AI does not need to fail for the AI bubble to burst.
The internet did not disappear when the dot-com bubble collapsed. What disappeared was the belief that every company with a domain name had discovered a new economy. The technology was real. Much of the market built around it was not.
AI is heading toward a similar reckoning.
The technology is useful, improving, and likely permanent. The financial story wrapped around it is something else. The market has not priced AI as another useful class of software. It has priced AI as a new supply of labor: cheaper than people, faster than people, available around the clock, and scalable without the usual limits of hiring.
Some companies acted on that promise before proving it inside their own operations. They cut roles after seeing AI perform the most visible parts of the work, then discovered that producing an answer was not the same as owning the outcome.
Now some of those roles are coming back.
Rehiring alone does not prove that AI is a bubble. It matters because it weakens the economic assumption beneath the boom: that companies can replace human labor with software that is cheaper, reliable enough to do the work, and scalable.
It is one of the clearest visible points where the AI sales pitch has met business reality. The costs, returns, and infrastructure surrounding that assumption are now facing the same test.
A job was never just the visible task
AI is exceptionally good at making work appear smaller than it is.
It can generate a customer response in seconds, produce a report from a stack of documents, or create hundreds of lines of code before a developer could finish describing the wider system. The output arrives quickly, and the speed is easy to measure.
Everything surrounding that output is harder to see.
A job also contains context, exceptions, judgment, relationships, verification, and responsibility for what happens next. Those parts rarely look impressive in a demonstration, but they are usually what separates a plausible answer from work a business can safely rely on.
Customer support makes this difference impossible to ignore.
A customer does not contact a business because they want text generated at them. They want their problem understood and resolved. Sometimes that requires account history. Sometimes it requires an exception to a normal process. Sometimes it requires noticing that the customer is frustrated for a reason the original question does not fully explain.
At Aliph Marketing Studio, AI will never be the customer support representative. We may use it elsewhere in our work, but never between us and a customer. The person responding must understand the situation, have the authority to make a decision, and remain responsible for the result. That is not inefficiency waiting to be automated. It is part of the service.
The corporate rush toward AI support is not producing the obvious victory its advocates expected. A 2026 Gartner survey of 3,566 customers found that customers were roughly three times more likely to use third-party AI tools than company-provided chatbots when trying to resolve a service issue. Use of company chatbots had remained statistically unchanged since 2022, and customers still expected the option to reach a human.
The spending is harder to defend. In a separate survey cited in the same report, service and support leaders had invested a median of 12% of their 2025 budgets in AI, the highest share among the ten business functions Gartner assessed. Only 24% demonstrated positive financial returns from their AI use cases.
This does not mean customers reject every use of AI in service. It means replacing accessible human support with standalone automation has not broadly produced the adoption or financial returns companies expected.
For that reason, I consider customer support the worst place to begin removing people. A weak internal summary can be corrected before it leaves the company. A dismissive, fabricated, or context-blind support response becomes the customer’s experience of the company.
This is the category error underneath many AI layoffs. Companies saw a machine perform the visible task and assumed the job had been automated. The context did not disappear. The exceptions did not disappear. Responsibility did not disappear.
Only the person carrying them did.
The boomerang is no longer hypothetical
There is not yet enough evidence to declare a market-wide reversal of AI layoffs. There is enough to say the failure pattern has moved beyond isolated anecdotes.
The clearest evidence of actual rehiring comes from Robert Half. In a survey of 2,000 US hiring managers reported by Fast Company, 32% said their organizations had eliminated a role because of expected gains from AI or automation and later rehired for that same role.
The reasons are more revealing than the headline number. Forty percent cited institutional knowledge or context that AI could not replace. Thirty-eight percent said the technology required more human oversight than expected. Thirty-five percent said the promised productivity gains were smaller than expected.
Those are not complaints about a missing feature that the next model release will necessarily fix. They describe a basic misunderstanding of what the role contained.
Gartner and Forrester expect the pattern to spread. Gartner predicts that by 2027, half of the companies that attributed customer-service headcount reductions to AI will rehire people to perform similar functions, often under different job titles.
Forrester forecasts that more than half of layoffs attributed to AI will be quietly reversed. The firm also warns that companies are presenting financially motivated cuts as AI transformation before they have mature systems capable of doing the work.
The boomerang also does not guarantee justice for the person who originally lost the job. Companies may restore the function under another title, hire someone cheaper, or divide the work across an already strained team. The business recovers capacity. The worker who paid for the experiment may never be invited back.
That makes the reversal more damning, not less. A company can recover from a bad automation decision. The person whose income was treated as an acceptable trial cost has to absorb the damage personally.
Then the cheap replacement got a meter
The case for replacing people with AI depends on more than capability. It depends on price.
The salary-versus-subscription comparison looks irresistible only while a flat monthly fee hides the cost of actual consumption.
I wrote about this dependency in AI Was Fast. Cheap Was the Part That Broke. Since then, the pricing model has become even more explicit.
On June 1, 2026, GitHub moved Copilot from request-based allowances to token-based AI Credits. In its announcement of the change, GitHub said agentic use had brought substantially higher compute and inference demands, making the previous model unsustainable.
