The Digital Employee Is Already at Work. Now We're Trying to Figure Out Who Pays for It — and Who Makes Money From It

2026-08-26 7614 words 36 minutes

AI agents are becoming a new form of labor. This essay examines who pays for digital work, who captures the savings, and why the economics remain unsettled.

Contents

AI agents are already doing work that, until recently, was done by people. Some companies are cutting staff, others are simply no longer opening new roles, and some are discovering that digital labor can be surprisingly expensive. Meanwhile, the people who keep their jobs are increasingly becoming operators and supervisors of machines. I tried to understand what is actually happening with the economics of AI agents by the end of summer 2026 — and where this story could go over the next nine months.

Everyone Is Waiting for Something

Late summer 2026 is a pretty nervous time for artificial intelligence. Investors are looking at the tens and hundreds of billions of dollars flowing into data centers, chips, power plants, and new models, and they are increasingly asking a question that would almost have been considered bad manners a year ago: when exactly is all of this going to pay off? Skeptics are waiting for the AI bubble to burst and for the next economic hangover to begin. Companies building models and infrastructure have to prove that their enormous investments will eventually turn into more than impressive demos — they need to turn into profit. Companies buying AI have to prove roughly the same thing to their own CFOs.

People are nervous for a different reason. They are waiting for layoffs. Some have already seen them. But the most interesting part of this story may be happening much more quietly: some workers are not being fired because they are never being hired in the first place. A role that three years ago would have appeared automatically as workload grew now has to pass through a new mental filter first: do we really need another person, or can we give this extra work to AI?

Then there is a third group — the people who bet on implementation. Executives, engineers, operations teams. They buy agents, connect them to systems, give them access to data, rebuild workflows, and quickly discover that autonomous AI differs from a nice-looking chatbot in roughly the same way a car at an auto show differs from a taxi that drives around a city twenty hours a day. While the car is sitting under spotlights, everything looks great. Once it starts working, you get fuel costs, maintenance, accidents, insurance, and eventually a person who has to explain all of it.

And finally, there are companies that prefer to wait. The strategy sounds reasonable: let Uber, Salesforce, banks, and tech giants make all the possible mistakes first and figure out security, integrations, cost, and control. Then everyone else can buy mature technology and take the shortcut.

But I am increasingly doubtful that this shortcut exists. Uber can learn how to use coding agents inside Uber. Santander can build them around its own banking infrastructure. A Japanese insurer can build them around its own sales processes. Other companies’ experience obviously helps, but the next company will still have to connect an agent to its own data, permissions, people, legacy systems, and those strange internal processes that have officially existed since 2014 even though nobody remembers why. Companies that are waiting may not be avoiding the learning curve. They may simply be delaying the moment when they will have to climb it anyway.

Against this background, I came across a Forbes headline: “KPMG Says Nearly Half Of Executives Pulled Back AI Agents Over Cost.” Nearly half of executives, it said, had pulled back on AI agents because of cost.

It sounded serious. And suspiciously simple.

So I started digging into what KPMG had actually found. And pretty quickly, I realized this story was not so much about expensive agents as it was about the arrival of a new kind of labor whose economics almost nobody really knows how to calculate yet.

Those 49 Percent

Let’s start with a small correction to the dramatic headline. KPMG did not find that half of businesses had given up on agents and decided to go back to the old way of doing things. In its survey of more than two thousand executives across twenty countries, 24% said they had narrowed or reduced the scope of deployment, while another 25% had delayed or paused further rollout after costs started looking worse relative to expected value. Those two numbers produced the famous 49%. Overall interest in AI and investment plans had not disappeared.

So this is not an escape. It is more like the moment companies started realizing that you cannot just rush this technology into production, and that a failed attempt can get very expensive.

At this point, the question “do AI agents work?” is starting to feel a little outdated. Yes, they work. They write code, handle correspondence, process requests, research documents, help sell products, work inside banking processes, and increasingly can do more than answer a question — they can carry out a sequence of actions without a person guiding every step. The debate about technical capability is slowly giving way to a more boring, and therefore more important, question: does the economics of this new kind of labor actually work?

