The Employee Who Never Sleeps

In 2025, Klarna reported something that would have sounded implausible only a few years earlier. Its artificial intelligence customer service assistant was handling most of the company’s customer service chats, operating around the clock across dozens of markets and languages. Klarna estimated that the system was doing work equivalent to more than 700 full-time agents and had delivered tens of millions of dollars in annual cost savings.

The obvious headline was replacement. The more important story was unbundling: separating a job into tasks that can be assigned differently. Software might answer routine questions while a person handles a difficult complaint.

Klarna had acquired a large amount of customer service capacity without hiring seven hundred additional people to provide it. The work could be performed at two in the morning, expanded across languages, and increased without recruiting a parallel human workforce. Yet the company also continued to offer human support, especially where customers wanted or needed it. The result was neither a purely human workforce nor a purely automated one. It was a different economic arrangement.

Klarna then supplied the second half of the story. In May 2025, chief executive Sebastian Siemiatkowski said the company had allowed cost to become too dominant a consideration in customer service and that quality had suffered. Klarna resumed recruiting human agents for more complex and premium interactions while keeping the AI assistant on the high-volume tier. By the third quarter of 2025, company materials estimated that the assistant was doing work equivalent to 853 full-time agents and delivering about $58 million in annualized cost savings. The point was no longer replacement versus retreat. It was rebalancing the bundle after learning which parts of customer service could be detached from employment and which still benefited from a human relationship.

The title, “The Employee Who Never Sleeps,” is a metaphor. An AI agent is not an employee. It has no employment contract, legal personhood, paycheck, family, pension, sick day, ambition, or right to workplace protection. It cannot be loyal or accept moral responsibility for the consequences of its work.

Artificial intelligence can increasingly provide something that once came bundled almost exclusively with human employment: useful labor capacity. The technology does not eliminate the need for people. It separates portions of work from the employment relationship that traditionally supplied them.

The first economic transformation of the AI age may therefore arrive before mass job elimination. It may be the emergence of labor without employment.

A Job Is More Than Its Tasks

A job looks simple on an organizational chart. A box has a title, a manager, a salary band, and a list of responsibilities. In practice, the box contains far more.

A human employee performs tasks, but also notices problems that were not assigned, interprets ambiguous instructions, remembers why a strange process exists, teaches new colleagues, manages relationships, absorbs organizational culture, complains when a goal makes no sense, and occasionally refuses to do something foolish. An employee develops judgment through repetition and error. The employee can be praised, disciplined, promoted, fired, sued in some circumstances, licensed in others, and asked why a decision was made.

Employment is therefore a bundle. We pay for work, but a job also gives people practice, relationships, shared knowledge, and a place in an organization. The worker develops judgment and has duties within a legal relationship with the employer. Moving a task to software does not automatically replace everything that came with the person doing it.

A drafting task can move to a model while the lawyer remains employed. A customer service exchange can move to an agent while human representatives handle exceptions. A software engineer can delegate test generation while keeping responsibility for the codebase. A finance team can assign reconciliation work to an automated system while humans retain approval authority. A manager can ask an agent to monitor reports overnight while the manager remains responsible for deciding what the reports mean.

The unit being automated is often smaller than the job.

This matters because public arguments about AI and employment are usually framed as a contest between two outcomes: the person keeps the job or the machine takes it. Real organizations rarely transform so neatly. They rearrange tasks. They remove some responsibilities, add others, change spans of control, combine jobs, eliminate positions, create new specialties, and discover that work once thought inseparable from a person can be purchased as software.

A job can survive while losing much of its old content. A job can also disappear even though many of its tasks remain, because those tasks are redistributed among machines and fewer people. Employment changes through recombination long before a profession vanishes from the economy.

Labor Without Employment

The phrase labor without employment sounds contradictory because industrial economies trained us to connect productive work with workers.

Machines have broken that connection before. A forklift supplies lifting capacity without adding a warehouse employee. A spreadsheet performs calculations that once occupied rooms of clerks. Industrial robots weld, paint, and move material without becoming members of the workforce.

A general-purpose AI agent can attempt many kinds of work, including reading, writing, searching, comparing information, scheduling, and using other software. Those tasks may belong to several different jobs. Its ability still needs checking for each assignment.