Agentic work consumes AI differently from autocomplete or a quick chat response. An agent may load a repository, run tools, fail, and retry. Every step consumes resources. The meter keeps running through the failures.
Gartner now predicts that AI coding costs could exceed the average developer's salary by 2028, citing rising token consumption, opaque billing, and autonomous workflows whose costs can outrun measurable productivity.
The providers are under pressure as well. OpenAI reportedly burned $3.7 billion in the first quarter of 2026 while generating $5.7 billion in revenue. Serving and developing these systems can consume extraordinary amounts of capital.
Agentic AI does not inherit the near-zero marginal cost associated with conventional software distribution. Its cost grows with usage, context, model choice, and failed attempts.
A replacement that still needs human review was never a clean replacement. If its operating cost also rises with every serious task, it may not remain the cheaper one either.
The spending already assumes the ending
The losses and rehiring might look like normal friction in an emerging market if the investment around AI were equally cautious. It is not.
Alphabet, Microsoft, Meta, and Amazon are now expected to spend more than $700 billion on AI in 2026. Analysts cited by Reuters put broader AI-related capital spending at roughly $800 billion this year, with one 2027 forecast reaching $1.12 trillion.
That is not experimental spending. It assumes demand will absorb an extraordinary buildout of data centers, chips, power, networking, and software, and that enough profitable use cases will emerge to justify it.
Major technological shifts need infrastructure, but that does not put every investment beyond criticism. The bubble appears when spending outruns proven demand and continues because nobody wants to fall behind if the most optimistic forecast is right.
That pressure gives layoffs a convenient story. Ordinary cost cutting sounds more strategic when announced as AI transformation. Payroll falls immediately, while the promised payoff is pushed into the future and management gets credit before the system proves it can carry the work.
The dot-com boom followed the same logic. Telecom companies built far more fiber than the market could support in the late 1990s. The internet later made enormous use of that infrastructure, but its eventual value did not save the companies, investors, or workers caught in the correction.
There is an important difference. Today's hyperscalers are highly profitable, less leveraged, and financially stronger than many telecom companies caught in the dot-com collapse. They can finance the buildout for much longer, even if returns disappoint. That makes an immediate collapse less likely, but it does not protect them from overcapacity, weak returns, shorter asset lives, or billions spent ahead of durable demand.
AI infrastructure can remain useful for decades and still be materially overbuilt at today's prices. A valuable asset does not make every investment in that asset sensible.
The company-wide gains are still missing
The strongest AI demonstrations happen at the task level. The financial promises are made at the company level. The gap between them is where a lot of the claimed productivity disappears.
A 2026 NBER working paper based on almost 6,000 executives across the United States, United Kingdom, Germany, and Australia found that around 70% of firms actively use AI. Yet more than 80% reported no impact on either employment or productivity over the previous three years.
The executives still expected AI to raise productivity by 1.4% and reduce employment by 0.7% over the next three years. In other words, adoption is already widespread, the realized company-level effect is usually absent, and the transformation remains safely located in the next planning horizon.
I see the task-level speed every day in software work. AI can write code much faster than I can type it. Typing was never the difficult part. The difficult part is understanding the system, deciding what belongs in it, catching a plausible but wrong assumption, testing the result, and owning what happens after deployment.
A tool can make each visible task faster while the project becomes harder to understand. It can produce more code while creating more review. It can increase activity without increasing finished value. I have written about this before as the illusion of progress in modern web development: movement is real, but movement and delivered value are not the same thing.
This is why a company can buy AI licenses for everyone, generate more material than ever, and still struggle to point to a corresponding improvement in profit, output, or customer experience. The machine completed tasks. The organization did not necessarily become more productive.
The bubble can burst without AI disappearing
“Bubble” describes the market built around a technology, not whether the technology works. Treating those as opposites protects the hype because any working product can then be offered as proof that the market around it must also be rational.
A bubble is a mismatch between real utility and the price paid for its imagined future. AI has plenty of utility. It also has providers burning billions, customers failing to demonstrate returns, employers rehiring people they thought the technology could replace, and a capital buildout already measured in hundreds of billions of dollars per year.
What breaks first may not be the technology. It may be the business models built on subsidized usage, the valuations that require extraordinary growth, the automation plans that removed people before removing the work, and the projects approved mainly because executives were afraid to look late.
A correction does not need to arrive as one clean event. It can look like tighter usage limits, higher prices, canceled data centers, written-down investments, failed AI vendors, and teams quietly restoring roles under new names. The boomerang layoffs matter because they show reality entering the calculation from the bottom up, one damaged workflow at a time.
AI will keep growing after the bubble bursts. The products and business models around it will still have to survive on actual business value, not fear of being left behind or the promise of free digital labor. The rehiring shows the correction starting inside real companies, with workers absorbing the cost when those assumptions fail.
AI may become as important as the internet, but that will not make the wreckage created along the way any less real.
Let’s Plan Your Next Project
If you’re building something fast, modern, and beautifully crafted — let’s talk. I help founders and teams turn ideas into polished, high-performance products.