To answer that, comparing token costs with an employee’s salary turned out to be nowhere near enough. You have to look at several sides at once. The person who gets fired or never gets hired. The person who stays and now has to be responsible for the machines. The company buying digital labor instead of human labor. And the company selling that digital labor.

Everyone in this story has a different balance sheet. And, more interestingly, the very same agent can be good news for one side and very bad news for another.

The Worker Who Wasn’t Fired

The easiest version of automation to turn into a headline is familiar. A company deploys AI, and then several hundred or several thousand people receive emails about “restructuring,” “efficiency,” and “resource reallocation.” That part of the story is already real. Salesforce has linked reduced demand for some customer-support work to AI. IBM has publicly talked about automating some HR functions. In 2026, there have also been larger corporate cuts where executives explicitly named new AI capabilities as one factor behind the decision. But macroeconomic research still does not show a world in which millions of white-collar workers are disappearing from the labor market all at once.

The more interesting channel works differently. The company fires nobody. It simply does not hire the next person.

Shopify turned this new logic into something close to official policy. Before asking for more headcount, a manager is expected to explain why the work cannot be done with AI. Salesforce has connected gains in engineering productivity with less need to expand its engineering workforce significantly. TCS has talked about slower future hiring as digital agents take on a larger share of additional work. In some companies, what disappears is not an existing job but a future line in the budget.

That could matter much more than it seems. A layoff shows up in statistics, press releases, government notices, and LinkedIn posts. A vacancy that was never opened leaves almost no trace. Nobody writes an obituary for the person a company planned to hire next fall and then decided it no longer needed.

That is why today’s labor-market picture can be misleading. If a company grows 30% while headcount rises only 5%, we see employment growth. Formally, everything looks fine. But under the old production model, the same business growth might have required the team to grow 20%. The gap between those two numbers is made up of people who never appeared on the payroll.

Recent research increasingly points to headcount compression as one of the most realistic scenarios for the current stage. Not a huge wave of layoffs happening all at once, but a gradual weakening of the link between business growth and employee growth.

Entry-level workers may feel this first and most painfully.

An analysis by the Stanford Digital Economy Lab found a noticeable relative decline in employment among young workers in occupations more exposed to generative AI. In the updated version of the research, by August 2026 the gap for workers aged 22–25 in the most AI-exposed occupations had reached roughly 19% relative to less exposed occupations, while more experienced workers looked more resilient. The authors also stress that this is a descriptive relationship, not a clean estimate of AI’s causal effect.

Even with that caveat, the mechanism is too logical to ignore. Entry-level work usually contains a lot of relatively structured tasks: initial research, data processing, first drafts, routine code, document checks, CRM updates, simple reports. This is exactly the layer of work that generative AI has learned to absorb fastest.

AI was supposed to take the hard jobs first. Instead, it learned office work surprisingly fast.

And here we get a fairly nasty paradox.

AI can make a junior employee much more productive. In one well-known study of customer-support workers, average productivity rose by around 15%, with the biggest gains going to less experienced employees. The machine was effectively passing some of the best workers’ knowledge to newer colleagues. It is one of the strongest arguments for AI as a tool that can democratize professional expertise.

But at the same time, companies are starting to hire fewer junior workers.

The result is almost comical: the technology makes a young employee much more effective, and then the company discovers that it needs far fewer young employees.

What is even more interesting is that expectations for the few people who do get hired are changing. Simple tasks are increasingly treated as “AI work,” so junior employees are expected to show mature judgment, client skills, independence, and the ability to check a machine much earlier in their careers. Labor-market research is already talking about the “seniorization” of entry-level roles: the vacancy may still be labeled junior, but the bar keeps moving upward.

In a sense, everyone got promoted. Except the people who did not get a job because of the promotion.

There is another hidden bill here that companies are barely paying yet. Professional judgment does not appear because somebody gives you a corporate login and adds AI to your job title. It develops over years of relatively simple work, mistakes, feedback, and watching more experienced colleagues.

A junior developer learns on small tasks before designing architecture. A young lawyer does research and document review before being trusted with a complex opinion. An analyst builds spreadsheets and simple models before becoming the person who can immediately spot that an entire presentation rests on a bad assumption.