I call this synthetic capacity: useful productive ability that can be added without adding a corresponding human position. Think of a small team using software to prepare routine reports overnight. The team gains help with a task; it still needs someone to judge and use the results.

The economic temptation is obvious. Human capacity is relatively fixed in the short term. Hiring takes time. Training takes time. People sleep. They take vacations. They leave. They negotiate compensation. They have rights. They deserve those rights because employment is a relationship between human beings and institutions.

An organization can often add software capacity faster than it can hire and train people. It might buy a larger subscription, pay for more computer processing, or run another copy of an agent. Software can also work through the night, although people still need to arrange support and decide what it may do while they are away.

That capacity has costs beyond the subscription. People must connect the system to the information and programs it needs, protect access, test its work, maintain it, and deal with failures. Mistakes create repair work. Automation can also repeat a mistake quickly and widely. Much of this expense sits outside the price of the AI service itself.

Cost is not only the price printed on the invoice. It is also the friction built into the delegation. A system may be useful enough to justify paying for, yet still make the human user stop repeatedly to buy small increments of access. That design changes the work. The user cannot easily plan the session, judge the real cost of completion, or maintain concentration through a long project. In that sense, the platform is not merely selling computation. It is controlling the rhythm of the delegation. A responsible delegation system should let the person understand the likely cost, authorize a meaningful work budget, and finish the task without artificial interruption.

The 168-Hour Temptation

There are 168 hours in a week.

A human worker does not and should not work all of them. An automated agent can remain available through all of them.

Availability is not the same thing as useful productivity. A confused system can produce nonsense continuously. An unreliable agent working overnight can create a morning full of cleanup. Yet continuous availability changes the economics of many tasks because time itself becomes less scarce.

A research request can continue after the person who assigned it goes home. A monitoring agent can watch incoming information through the night. Software testing can continue between engineering shifts. Customer questions can be answered across time zones without building an identical human team in each one. Reports can be assembled before a manager arrives.

The attraction is stronger when the work is uneven. Human staffing requires organizations to plan around peaks. A company needs enough people for the busiest hour even if the quiet hours waste capacity. Software can often scale more elastically.

The comparison changes before the human worker becomes obsolete. Managers begin asking not only whether AI can perform a task as well as a person, but whether the organization needs another person to supply the next unit of capacity at all.

A team that would once have grown from ten people to twelve may remain at ten and add agents. A department that loses two people may replace one and automate part of the remaining work. A successful company may increase output without increasing headcount at the old rate. None of those decisions looks like a robot walking into an office and taking someone’s chair. Collectively, they can reshape the labor market.

The employee who never sleeps can change the next hiring decision, even when nobody loses an existing job. A manager may buy more software capacity instead of filling the next vacancy.

From Headcount to Capacity

For generations, managers have used headcount as a rough proxy for organizational capacity. More customers require more service representatives. More software requires more engineers. More contracts require more lawyers. More reports require more analysts. AI weakens that relationship.

A future manager may supervise ten people and fifty agents, although the word supervise will mean something different for each. The people bring judgment, relationships, accountability, institutional knowledge, creativity, and the capacity to decide when the plan itself is wrong. The agents bring speed, persistence, replication, and inexpensive parallel effort. The managerial problem shifts from allocating people to allocating capability.

That sounds efficient, and it can be. It can also create a new bottleneck. Fifty agents can generate more recommendations, drafts, alerts, code changes, leads, summaries, and exceptions than ten humans can intelligently evaluate. Adding machine capacity can increase the demand for human judgment rather than reduce it.

This is one reason the number of agents is a poor measure of organizational maturity. A company with hundreds of autonomous systems may be less capable than a company with twenty well-bounded systems whose outputs feed clear human decisions. Quantity of delegated activity is not the same thing as quality of delegation.

The organization of work therefore becomes an architecture problem. Which tasks should remain bundled inside a human job? Which can be separated safely? Which require a person because the work itself depends on trust or legitimacy? Which can be automated only if an accountable human remains close enough to recognize an exception? Which activities should never be optimized merely because the software can perform them cheaply?

