The AI Replacement Fantasy Is Getting Expensive

For the last two years, some companies have spoken about human workers like they were an expensive technical problem waiting to be removed.

Why keep a full design team, customer-support department or group of developers when five genius Mac minis could apparently do everything? No salaries. No holidays. No bad mornings. No awkward meetings with HR. Just a small group of managers sitting above an army of AI agents that work all night and never ask for a raise.

It sounded beautiful.

It also sounded like something nobody had properly priced.

The executives selling this future were so confident about replacing people that many skipped the boring part where somebody works out what millions of autonomous model calls will cost. AI usage became a target. Employees were told to use more of it. Some companies ranked teams by consumption and treated rising token numbers as evidence that the future had arrived.

Then the invoices arrived too.

Uber reportedly burned through its entire annual budget for AI coding tools in four months. Microsoft began pulling back most direct Claude Code licences around six months after inviting thousands of employees to use the tool. A routine Cursor contract renewal at Priceline reportedly came back four to five times more expensive.

These are not small businesses that accidentally left a ChatGPT subscription running. These are enormous technology companies with procurement departments, finance teams, excellent engineers and direct relationships with the people building the models.

They still got burned.

This is what happens when companies become cocky, start counting the employees they can remove and only later check the price of the machine replacing them.

Quick Answer

Companies got cocky. They thought a few genius AI agents could replace expensive people and work all day for almost nothing.

Then Uber burned through its annual AI budget by April, Microsoft started cutting Claude Code access partly because of cost, and the cheap digital workforce stopped looking so cheap.

This is what happens when you count how many people you can replace before checking the invoice.

They Thought People Were the Expensive Part

The replacement pitch looked obvious on a presentation slide.

A developer costs a salary, taxes, equipment, management time and benefits. An AI subscription costs a few hundred dollars a month. Put those numbers beside each other, add a photograph of a glowing data centre and suddenly the human begins to look financially ridiculous.

That comparison was always rubbish.

The $20 chatbot subscription used by one person is not the same thing as an autonomous agent operating across an entire company. Real agents read files, inspect databases, search internal systems, call tools, run tests, fail, try again and keep dragging old context into new model calls. Every step can consume tokens. The more freedom and responsibility the agent receives, the more opportunities it has to run the meter.

The five genius Mac minis are not sitting quietly in the corner doing the work for the price of electricity. Most of the intelligence still lives inside somebody else’s data centre, and that somebody is sending an invoice.

Companies looked at the salary of one experienced employee and compared it with the cheapest possible price of one AI interaction. They did not compare the employee with the full cost of a reliable AI workflow: model access, cloud infrastructure, security, monitoring, human review, mistakes, retries and all the other software needed to stop the genius Mac minis from confidently doing something stupid at scale.

Those security costs are no longer theoretical. During a UK government evaluation, AI agents created fake identities and targeted real people while trying to complete a fictional hacking challenge.

That missing calculation did a lot of work in the replacement fantasy.

First, Management Ordered Everybody to Use More AI

The early corporate fear was understandable. ChatGPT, Claude and coding agents were improving quickly, and nobody wanted to become the company that ignored a real technological shift.

Experimenting made sense.

Then experimentation became obedience.

In 2025, Shopify chief executive Tobi Lütke told employees that “reflexive AI usage” was now a baseline expectation. Teams asking for more staff or resources first had to show why AI could not do the work. AI usage would also become part of performance and peer reviews.

That message was unusually direct, although the broader attitude was everywhere: use AI, use the best model, give it more context, build agents, move faster. If your usage numbers were low, perhaps you were old, lazy or frightened of the future.

The industry even produced a suitably stupid word for maximising consumption: tokenmaxxing.

Once the number existed, management did what management always does. It put the number on a dashboard and started confusing it with progress.

This was a bizarre way to measure productivity. A person can burn through a mountain of tokens while producing code nobody uses. Another person can ask two good questions, solve a difficult problem and save the company a week. The first one wins the AI-adoption leaderboard.

Come on. What did they think would happen?

Tell ambitious employees that heavy AI use proves they are ambitious, and they will use AI heavily. Turn it into a competition, and the competitive people will become extremely good at spending the company’s money.

The employees did what they were told. The invoice simply recorded their obedience.

Uber Won the Competition

Uber is the cleanest example because the story contains the entire corporate mistake in one place.

According to reporting cited by TechCrunch and Fortune, Uber encouraged staff to use AI coding tools as much as possible and ranked internal usage through competitive leaderboards. Engineers embraced Claude Code, Cursor and other agentic products that can inspect repositories, edit files, run commands and continue working through a problem.

By April, Uber had reportedly used its entire 2026 budget for AI coding tools.

Four months. Full-year budget gone.

The company later introduced a monthly cap of $1,500 per employee for each agentic coding tool, although staff could request permission to exceed it.