If agents take over that first layer of a career, the company really does save money today. But a few years later, an uncomfortable question may appear: where do you get the senior people who are supposed to supervise those same agents?

Automating the ladder turned out to be easier than figuring out how people will climb without it.

The Person Who Stayed

Let’s say your job did not disappear. Better yet, the company bought you several digital assistants. One collects data, another does research, a third writes, a fourth analyzes documents, a fifth codes. In a presentation, this looks almost liberating: finally, the machine will take the routine work and the human can focus on what the human brain was made for — strategy, creativity, and preferably something high-margin.

In a real company, there is one more step between “the agent did the work” and “the work is done.”

Someone has to look at what it actually did.

In one banking-app modernization case described by McKinsey, developers worked with 15–20 specialized agents per person. Some agents explored the legacy codebase, others designed the target architecture, others migrated code and ran tests. Humans did not disappear. Their work moved upward: planning, coordination, review, testing, and decision-making.

From the outside, this looks like an extraordinary productivity boost. One developer gets a small digital team.

One person can now manage a digital team. That doesn’t necessarily mean less work.

But machine output and human attention do not scale at the same speed. If twenty agents can produce work at the same time, they can also produce work that needs to be checked at the same time. Sometimes it is wrong. Sometimes it is partly right, which can be worse: obvious nonsense is easy to throw away; a convincing mistake has to be noticed first.

This creates a hidden category of work that companies still measure badly: the supervision tax. The employee spends less time producing the output directly and more time giving machines context, assigning tasks, checking results, fixing mistakes, restarting failed processes, and handling exceptions.

Imagine a normal corporate calculation. An analyst used to need eight hours to prepare a report. Now an agent creates a draft in eight minutes. The presentation shows almost eight hours saved. But if a senior analyst then spends forty minutes checking sources, correcting calculations, and trying to understand why section three contains a completely confident number that exists nowhere in the source material, those forty minutes often disappear between the lines.

AI looks almost free if human distrust is not included in the cost.

And the higher the cost of an error, the more expensive supervision becomes. Reviewing five ad drafts is relatively easy. Reviewing five credit recommendations or five decisions inside a regulated process is a completely different job. That is why the question “how many agents can one person supervise?” is almost meaningless by itself. Five agents writing first drafts of marketing copy and five agents capable of making irreversible financial actions are arithmetically the same number, but they create completely different jobs.

There is another unpleasant possibility here: AI can reduce the amount of human work while increasing its average difficulty.

If an AI support agent handles the 80 easiest cases out of 100, the human gets the remaining twenty. Great. Except now the entire day consists of twenty customers the machine could not help: unusual situations, conflicts, exceptions, errors, anger, money that went missing somewhere. Complex cases used to be mixed with simple ones. Now the simple ones are gone.

There are fewer tasks. The density of unpleasantness goes up.

This matters especially for managers. AI does not necessarily reduce management work; it can increase the number of people and processes one manager is responsible for. We are increasingly moving toward a model where one person manages not only people but groups of agents — and is still responsible for whatever comes out the other end.

Microsoft calls this person an agent boss. The phrase sounds a little like a job title a child might invent after watching a cartoon about the future. But the production model behind it is quite serious: one person controls a much larger volume of work because more and more of the direct execution is being handed to machines.

The only question is whether that means the person works less.

So far, there is not much evidence that it does.

Who Owns the Hour AI Saved?

The Bank of Korea produced one of the best numbers for understanding the current stage of AI. Researchers found that using generative tools really does reduce time spent on work by about 3.8% — roughly an hour and a half in a standard 40-hour week.

It is a great example of how easy it is to mix up three completely different things: task speed, worker productivity, and organizational productivity.

If I write an email in five minutes instead of ten, the task became twice as fast. That does not mean my working day became five minutes shorter. And it certainly does not mean the company made more money.

So where did the saved hour go?

Maybe into waiting for a manager’s decision. Into the next meeting. Into another project. Into ten more emails. Into checking an agent’s work. Sometimes just into a small break, which, to be fair, is not the worst possible use of new technology either.

But corporate logic usually sees that hour differently. If technology frees up capacity, there is a strong temptation to fill it.