The answers will differ by industry, but the questions travel well.

The New Span of Control

Traditional management assumes that span of control is mostly a human question. How many people can one manager supervise before communication, coaching, review, and coordination begin to fail? AI adds another dimension. A manager may soon control far more productive capacity than the number of people on the team suggests.

That can make one capable person extraordinarily powerful. A project leader can assign parallel research tasks, ask one agent to test assumptions, another to monitor competitors, another to draft alternatives, and another to check the work. The human no longer performs each intermediate step. The human designs the sequence, decides which outputs matter, resolves conflicts, and determines when the work is good enough to act on.

The gain can be enormous, but so can the management burden. Every agent creates another stream of activity that can drift, duplicate work, misunderstand context, or confidently optimize the wrong objective. A manager with too many machine delegates can become the organizational equivalent of an air traffic controller with too many screens and no time to look closely at any of them.

The meaningful span of control is the amount of delegated activity a person can still govern, not the number of agents that person can launch. The limit depends on stakes, reversibility, monitoring quality, and the cost of an error. A hundred low-risk research agents may require less supervision than one system authorized to move money or alter a customer’s legal rights.

This changes the meaning of managerial leverage. The best manager may not be the person who uses the most AI. It may be the person who knows what to delegate, what to keep, how to design the boundary, and when additional machine capacity would create more noise than value.

Headcount once constrained the scale of a manager’s reach. Judgment may become the tighter constraint.

The Cost Illusion

The phrase automated labor encourages another mistake: comparing a salary with a software bill.

A capable agent may cost far less than an employee performing the same narrow task. Organizations, however, buy systems of work rather than isolated tasks.

The full cost of an agent includes preparing its information, connecting it to other software, controlling access, checking its work, repairing mistakes, and protecting the organization against the new security risks it creates. Suppose an agent saves 100 hours of a task but creates 50 hours of checking and repair. That leaves 50 hours saved before other costs are counted. Whether the change is worthwhile also depends on what those hours cost, the quality of the result, and the risks people must carry.

Meta’s 2026 workforce experiment illustrates the danger of treating machine output as a simple substitute for people. The durable lesson is the cost illusion itself: more automated activity can create new categories of review, repair, coordination, and risk that a simple salary-versus-software comparison misses. The dated details of that experiment belong in the Current Evidence layer below.

The employee who never sleeps can also create work while everyone else is asleep. Time saved in one part of a system does not automatically become productive output somewhere else; organizations can absorb the apparent gain in verification, exception handling, incident response, and rework.

Current Evidence: Transformation Before Replacement

Reviewed October 6, 2026.

The strongest available evidence does not yet support a simple story of economy-wide job replacement. It supports rapid task adoption, uneven productivity, organizational experimentation, and early changes in how firms combine human and machine work.

A 2026 U.S. Census Bureau working paper, based on the survey introduced in Chapter Two, found that 18% of firms reported using AI in a business function during the November 2025 to January 2026 reference period. The employment-weighted rate was 32%, reflecting much heavier adoption by larger companies. Among AI-using firms, 66% reported using AI solely to augment tasks, while AI-related employment decreases were reported by only 2% of AI-using firms.

Those numbers should not be read as reassurance that employment will remain unchanged. They show that the current transition is occurring mainly inside jobs and workflows rather than through immediate mass headcount elimination. The same Census research found that broader functional integration and operational investment were associated with a greater likelihood of employment decreases, suggesting that organizational redesign matters more than casual use by individual workers.

The International Labour Organization reached a similar but broader conclusion in a June 2026 review of empirical evidence. It found real but uneven productivity gains and limited large-scale job displacement so far. The review also noted that worker-reported time savings of a few percent of working hours had not yet translated reliably into higher measured output, earnings, or employment. More immediately, the evidence pointed to changes in work organization, coordination, worker autonomy, job quality, and the distribution of opportunity.

Anthropic’s September 30, 2026 robotics study illustrates the distance between demonstrated capability and practical replacement. Using Claude-assisted assessments of occupational tasks and costs, the researchers estimate that robots could perform tasks representing 34% of U.S. working time under some conditions, but would be cost-competitive for only 0.3%. These are model-assisted estimates, not observed job losses. Many capabilities depend on specially arranged workplaces.