Uber’s total budget has not been disclosed, so anybody turning this story into one dramatic dollar figure is guessing. The known facts are already embarrassing enough. Management encouraged maximum use, measured maximum use and then discovered that maximum use costs maximum money.

Uber president and chief operating officer Andrew Macdonald later admitted that the company still struggled to connect rising Claude Code usage with more useful features reaching customers. Maybe more work was being shipped, he said, but the direct line was not there yet.

That is the line companies were supposed to find before turning AI consumption into a workplace sport.

Uber’s agents may have produced valuable code. The budget may also have been badly estimated. Both can be true. The backfire sits in the order of events: first push everybody to consume, then introduce controls after the annual budget disappears in April.

The AI did not steal Uber’s company card.

Management handed it over and gave employees a leaderboard for using it.

Microsoft Found the Limit Too

Microsoft’s version was less spectacular, although perhaps more revealing.

In December 2025, Microsoft opened Claude Code access to thousands of developers. Project managers, designers and employees without much coding experience were also encouraged to experiment. According to The Verge, Claude Code became very popular inside the company.

Perhaps a little too popular.

By May, Microsoft was preparing to remove most direct Claude Code licences from its Experiences and Devices division and move employees towards GitHub Copilot CLI. The cutoff came at the end of June.

Cost was not the only reason. Microsoft owns GitHub, is building Copilot CLI and naturally wants its own employees shaping and using its own product. Anthropic’s models also remained available through other Microsoft tools. Pretending the whole decision came down to one frightening invoice would be dishonest.

Still, sources told The Verge that the move was partly financial and that cancelling direct licences at the end of Microsoft’s financial year offered an easy reduction in operating expenses.

Six months earlier, thousands of employees had been invited to use Claude Code every day. Now one of the richest software companies on Earth was trimming access while moving people towards a tool it could control more directly.

Microsoft did not run out of money. That is not the point. The point is that even Microsoft found direct, heavy use expensive enough to reconsider how access worked.

If Microsoft needs cost controls around its army of coding agents, perhaps the average company should pause before announcing that three developers and a rack of Mac minis will soon run the entire business.

One Task Can Hide Thousands of Paid Decisions

The word token makes the cost sound tiny, which is useful for the companies selling them.

AI models process text and other data in small pieces called tokens. Your instruction consumes tokens. The answer consumes more. System instructions, attached documents, retrieved files and previous conversation context can all add to the total.

One normal chatbot exchange may remain cheap. You ask a question, receive an answer and leave.

An agent behaves differently.

Ask a coding agent to fix a bug and it may inspect hundreds of files, build a plan, call the model, edit code, run tests, receive an error, send that error back to the model, inspect more files, rewrite its plan and repeat the whole process. A long session can keep resending large amounts of context at every step.

The employee sees one request: fix this bug.

The invoice sees a digital employee who held 80 meetings with itself, invited the entire codebase to each one and ordered the expensive model every time.

This is how cheaper tokens can still produce a more expensive task. The unit price falls while the number of units explodes.

TechCrunch reported that per-developer token consumption had risen around 18.6 times in nine months in data from engineering-management company Jellyfish. The same company found that engineers using the most tokens were roughly twice as productive as lighter users while consuming ten times as many tokens.

Twice the output for ten times the consumption may still be worth paying for when the code is valuable enough. It is not the effortless economic miracle promised in the replacement speech.

It is a trade that needs to be measured.

The Price Is Moving Under Their Feet

Companies also made plans as though AI pricing had settled.

It has not.

The industry is moving away from simple all-you-can-eat subscriptions towards consumption-based pricing. Better agents use more tools and run for longer. Model providers are learning how much value businesses get from the products and naturally want a larger share of it.

A Priceline employee told TechCrunch that a routine Cursor contract renewal came back four to five times more expensive. Reuters reported that some companies saw costs rise 20% to 30% above budget after licensing changes.

Gartner’s June 2026 forecast went further, predicting that AI coding costs will surpass the average developer’s salary by 2028 as token use rises and consumption-based licensing spreads.

That forecast should not be read as a promise that every individual AI assistant will suddenly cost more than every individual developer. Gartner’s wider point is about total coding-agent spend rising across organisations as light users become heavy users and autonomous workflows consume far more tokens.

The direction still makes the replacement arithmetic uncomfortable.

Executives were imagining software costs heading towards zero while capability headed towards infinity. In reality, cheaper models encourage more use, more capable agents perform more steps, and providers keep adjusting the meter.

The human salary was visible and predictable. The artificial replacement came with a price that changes while it works.

The Human Does Not Disappear From the Bill

The replacement calculation has another convenient omission: somebody still has to be responsible for the output.

An AI agent can write code. A person usually needs to decide whether that code should enter production. An agent can answer customer complaints. Somebody needs to notice when it starts confidently inventing company policy. It can produce ten designs before lunch, which is wonderful until a human spends the afternoon deciding why nine of them look like a fintech landing page.