There is something here that looks a lot like the Jevons paradox. When producing something becomes cheaper, society often starts consuming more of it. AI made presentations cheaper — so we get more presentations. It made research and analysis cheaper — so we get more research. It made writing code easier — so we get more code. The machine removed some work, and the organization responded by increasing the amount of work it can now produce.

That may be good news for the economy and less good news for an employee’s calendar.

So the more interesting question is not “how much time does AI save?” but who owns the time it saved?

If the worker owns it, they can work less. If the company owns it, the worker can handle more clients, projects, and processes. If the savings lead to lower headcount, some of that time simply turns into work redistributed among the people who remain.

I have found very few major corporate cases where the headline result of AI adoption is: “our employees now work five hours less for the same salary.”

There are plenty of examples of “we freed up tens of thousands of hours so employees can spend more time with customers.”

Technology gives a person an hour back. The employer is already standing there with the next task.

Uber and the Problem of AI Being Too Useful

Uber is a particularly good example of how strange the economics of digital labor can become.

The company rolled out AI coding tools widely across thousands of developers. Usage grew at enormous speed. From an adoption point of view, this is close to the perfect result. Nobody had to force employees to use the new tool. They wanted to use it.

And that became the problem.

The heaviest users could spend hundreds or even thousands of dollars a month on AI coding. A budget intended to cover roughly a year of AI developer tools was used up in about the first four months. The company started introducing limits and learning how to manage consumption.

With traditional software, this kind of success would look like a dream. You buy a product, employees use it every day, adoption is nearly universal. The more they use it, the easier it is to justify the license.

AI adds an awkward detail: every time an employee enthusiastically uses the product, somewhere a meter starts running again.

But even a huge Uber bill tells us nothing about whether the decision was bad. If a developer using an agent creates twice as much genuinely useful product, an extra $1,500 a month per employee could be one of the best investments the company ever made.

The problem appears at the next level. Uber could see rising usage, more code, and faster work. But executives publicly acknowledged that drawing an equally clean line from extra code to genuinely useful features for customers, and then from those features to extra revenue, is much harder.

This is one of the central conflicts in today’s AI economy. Producing more is not the same as creating more value. More code does not necessarily mean a better product. More leads do not necessarily mean more sales. More documents do not necessarily mean better decisions. And “hours saved” definitely do not equal money saved until the company does something economically useful with those hours.

The Digital Employee and Its Corporate Credit Card

In a presentation, replacing a human with an agent can look almost indecently attractive. A person needs salary, taxes, benefits, equipment, vacation, sick leave. They sleep, sometimes quit, and for some mysterious reason believe they should earn more after a few years.

An agent asks for none of that.

It does, however, come with a kind of corporate credit card.

One of the most interesting academic papers of 2026, How Do AI Agents Spend Your Money?, shows how quickly costs get strange once a model stops answering a single question and starts carrying out a long task on its own.

Researchers tested eight leading models on hundreds of real software-engineering tasks. In the experimental setup, complex multi-step tasks performed autonomously by agents consumed roughly a thousand times more tokens than a normal one-prompt, one-response interaction with AI. This does not mean that “every AI agent is a thousand times more expensive than ChatGPT.” It refers to a specific type of complex programming task where an agent had to perform many actions in sequence. But the mechanism matters far beyond software development.

The agent reads data, calls a tool, gets a result, adds it to context, thinks again, reads something else, tries an action, gets an error, fixes it, runs a test, and continues. As it moves forward, it drags a growing history of previous steps behind it. The context grows like a snowball, except someone pays for every new layer.

Digital workers don’t need salaries or vacations. But they still send a bill.

The spread is even more interesting. In the same study, repeated runs of the exact same task could differ in token use by as much as 30 times in extreme cases. The models were also poor at predicting their own future cost.

Imagine giving an employee the same task twice. On Monday, they send you a bill for $10. On Thursday, it is $300. When you ask what it will cost next time, they answer with great enthusiasm and very little correlation with reality.

The CFO starts to get a little nervous.

That is why the real cost of a digital employee is not the token price. You also have model calls, external tools, compute infrastructure, corporate data retrieval, permissions, monitoring, security, auditing, integration, retries, error correction, and the human supervision we just discussed.