Before replacing a role, an employer must establish whether the machine can perform the required work reliably in the actual setting, what supervision and recovery will cost, and who will handle exceptions. A demonstration answers only part of that decision. Anthropic systems also assist this book’s external critique workflow; see the transparency statement.

A labor market can therefore be transformed before unemployment statistics announce a technological revolution.

Reuters reported in September 2026 that Meta had developed an internal plan known as Project OT, for Organization Transformation, that modeled scenarios in which some teams would shrink by as much as 60% across two waves as the company moved toward what it called an AI-native operating model. The first wave proceeded in May and cut roughly 10% of the workforce. Hours before it began, chief executive Mark Zuckerberg abandoned the second company-wide wave. Meta confirmed the project, said the 60% figure applied only to scenarios for certain teams, and said it had not moved forward with every scenario.

According to internal posts reviewed by Reuters, on which Meta declined to comment, major technical and security incidents rose 40% from the prior year and the time employees spent responding to them rose 70%. Reuters also reported that AI-assisted code changes were up 220% year-on-year while product improvements reaching users increased much more slowly. Those figures do not establish that AI caused the incidents or that the strategy cannot work. They do show why quantity of delegated activity is not the same thing as quality of delegation.

California has written a boundary against algorithmic abdication into employment law. Signed on September 30, 2026, SB 947—the “No Robo Bosses Act”—will prohibit employers from relying solely on an automated decision system to discipline or dismiss an employee. Beginning July 1, 2027, an employer relying primarily on the system must have a human corroborate the decision. If its output cannot be corroborated, or the reviewer finds it inaccurate, incomplete, or misleading, the employer cannot use it for that decision. The book’s distinction now has a concrete legal expression: analysis may be delegated; responsibility for judgment remains with the employer.

On October 7, 2026, government, employer and trade-union representatives signed the Astana tripartite declaration on digital labor and artificial intelligence. The agreed principles call for workers to receive information about AI decisions affecting their employment and working conditions, with an opportunity for human review. The parties also agreed to establish a national platform, working groups and monitoring arrangements. This is a declaration of shared principles; its adoption does not establish effective enforcement. The practical question remains whether workers can reach someone with the authority to reconsider a decision.

At this stage, the most defensible conclusion is that firms are learning how much work can be detached from employment, and the answer is increasing. The labor-market consequences depend on what organizations do with the capacity they gain: expand output, reduce hiring, cut headcount, improve quality, or spend the savings on new forms of work.

What Employment Still Supplies

The more work machines can perform, the easier it becomes to underestimate everything employment supplies besides output.

A workplace is a training system. Senior employees teach junior ones. Colleagues notice mistakes. People absorb standards that were never written down. Informal conversations reveal problems before they reach a dashboard. Workers develop judgment because they experience consequences over time.

A workplace is also a social system. People build trust. They understand which customer is frightened rather than angry, which executive is asking the wrong question, which colleague needs help, and which procedure should be ignored because the situation has changed.

An AI system can increasingly imitate parts of those functions. It can store institutional knowledge, monitor patterns, coach workers, and communicate fluently. An organization should still distinguish imitation of a social function from ownership of a social role.

The human employee is situated inside relationships of authority, duty, law, and consequence. The agent is situated inside a technical architecture created by people.

This brings the economic argument back to the Doctrine. Human involvement is not human accountability. The accountable human must have authority over the decision itself. When a company replaces human tasks with agents, accountability does not migrate into the software. It moves to the people who authorize the workflow, define the boundaries, choose the monitoring system, and decide what happens when the machine encounters something outside its competence.

The manager of an automated workforce therefore has a different job from the manager of a human workforce. The manager governs authority rather than motivating the agent or protecting its career.

Accountability Analysis

The title of this chapter contains a trap. Calling an AI system an employee makes the technology easier to understand, but it can also smuggle in assumptions that do not belong.

An employee can be assigned responsibility because the employee is a person embedded in an institution. An AI agent can be assigned work and operational discretion, but performing more of the work does not make it the accountable human.