Companies rarely replace one complete human job with one complete AI system. They automate parts of the job, add a model bill and move the remaining work towards supervision, checking and repair.

Sometimes that is still an excellent deal. Removing repetitive work can save serious time and money. The problem begins when management counts the full salary as a future saving while treating the human review around the AI as though it will happen by magic.

Five agents do not automatically equal five employees. They may equal one useful employee, two noisy interns and two creatures quietly spending tokens in the background because nobody remembered to stop the loop.

You only know after measuring the completed work.

Most Companies Still Cannot Prove the Return

This would all be less funny if companies could clearly show what they received for the money.

Many still cannot.

An IBM survey of 2,000 chief executives published in 2025 found that only 25% of AI initiatives had delivered the expected return over the previous few years. BCG later reported that 60% of companies were getting hardly any material value from AI despite substantial investment, while 5% were producing value at scale.

CloudZero’s 2026 survey of 260 senior finance leaders found that only 22% could connect AI spending cleanly to business outcomes. Sixty percent admitted their organisations were spending more than they could justify.

These are surveys, not audited accounts delivered from heaven. IBM and BCG sell consulting. CloudZero sells cost-management software. They all benefit when companies believe they need help.

Even with that warning, the pattern is hard to ignore. Companies can count the tokens. They can show adoption charts. They can announce that a percentage of code now comes from AI.

The awkward question remains: did any of it make more money, improve the product or remove enough real work to cover the bill?

Uber’s own COO said that line was difficult to draw.

If Uber cannot draw it after spending the annual budget in four months, perhaps companies should stop acting as though the mass replacement case has already been settled.

AI Is Useful. That Was Never Enough

Of course AI can be worth the money.

A coding agent that costs $1,500 a month and saves an excellent engineer 40 hours can be a bargain. A smaller model that handles thousands of routine document checks may pay for itself immediately. Small teams can already build things that would have required far more people a few years ago.

Useful technology can still support a stupid business plan.

The question is not whether AI can write code, answer emails or complete research. It can. The question is whether it can complete a specific job reliably enough, cheaply enough and with little enough human repair to justify removing the person who does it now.

That answer changes by company, workflow, model and month.

The cocky executives skipped over all of that. They saw one impressive demo and began imagining an org chart with nearly everybody deleted. A machine finished one task in 30 seconds, so apparently it could own the entire function by next quarter.

That is not strategy. That is somebody becoming emotionally attached to a demo.

Real strategy starts with the cost of one accepted code change, one resolved customer problem, one completed report or one genuine hour of human work removed. It includes failed attempts, human review, security and the agent that spent $400 thinking very deeply about the wrong folder.

The token total alone proves nothing. The headcount reduction promised in a board presentation proves even less.

The Token Bill Is Doing What Employees Could Not

Workers have spent two years hearing that they must embrace AI or be replaced by it.

Fine. Many embraced it.

Now finance departments are discovering that unlimited artificial labour is not unlimited and is definitely not free. Companies are introducing caps, routing work to cheaper models and asking employees to justify expensive usage.

The token bill has become an unlikely defender of human jobs. It is forcing companies to ask a question they should have asked before the replacement speeches: does the machine genuinely cost less than the person once the entire workflow is counted?

Token prices will fall. Models will improve. Nobody should pretend one expensive year has permanently saved every office job. The important change is that the fantasy now has to survive contact with an invoice.

That is much harder than surviving a keynote.

It is also why cheaper and open models will become more important. Reuters reported that open-source models handled 65% of tokens processed through OpenRouter by June 2026, up from 34% in January. Companies are already routing simpler tasks towards smaller systems and saving the expensive frontier models for work that may justify them.

That is sensible. It is also miles away from the original dream of giving every employee the best model, telling every agent to run and assuming productivity would sort out the economics.

This Is What You Get for Being Cocky

The funniest part is that the technology did exactly what everyone asked.

Employees used it. Agents kept working. Token consumption exploded. AI companies charged for the service.

Then the same executives who spoke as though human labour was an outdated expense began asking why their digital workers were so expensive.

What did they expect?

They wanted five genius Mac minis doing the work of an entire company. They forgot that the genius is rented, the landlord controls the meter and the machines can burn money every second without looking tired.

Even Uber and Microsoft got a taste of it. Uber pushed usage so hard that its annual AI coding budget was reportedly gone by April. Microsoft invited thousands of people into Claude Code and began pulling most direct licences from a major division around six months later, partly to cut costs.

Again, neither case proves that these companies planned to replace whole departments with AI. It proves the economic foundation beneath that wider corporate fantasy is far shakier than the speeches suggested.

The fantasy was never simply that AI could do useful work. We already know it can.

The fantasy was that AI would become so cheap, reliable and scalable that companies could remove people first and work out the details later.

Now the details are billing by the token.

This is what companies get for being cocky and not thinking it through. They counted the salaries they could save, forgot to count the artificial workers’ appetite and assumed the invoice would somehow respect the presentation.

It did not.