We have all seen companies look harder and harder into AI for the work that used to belong to people. The pitch sounds irresistible: these systems run all day, they do not call in sick, they do not ask for raises, and they do not complain when you dump more work on them. As a human myself (in case you didn't know, haha), those are all definitely flaws my colleagues and I have compared to the AI gunning for our jobs eventually. On a spreadsheet, that can look like the cleanest headcount decision in the world. In practice though, the savings are not what was expected, the repair work is coming in larger, and the real bill is hitting a lot of folks later than expected.

The layoff numbers by themselves are already a warning sign. Challenger, Gray & Christmas reported that employers announced 49,135 AI-related job cuts in the first four months of 2026, and AI accounted for 26 percent of all announced layoffs in April alone. What makes the story more revealing is that the financial payoff still is not showing up the way many executives thought it would. MIT's NANDA research found that 95 percent of enterprise generative AI pilots delivered no measurable profit and loss impact, which is about as far from a slam dunk as y'all can get.

That gap matters because the cost of replacing people with AI is not just the subscription fee or the API line item somebody puts in a deck. It is electricity, the cooling, the GPU access, the cloud capacity, the data storage, and the security review. It is the integration with old internal systems that were never built for this. It is the human oversight layer that still has to catch the mistakes when the model sounds polished but gets the substance wrong. A lot of companies are discovering that they did not replace labor so much as trade a visible salary expense for a messier pile of infrastructure and supervision costs.

The electricity piece alone deserves a lot more attention than it gets. A California policy report by CalMatters warned that large data centers tied to AI growth may require major new power infrastructure and that if regulators are not careful, ordinary ratepayers can end up subsidizing those costs. The same report argued that data centers should pay the full costs they impose on the grid, including prepaying for new infrastructure and paying for the power capacity they reserve. That is a big clue about where this is going. The AI bill is not staying inside the AI budget. It is spreading outward into utilities, permitting, land, transformers, cooling systems, and local power planning.

Then there is the chip and compute problem. The whole market keeps acting like AI can scale infinitely as long as management wants it badly enough, but computing is not abstract. It runs on expensive hardware, limited supply, and energy hungry facilities. Latest research studies show that many firms are badly underestimating AI total cost of ownership, from enterprise setup and licensing to retraining, observability, model retuning, and reversal costs when the rollout disappoints. One of the most useful ways to frame this is simple: every company chasing "AI savings" is also joining the same fight for GPU capacity, server space, engineering talent, and reliable electricity. None of those are cheap, and none of them are infinite.

Even when the tools work more or less as intended, there is another cost that barely shows up in the executive story: cleanup. Harvard Business Review highlighted what researchers called "workslop," meaning AI-generated work that looks acceptable at first glance but does not meaningfully move the task forward. In the research, desk workers spent an average of 1 hour and 56 minutes fixing each incident, with an estimated cost of $186 per employee per month, or about $9 million a year for a 10,000-person company. The model gave them something fast, but now a human has to check it, correct it, rewrite it, de-risk it, and sometimes apologize for it.

That is why the replacement case keeps looking weaker once it hits real workflows. Customer service is a good example. On a dashboard, the resolution rate may appear acceptable while customer satisfaction declines in the background. This cites a 2026 Qualtrics report finding that AI-powered customer service fails at four times the rate of other AI use cases and that one in five consumers said it provided no benefit at all. In plain English, folks do not judge support by whether a bot answered. They judge it by whether the problem got solved, whether the answer made sense, and whether they could get to a real person when things got messy.

Software and knowledge work are hitting a similar wall. AI can be genuinely useful for drafting, summarizing, searching, and speeding up repetitive work. It can absolutely help a good employee move faster. But that is not the same thing as replacing judgment. The project research notes that employment for younger software developers dropped sharply after the big generative AI push, while multiple benchmarks still show that AI agents fail a large share of ordinary office tasks and code-related work without supervision. So companies are cutting the entry level of the ladder before the tools are reliable enough to stand on their own.

That may look efficient for a quarter or two, but it creates a long-term talent problem and pushes more review work onto the people who remain.

And this is the part that keeps getting skipped: layoffs are not free either. Severance costs money. The companies that go through buyer's remorse and end up rehiring cost money. Training those replacements costs money. Burned customer trust costs money. Internal confusion costs money. Many leaders in several industries who made AI workforce moves regretted the move, and a significant share had already been rehired after realizing the technology could not carry as much of the workload as expected. That is the kind of quiet reversal that rarely gets the same attention as the original AI announcement, but it tells you a lot about what is really happening behind the scenes.

The core issue is straightforward. These models do not replace judgment, context, accountability, or ownership. They generate the next likely token. Sometimes that is incredibly useful. But the second the situation gets expensive, sensitive, emotional, ambiguous, or high-stakes, somebody still has to own the answer. The machine does not care if the customer feels a certain kind of way, or if the legal language is wrong or doesn't have the right tone. It also doesn't care if the analysis is misleading or if the recommendation causes damage. A person, on the other hand, still carries that weight.

That does not make AI useless; let's get that straight. You can use these tools every day for research, drafting, comparing options, pressure-testing ideas, and cleaning up ugly first passes. They are useful, fast, and can make a strong employee stronger. But I still treat them like a very bright and very confident intern who occasionally makes something up, says it smoothly, and hopes nobody checks too closely. We should all be checking though. Because the human is the one who has to live with the outcome.

That is why the companies that will win here probably are not the ones trying hardest to replace people. They are the ones using AI to make good people more effective while staying honest about the total cost of the system around it. It also means admitting that a tool can be impressive and still be the wrong substitute for a person.

The window to learn how to use these systems well is still open. But the lesson coming out of 2026 so far is not that AI employees are here to take over. It is that a lot of companies tried to skip straight to replacement before they understood the real operating costs, the quality tradeoffs, and the amount of human judgment the work still requires. At the end of the day, humans still hold the edge where it actually matters: judgment, accountability, context, restraint, and the messy ability to show up when something real is on the line. AI is an incredible tool. It is just not a teammate yet.

— Eduardo Cestaro