Sometimes there is one more cost: the person you have to bring back after deciding a little too optimistically that you no longer needed them.

Klarna and the Price of a Human Conversation

Klarna became one of the biggest symbols of the early wave of replacing people with AI. The company said its customer-service assistant could handle a workload comparable to around 700 full-time employees. The story instantly became proof that much of customer-support work was almost over.

Then the picture became more complicated.

Klarna’s leadership acknowledged that focusing too aggressively on cutting costs had hurt service quality, and the company began strengthening human support again for cases where customers really needed a person.

It is easy to turn this into the opposite myth: “AI failed, humans won.” That would be just as wrong.

AI at Klarna still handles a huge amount of work. The company did not go back to 2022 and switch automation off. It found a boundary: a standard support request and a human relationship with a customer are not the same thing.

This is probably what mature AI-agent systems will look like in many companies. The machine takes everything that is repetitive enough, measurable enough, and safe enough. Humans remain for exceptions, conflicts, emotionally difficult situations, and decisions where the cost of an error rises sharply.

In that model, the human does not disappear. The human becomes the expensive final support layer.

And again, this brings us back to job complexity. The more successfully AI filters out simple cases, the higher the concentration of difficult ones for the remaining team. The company saves human hours, but every human hour that remains may become more expensive and more exhausting.

But Do Agents Actually Make Money?

After KPMG, Uber, rising bills, and Klarna, it would be very easy to write an article arguing that corporate AI has turned out to be another expensive illusion. That is exactly why I went looking specifically for evidence pointing the other way: not “how many hours did AI save?” but where are agents actually generating additional revenue or other measurable economic value?

Those cases do exist.

Santander reported around €35 million in economic value from AI in the first quarter of 2026 and set much larger targets for the following periods. The bank is also using hundreds of autonomous agents across different processes. But precision matters here: €35 million is not published, audited net profit from agents. Santander calls it business value, which includes different kinds of savings, efficiency gains, and financial impact.

Uber for Business used Salesforce Agentforce to automate work with incoming leads and reported roughly a 60% increase in conversion. The company expects an additional revenue impact in the seven figures. This is much closer to the kind of evidence I was looking for: the agent is operating close to the point where a sale happens. But the public case still does not show the system’s full cost, so “conversion increased” cannot automatically be translated into “the agent generated this much net profit.”

Walmart reported that customers using its AI assistant Sparky build baskets that are around 35% larger. The number looks impressive, but it is an observed correlation: it is entirely possible that people willing to actively use an assistant while shopping are already more motivated buyers.

And this is where I caught myself thinking that I had been asking a slightly wrong question all along. What matters is not only how smart an agent is or how much it costs. What may matter even more is where exactly the company puts it — and how close it sits to the place where real money is created for the business.

An agent that saves an employee twenty minutes while preparing an internal memo may be very cheap and create almost no financial result. An agent that is expensive but increases the chance of selling a high-value product may have fantastic economics.

So the question “are agents more expensive or cheaper than people?” is starting to sound about as useful as asking, “is an employee who costs $200,000 a year expensive?”

Expensive for what?

If they bring the company a million dollars, no.

If they spend the entire year rearranging columns in one report, maybe.

The Most Important Number Almost Nobody Shows

The longer I worked on this research, the more one thing about corporate AI reporting started to annoy me. Companies are very good at talking about token counts, the share of employees using a new system, the amount of code produced, the number of cases resolved autonomously, hours saved, and productivity gains.

But they very rarely show the full path to money.

How much did the system cost after integration? How much human time was needed for review? What did errors cost? How much additional revenue appeared specifically because of the agent? How much of that revenue remained as gross profit? And what was left after all additional costs?

Uber has an impressive story about adoption and code output, but there is no clean public line from AI spending to profit and loss. Santander has business value, but that is a management metric. Walmart has bigger baskets, but causality remains a question. The Bank of Korea has time saved, but not a corresponding increase in overall output.

That does not mean AI is not paying off. It means something more uncomfortable: the market is already spending enormous amounts on a new form of labor, but it is still much better at measuring its activity than its economic result.

And this is not only a problem for buyers.

Because there is another side that already seems to have become pretty good at monetizing what is happening.

