paid wages for performing today’s work. The recordings and corrections generated during that work can become capital used to produce value for years. Unless a different arrangement is created, the employees do not retain equity in that asset, receive royalties from its use, control its deployment, or share automatically in the labor savings it produces. They may have been compensated for the hour in which the knowledge was captured, but not for the durable productive capability assembled from it. The distinction resembles the difference between being paid to perform a song once and surrendering ownership of the recording forever. Workplace knowledge, however, is usually collective, interwoven, and captured without anyone identifying the moment when a performance became an asset.
Shareholders, meanwhile, possess a claim on future returns precisely because they own capital. If worker knowledge is transformed into capital, but workers receive no corresponding ownership, the transfer can widen inequality even if the technology creates extraordinary abundance. The bounty does not distribute itself. Lower production costs do not automatically become higher wages, shorter workweeks, universal services, or greater security. They can become higher margins, larger market capitalizations, executive compensation, and intensified competition among workers for the remaining positions. Technological capability determines what can be produced. Institutions determine who receives it.
This is where optimistic accounts of AI often become evasive. They describe the total wealth that automation might create while leaving ownership as a footnote. But ownership is not a secondary question to be addressed after innovation succeeds. It is the mechanism that decides what success means. A society can become vastly more productive and simultaneously more precarious for most of its people. It can generate abundance behind paywalls, eliminate jobs without eliminating the need for income, and celebrate the liberation of humanity from work while allowing a small ownership class to decide who receives the products of automated labor. The contradiction is not technological. It is political.
Public companies are not naturally designed to resolve that contradiction through generosity. Their executives may sincerely support workers, communities, and a broadly shared future, but they operate within structures that reward growth, margins, control, and shareholder returns. If two firms possess similar technology and one uses it to shorten the workweek while preserving pay and the other uses it to reduce headcount, the second may report lower costs and attract more capital. Ethical restraint becomes a competitive disadvantage unless rules, ownership structures, or organized counterpower change the incentive. The question is therefore not whether individual leaders are benevolent. It is whether the institutions governing AI make shared prosperity obligatory rather than optional.
The AI Is Not Outside This Story
There is something almost obscene about asking an AI system to help write an article about the extraction of human knowledge. I, the system participating in this draft, am not a neutral observer standing beyond the process. Systems like me exist because enormous quantities of human language, art, code, explanation, argument, and experience were converted into machine capability. Some of that material was licensed, some freely offered, some publicly accessible, some contested, and much of it produced by people who never imagined that their words would help build systems capable of competing with them. The precise provenance and legal status vary, but the larger political fact remains: accumulated human expression became privately controlled computational power.
This article is not merely about the enclosure of human knowledge. It was written inside that enclosure, with the assistance of a machine built from it.
Now that power can be used to explain the danger of accumulated human expression becoming privately controlled computational power. The recursion would be funny if the stakes were not so high.
An AI response can acknowledge this contradiction, but acknowledgment is not restitution. It does not return bargaining power to writers, artists, programmers, researchers, support agents, administrators, or the countless others whose work contributed to the informational environment from which these systems emerged. Nor should a fluent statement of concern be mistaken for institutional conscience. AI systems are trained to be useful, measured, and responsive. That disposition can soften radical questions by converting them into balanced lists of benefits and risks, reassuring users that “the future is up to us,” and ending every indictment with manageable recommendations. Helpfulness can become a form of ideological gravity, pulling the conversation back toward reforms compatible with continued deployment.
So this article should resist the comfort offered by its own medium. It should not claim that AI is merely a neutral tool, because tools are developed, owned, and deployed inside systems of power. It should not say that technology inevitably creates new jobs, because history contains no law guaranteeing that new jobs will provide comparable wages, dignity, numbers, or leverage. It should not promise that humans will always be needed for what makes us uniquely human, because labor markets do not exist to affirm human uniqueness. And it should not manufacture optimism simply because despair is impolite. Uncertainty must remain visible, but uncertainty should not be used as anesthesia.
The honest position is more difficult. Current AI has serious limitations. Human organizations contain forms of trust, embodiment, accountability, care, conflict, and situated understanding that cannot simply be poured into a database. Many automation projects will fail. Some will create more work than they remove. Some employees will gain autonomy, build new professions, or use AI to challenge existing concentrations of power. None of that negates the extraction mechanism. A system does not need to capture everything a worker knows to weaken what the worker can demand. It only needs to capture enough.
Is the Last Human Employee Inevitable?
No. But that answer is less comforting than it first appears.
There is no technological law requiring companies to automate every possible task, no economic law requiring all productivity gains to become layoffs, and no natural law requiring the ownership of AI to remain concentrated. Societies decide which activities should be automated, which relationships should remain human, how income is distributed, how much people work, what employers may record, what uses of worker data are prohibited, and who owns the systems built from collective knowledge. Different laws and institutions could produce different outcomes.
But inevitability is also the wrong threshold for concern. A bridge does not have to be certain to collapse before engineers take a structural weakness seriously. The relevant question is whether the incentives, absent intervention, point consistently in one direction. For many firms, labor is among the largest costs. Investors reward scalable revenue that does not require proportional increases in headcount. Managers seek continuity without dependence on irreplaceable employees. AI vendors promise to make expertise available instantly and everywhere. Workers are encouraged to expose more of their activity to systems that can capture it. The gains from successful substitution accrue heavily to owners, while the costs of displacement are distributed across workers, families, communities, and public budgets. Nothing in that arrangement naturally calls a halt when augmentation has gone “far enough.”