The first accountability question remains: who answers? The accountable owner must sit where authority over the real workflow sits.

If a customer service agent denies a request, the answer cannot be the model provider merely because it built the underlying system, unless the provider actually owns that decision under the relevant arrangement. It cannot be a remote person who helped train the model. It cannot be an annotator who once labeled similar data. It cannot be a generic reviewer whose only connection to the outcome is that a human touched the technology somewhere upstream.

In a company, that may be the executive who approved the deployment, the operational leader who owns the process, a licensed professional responsible for the specific decision, or several identifiable roles whose duties form a traceable chain. The chain may be distributed. It may not disappear into the phrase AI employee.

The practical test is simple. If the system makes an important mistake tomorrow, who can change the result, who can change the rule, who can suspend the system, and who must explain why the system had that authority? If nobody with those powers can be named, the organization did not hire a tireless employee. It created unowned authority.

Accountability Verdict

Artificial intelligence is changing employment before it eliminates employment.

It is separating productive capacity from the human relationship that once supplied most cognitive work. That separation can make companies more productive, make small teams more powerful, extend service across time and language, and remove tedious work from human jobs. It can also compress hiring, concentrate responsibility in fewer people, hide new categories of work, and tempt organizations to mistake continuous machine activity for useful capacity.

The economic opportunity is real, and so is the conceptual error.

An AI agent can supply labor without employment. It cannot supply accountability without humans.

Count the whole job

For one task you plan to hand to AI, count the time spent preparing instructions, checking results, fixing mistakes, and handling unusual cases. Compare that with the time saved. Also ask whether the result is good enough and whether someone can correct it. A shorter software bill or a faster first draft cannot answer those questions alone.

The next chapter examines the group most exposed to that distinction: the people who traditionally entered a profession by doing the very tasks AI is learning to perform first.

Sources and Notes

1. Klarna Group plc, Annual Report on Form 20-F for the year ended December 31, 2025, filed with the U.S. Securities and Exchange Commission in 2026. https://www.sec.gov/Archives/edgar/data/2003292/000200329226000007/klar-20251231.htm. Accessed September 9, 2026. Klarna reports that its AI assistant handled 80% of customer service chats in 2025, had completed 31 million conversations since launch, operated 24/7 in 23 markets and more than 35 languages, and was estimated to perform work equivalent to more than 700 full-time agents. The filing reports $39 million in 2024 cost savings from the assistant. The figures are company reported. Source record: DA-CH5-S1. Archive reference: pending.

2. Kathryn Bonney, Cory Breaux, Emin Dinlersoz, Lucia Foster, John Haltiwanger, and Aditya Pande, “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks,” NBER Working Paper 35141, 2026, revised July 9, 2026; companion Census Bureau Working Paper CES-WP-26-25. https://www.nber.org/papers/w35141 and https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-25.html. Accessed September 9, 2026. The nationally representative data show 18% firm adoption, 32% employment-weighted adoption, 66% augmentation-only use among adopters, and employment reductions in 2% of AI-using firms. Source record: DA-CH5-S2. Archive reference: pending.

3. Rossana Merola, Ekkehard Ernst, Daniel Samaan, Maria del Rio Chanona, and Ole Teutloff, International Labour Organization, “The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence,” June 1, 2026. https://www.ilo.org/publications/impact-genai-jobs-productivity-and-work-organization-review-empirical. Accessed September 9, 2026. The review finds real but uneven productivity gains, limited large-scale job displacement to date, and meaningful changes in work organization and job quality; it also notes that reported time savings have not yet translated consistently into higher measured output, earnings, or employment. Source record: DA-CH5-S3. Archive reference: pending.

4. Katie Paul, Reuters, “Mark Zuckerberg had a bold plan to replace Meta staff with AI, and it imploded. Answering readers’ questions on Reddit,” September 3, 2026. https://www.reuters.com/technology/artificial-intelligence/mark-zuckerberg-had-bold-plan-replace-meta-staff-with-ai-it-imploded-answering-2026-09-03/. Accessed September 9, 2026. Reuters reports Project OT from internal documents, recordings, and interviews: the first wave reduced Meta’s workforce by roughly 10%; the second company-wide wave was abandoned; 60% referred to scenarios for certain teams; Meta said it did not pursue every scenario. Internal posts reviewed by Reuters reported technical and security incidents up 40%, response time up 70%, and AI-assisted code changes up 220% year on year; Meta declined to comment on those internal incident figures. Source record: DA-CH5-S4. Archive reference: pending.