The Company Selling You an Employee

Traditional software had a fairly simple role. It sold tools to people. A CRM helped a salesperson sell. A support system helped an agent answer customers. A development environment helped a programmer write code.

AI-agent software is starting to reach for a different part of the budget.

Not just the software budget.

The payroll budget.

Salesforce Agentforce has already become a large and fast-growing source of recurring revenue. Sierra sells agents for customer service and experiments with outcome-based pricing, where the customer does not pay for a seat in an interface or even for tokens, but for completed work. This is a much deeper change in the business model than adding another feature to a CRM.

A software company used to tell an employer: “buy our tool for your ten employees.”

Now it can say: “maybe you do not need ten employees.”

If this logic really scales, AI-agent vendors gain access to a market much larger than the traditional software-subscription market. They are no longer selling only the means to produce work. They are selling the ability to perform the work itself.

Labor as a service, basically.

But vendors have their own problem. Traditional SaaS loved active users. The cost of a customer clicking a button in your application another hundred times was usually tiny. With AI-agent software, every extra action has a real compute cost.

A successful user can become an expensive user.

That is why the market is frantically experimenting with pricing: by tokens, actions, completed tasks, outcomes, credit packages, and different hybrid models. If the customer pays for usage, much of the risk of unpredictable costs stays with the customer. If the customer pays for an outcome, the risk moves to the vendor: the agent can make five extra attempts, call an expensive model, fail, try again, and the customer still wants to pay only for the solved problem.

And this creates a funny question that classic SaaS almost never had to ask:

Who should pay for a machine’s thinking when all that thinking leads nowhere?

That is why huge recurring revenue at companies selling AI agents does not automatically mean equally huge profits. We can already see fairly well how much money this new market is generating. It is much harder to know how much vendors actually keep after paying for compute and model usage.

So even the most obvious winners have not yet revealed the full economics of their new business.

Where Did the Old Salary Go?

Now imagine an extremely simplified situation. A company used to spend $100,000 a year on an employee. After automation, a large share of the same work can be done for $20,000 in additional technology costs.

On paper, $80,000 just appeared.

But that money does not vanish, and it does not necessarily become pure employer profit. Some goes to the company selling the AI-agent platform. Some goes to the developer of the underlying model. Further down the chain sit cloud and compute infrastructure. Integration, monitoring, and security often add more. Maybe the company now needs a more expensive senior specialist who supervises several digital workers. Whatever remains after all of that can indeed become additional profit.

The person whose work was automated is the only participant in this new chain who is not automatically entitled to a share.

That is why the debate about AI productivity is slowly turning into a debate about who gets the AI dividend.

If technology makes society more productive, who gets the gain? Workers through higher wages and more free time? Consumers through lower prices? Companies through profit? Shareholders through higher valuations? AI vendors through usage fees?

Economically, it is perfectly possible for AI to create enormous real value while the position of a large share of workers gets worse at the same time. There is no mathematical contradiction here. GDP and your salary have never been the same metric.

This becomes especially clear when you look at other countries.

One Digital Worker, Two Opposite Problems

The American AI debate quickly turns into “who is going to lose their job?” In Japan, the same agent enters an economy with the opposite problem: there simply are not enough workers.

Japanese companies increasingly discuss AI through 人手不足 — chronic labor shortages. In that environment, a digital employee may not be a way to remove a person from payroll. It may be a way to get work done when the company cannot find a person at all.

Meiji Yasuda Life is investing heavily in generative AI deployment and in developing its own AI-skilled workforce, while the Japanese market more broadly increasingly sees AI as one way to compensate for labor shortages.

Same technology.

In California, using it may mean: “we will not hire another five people.”

In Japan: “thank God we do not have to find another five people.”

This is an important counterweight to attempts to build one universal moral story around AI. The social outcome of automation depends not only on what the model can do. It depends on demographics, labor markets, the cost of human time, institutions, and how badly the economy needs more workers in the first place.

South Korea offers a third version. Individual AI adoption is very high, but corporate restructuring is moving more slowly. Banks are starting to deploy agents in document workflows and lending processes, freeing tens of thousands of human work hours. But those hours are generally expected to go not toward shorter working weeks, but toward more customer-facing and commercial work.