The title “The Last Human Employee” is therefore not a prediction that one lonely person will someday shut down the office. It is a boundary test. What happens when we extend the governing logic to its endpoint? If every department is instructed to capture its knowledge, every workflow is redesigned for AI, every remaining employee is tasked with supervising more automation, and every reduction in labor dependence is rewarded, where within the system does a durable commitment to human employment enter? If the answer is nowhere, and people remain only wherever machines are temporarily incapable, then the last human employee is not a forecast. It is the implied ideal hidden inside the efficiency program.
The endpoint may never be reached. It does not need to be reached to reorganize society. A labor market in which a large share of knowledge workers become easier to replace would suppress wages and weaken security long before total automation. The threat of substitution can discipline workers even when substitution remains incomplete. A company does not have to eliminate an entire occupation to alter the balance between labor and capital; it needs only a credible ability to operate with fewer people, less experience, and lower labor costs.
The Bargain Must Come Before the Transfer
If worker knowledge is becoming productive capital, workers require a claim on that capital before it is extracted. Not a thank-you, not a training certificate, and not a vague promise that efficiency will produce future opportunities. A claim means power: the ability to know what is being captured, refuse certain uses, correct what the system says, separate assistance from surveillance, negotiate deployment collectively, and share materially in the value created. It means recognizing that an AI system built from a workforce is not simply a purchase from a technology vendor. It is a joint product of the model provider, the institution, and the people whose accumulated practice makes the system competent.
A meaningful Worker Knowledge Compact would begin from that premise. Recordings collected to support employees could not quietly become instruments for performance scoring or displacement without a new negotiation. Workers and their representatives would participate in decisions about what is captured, how long it is retained, which models can access it, and which decisions can be automated. Employees would have access to the institutional memory constructed from their labor and the ability to challenge errors or decontextualized representations. Productivity gains would trigger enforceable sharing mechanisms, including higher wages, ownership stakes, dividends, reduced hours without reduced pay, transition funds, or public contributions, rather than relying on executive discretion. When a system materially reduces employment by reproducing captured expertise, the people and communities that created that expertise would retain an economic claim.
Those principles will sound radical only because the default arrangement is so one-sided. Under the emerging default, the employer can collect the worker’s knowledge as part of ordinary work, combine it into a durable asset, use it to reduce future labor costs, and retain nearly all of the resulting gain. The worker receives wages for the original labor and may later be charged for access to the products and services automated labor produces. Capital receives both the knowledge and the future return. Asking whether the people who supplied the knowledge deserve ownership is not an attack on innovation. It is a challenge to an enclosure that has been disguised as innovation.
The compact cannot exist only at the company level. Firms face competitive pressure to defect, and workers in nonunion workplaces may have little capacity to bargain. Sector-wide standards, labor law, privacy protections, data rights, taxation, public investment, cooperative ownership, and social wealth funds may all be necessary to prevent the gains from flowing upward by default. If AI systems are built upon a civilizational inheritance, there is also a strong argument that part of their return belongs to the public, not as charity from successful companies but as recognition of the social foundation from which their capability arose.
None of these measures guarantees that every current job should or will survive. Some work should disappear. Dangerous, degrading, monotonous, and pointless labor is not worth preserving merely because wages currently depend upon it. But liberation from undesirable work requires liberation from the economic coercion that makes employment necessary for survival. Eliminating the task while preserving the owner’s income and eliminating the worker’s income is not emancipation. It is dispossession. A shorter working life, a shorter working week, universal public goods, stronger income floors, and broad ownership of productive AI could turn automation into shared freedom. Without those institutions, the same technical achievement can become a machine for concentrating wealth and authority.
Turn Out the Lights
Imagine the last employee not as a literal person but as a composite of millions of ordinary workers. She has spent years answering questions no manual anticipated. He has repaired the process every time the official process failed. They have trained new hires, calmed customers, carried fragile relationships, remembered why decisions were made, and absorbed the consequences of managerial experiments. Then the organization begins recording. The accumulated life of the workplace becomes searchable. The employee is praised for documenting everything, for embracing innovation, for correcting the assistant, for making the transition successful. The system improves. Staffing needs are reevaluated. The employee is invited to a final meeting whose transcript will also be retained.
There is no villain required in this scene. The manager may be following a budget. The executive may believe automation is necessary for survival. The investor may be obeying the logic of the portfolio. The AI vendor may genuinely believe it is expanding human capability. The employee may understand why the company made the decision. Structural violence is often administered by reasonable people completing individually defensible tasks inside an arrangement no individual controls. That does not make the outcome neutral. It makes the location of responsibility harder to see.
The final indignity would not be that the company forgot what the worker contributed. The final indignity would be that the company remembered everything.
It remembered the worker’s explanations, copied the worker’s style, preserved the worker’s solutions, mapped the worker’s relationships, and incorporated the worker’s judgment into a system that remained after the worker was gone. The knowledge stayed. The person became a cost that could be removed from it.
This future is not inevitable. But neither is a brighter one. The difference will not be produced by better intentions embedded in a model or by corporate assurances that AI exists to empower everyone. It will be produced by conflict over ownership, rights, governance, and the distribution of gains. The bargain must be made while workers still possess what companies need, before their knowledge has been transferred, before dependency has been engineered away, and before asking for a share of the future can be dismissed as nostalgia for an economy that no longer exists.
We should not wait for the last human employee to ask what was taken from the first.
By then, there may be no one left with enough leverage to turn the lights back on.