5. George L. Taylor, author production note on metered AI workflow friction, September 20-21, 2026. The note distinguishes total project cost from repeated small-block credit purchases during a long AI-assisted publishing workflow. The passage treats the experience as a design and accountability example, not as an audited pricing study or a claim about every AI service. Source record: DA-CH5-S5. Archive reference: pending.

6. Charles Daly, Bloomberg, “Klarna Slows AI-Driven Job Cuts With Call for Real People,” May 8, 2025. https://www.bloomberg.com/news/articles/2025-05-08/klarna-turns-from-ai-to-real-person-customer-service. Accessed September 9, 2026. Bloomberg reports Sebastian Siemiatkowski’s view that cost had become too predominant in customer service and Klarna’s renewed hiring of human agents so customers could reach a real person. Source record: DA-CH5-S6. Archive reference: pending.

7. Klarna Group plc, Third Quarter 2025 Results presentation filed with the U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/2003292/000162828025052805/exhibit992klarnapresenta.htm. Accessed September 9, 2026. The presentation reports 28 million annual conversations, 81% of customer service chats handled by AI, work equivalent to 853 full-time agents, and $58 million in annualized cost savings. These are company reported estimates. Source record: DA-CH5-S7. Archive reference: pending.

8. Anthropic, “What work can robots do?”, September 30, 2026. https://www.anthropic.com/research/what-work-can-robots-do. Accessed October 1, 2026. Claude-assisted occupational task and cost assessments estimate technical exposure covering 34% of U.S. working time under some conditions and cost competitiveness for 0.3%. These are model-assisted estimates, not observed job losses; many capabilities require structured workplaces. Anthropic systems assist this book’s external critique workflow. Source record: DA-CH5-S8. Archive reference: pending.

9. California Legislature, SB 947, Chapter 859, Statutes of 2026, approved and filed September 30, 2026; Labor Code Sections 1522(b)–(c) and 1526.7. https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB947. Chaptered legislative text verified through https://legiscan.com/CA/text/SB947/2025 because the official bill-text page was inaccessible. Senator Jerry McNerney, “Newsom Signs McNerney’s No Robo Bosses Act of 2026 Requiring Human Oversight of AI in the Workplace,” September 30, 2026. https://sd05.senate.ca.gov/news/newsom-signs-mcnerneys-no-robo-bosses-act-2026-requiring-human-oversight-ai-workplace. Accessed October 6, 2026. The provisions become operative July 1, 2027. Corroboration may use data collected or used to produce the system’s output or other relevant supporting information; independent evidence is not required by the chaptered text. Enactment does not establish effectiveness in practice; statutory exceptions remain. The connection to retained judgment is the author’s analysis. Source record: DA-CH5-S9. Update 071. Archive reference: pending.

10. Federation of Trade Unions of the Republic of Kazakhstan, “First digital forum of the federation: social partners signed the Astana tripartite declaration on digital labor and AI” (Russian-language title translated), October 7, 2026: https://www.kasipodaq.kz/ru/news/779/. Ministry of Labour and Social Protection of the Population of Kazakhstan, “The state, employers and trade unions agreed principles for using AI in employment” (Russian-language title translated): https://www.gov.kz/org/enbek/press/news/details/1307019?lang=ru. Accessed October 12, 2026. Official participant accounts confirm signing by government, employer and union representatives, worker information and human-review principles, and plans for a national platform, working groups and monitoring. The ministry states that decisions affecting rights and interests should involve a human. These are accounts of a declaration and implementation plans, not independent evidence of compliance, effective remedies or a newly enforceable statutory obligation. Employer-accountability comments in the federation account are attributed to its chairman rather than presented as verbatim provisions of the complete declaration. The practical question about authority to reconsider is the author’s analysis. Source record: DA-CH5-S10. Update 091. Archive reference: pending.