In Spain, research on the financial sector points to another limit of automation: customers are willing to accept AI while everything is going smoothly, but when something goes wrong, the desire to speak to a human rises sharply. In one recent survey, 60% of Spanish customers said they would not sign up for a banking product without human involvement, and when a problem occurs, 85% would prefer to deal with a person. Here, the human employee is not an expensive leftover from the old model. The human is part of the product — a carrier of trust.

You can automate a transaction.

Trust is much harder to automate.

Who Is Winning Right Now?

By the end of August 2026, the picture is clear enough to reject both extremes. There is no mass destruction of human employment by agents yet. But it is also impossible to say that all of this is just another wave of hype with no real effect on work.

For people directly displaced from specific roles by AI, the result is obvious. They lost income and their place in the organization.

For people who were never hired, the effect is even less visible but potentially larger. Headcount compression, extra barriers before opening new roles, and a higher bar for entry-level jobs may become the main labor-market effect of the next stage.

For the people who remain, the picture is mixed. They get tools that can sharply increase individual productivity, but they also get more responsibility, more supervision, and more difficult exceptions. Their value may rise, but I do not yet see any guarantee that their salary or free time rises proportionally with it.

Managers are becoming a particularly interesting group. Some management layers may disappear, but the managers who remain get much more responsibility: more people, more processes, and gradually more digital workers. The manager of the future may not be someone who manages ten people, but someone who manages five people and fifty machines — and is still accountable when one of those machines does something stupid.

For companies buying agents, results vary dramatically. Some get real savings and the ability to grow without proportional payroll growth. Others see more activity but cannot connect it to profit. Others discover an unexpected AI bill or have to bring back human work they declared unnecessary too early.

The companies selling agents look like the most obvious winners in terms of revenue. But even for them, profitability remains an open question because digital labor has a real variable cost.

So the phrase “AI wins, people lose” explains almost nothing. What is actually happening is a much more complicated redistribution between different kinds of workers, employers, software vendors, model owners, and capital.

And all of this is only beginning.

Maybe We’re Looking for the AI Bubble in the Wrong Place

Let’s go back to the main fear of late summer 2026.

When people talk about an AI bubble, they usually imagine the classic scenario. Enormous investment, inflated valuations, promises that cannot be delivered, then a sudden reality check and capital disappearing.

But corporate AI may hit reality in a much less cinematic way.

Companies will simply switch off bad use cases. Limit autonomy where it does not create enough value. Use cheap models for simple tasks and expensive frontier models only where the extra intelligence actually matters. Put spending limits in place. Stop loops where an agent keeps repeating the same pointless actions. Measure the cost of a successfully completed task rather than the cost of a token. Put ordinary predictable software back in places where a model is not needed at all. And keep humans where the cost of an error is higher than the savings.

This is already happening.

Uber did not turn off AI after discovering its budget problem. It started managing consumption. KPMG did not see companies abandoning agents en masse. It saw deployment plans being revised and costs being treated much more seriously.

Maybe there really is a bubble somewhere in company valuations and expectations around infrastructure spending. But at the level of corporate adoption, what is happening looks less like a collapse and more like growing up.

Companies have simply started doing the math.

And sometimes that is exactly the sign that a technology has stopped being a toy.

What Could Change by Summer 2027?

We do not need to predict artificial superintelligence to get a serious change in the labor market. A much more boring scenario is enough.

Imagine that over the next nine months today’s agents simply become 30% better economically. Models get a little cheaper. Errors become less frequent. Context handling improves. Companies get better at routing tasks between cheap and expensive models and limiting unnecessary calls. Human supervision becomes a little better integrated into workflows.

No conscious superintelligence.

Just today’s digital employee, a little cheaper, a little more reliable, and a little easier to manage.

That alone may be enough to change the decision about hiring the next person in many companies.

So I would not bet on a scenario where white-collar workers around the world are suddenly hit by mass layoffs in the first half of 2027. The current evidence gives us too little reason to predict that.

Continued headcount compression looks much more likely.

The company grows, but its human workforce grows much more slowly. New workload is first distributed between the existing team and AI. When someone leaves, the company increasingly decides not to hire a replacement. Several processes are grouped around one stronger specialist and a set of agents. Direct layoffs appear mainly where the economics of that model have already been proven well enough.

The most visible social effect may show up among young workers. If a company can hire one junior employee instead of two and give that person AI, there are fewer openings and the expectations for every remaining candidate go up. A few years later, that logic may hurt the company itself: you cannot keep importing experienced specialists from the future if you stopped developing people today.

At the same time, a new class of “digital labor operators” will grow. Not necessarily people with a job title like AI Manager. More likely ordinary developers, analysts, salespeople, marketers, and managers who gradually discover that a large part of their job is no longer producing the output directly, but organizing and checking the output of machines.

The main limit of this kind of organization may turn out to be something surprisingly old-fashioned.

Human attention.

Compute can be bought.

One person’s ability to remain responsible for the decisions of a hundred autonomous processes scales much less easily.

The Smartest Agent Won’t Necessarily Win

The first stage of generative AI was a competition between models. Who reasons better, who has the larger context window, who wins the next benchmark.

The corporate market needs a slightly different winner.

Not the smartest agent in the world, but an agent that is good enough and has understandable economics.

A CFO does not necessarily care that a model scored two points higher on a test of abstract reasoning. It is much more useful to know that a specific operation costs $1.20, finishes without human help 96% of the time, needs review another 3% of the time, gets escalated to a specialist 1% of the time, and saves the company $430 per thousand operations after all costs.

At that point, the agent stops being magic.

It becomes a line in the budget.

And that is when the really interesting phase begins.

Because today we are still mostly asking what AI can do. Which profession it can replace, how much code it can write, how long it can work autonomously.

But the economic question is slowly becoming a different one.

Who gets what it saves?

If a machine frees up an hour for a person, that hour can go to the person, the company, the customer, or the next task.

If a company avoids hiring another employee, the money can become profit, go to the AI vendor, the model developer, the data center, or the salary of a more expensive specialist supervising the machines.

If one worker can now serve twice as many customers, that worker may become twice as valuable. That does not mean their salary automatically doubles.

If a junior employee is no longer needed, the company saves money today. But a few years later, it will still need an experienced specialist — and somebody had to teach that person how to be a junior first.

Technology may solve the problem of producing work long before society solves the problem of distributing the value it creates.

And that part of the problem is no longer technological.

The Digital Employee Is Already Here

That is probably the most important thing I took away from this whole story.

We are no longer talking about a hypothetical world where AI will someday enter the workforce.

It already has.

The first stage just looks much less dramatic than either optimists or pessimists imagined.

Millions of people are not all walking out of offices carrying cardboard boxes at the same time. Somewhere, a company does not open another junior role. Somewhere, a manager gets a few more processes. Somewhere, a developer works with a dozen agents. Somewhere, a support employee is left with only the hardest calls. Somewhere, a CFO sees a token-usage report for the first time. And somewhere, Salesforce or Sierra turns somebody else’s savings on human labor into its own revenue.

So far, we have got neither the promised world where AI takes the routine work and humans do nothing but creative tasks and go home early, nor the mass unemployment predicted by the darkest scenarios of the last few years.

Something less spectacular, and possibly more important, is happening.

The price of human labor is changing relative to digital labor.

Today, the digital employee is still fairly expensive, unpredictable, occasionally wrong, and in need of supervision. That leaves humans a huge amount of economic space.

But I am not sure that advantage is permanent.

So by the middle of 2027, I will be more interested in boring numbers than in the next model benchmark or even the number of new AI models. I will be watching how many companies stop replacing employees who leave, how many entry-level jobs disappear, how many processes each remaining employee is responsible for, how much human supervision costs, and what share of AI projects can finally show real profit after all expenses.

Because the most important question of the next stage may not be “will AI replace humans?”

It is much more down to earth.

What happens if a worker who needs no salary, vacation, or sleep becomes another third cheaper — and companies finally learn how to calculate its real cost?

Then we may find out whether 2026 was the beginning of another technology bubble.

Or the first year a new kind of labor quietly appeared in corporate accounting.

Sources and Research

Below are the main primary studies, corporate materials, and journalistic sources used in this article.