The Real Cost of AI

The date is July 24th, 2026. A Friday.

Sarah (yes, her again), VP of Operations at a mid-size fintech in Frankfurt, adjusts her presentation remote and takes a slow breath before the final slide. The numbers are good. Better than projected, actually. The AI customer service platform went live in March. By May, forty agents had been replaced. Projected annual savings: 3.2 million euros. Customer complaints are up slightly, but the Q4 report does not include customer satisfaction scores. Nobody asked for them.

The board applauds. Sarah gets a raise. However, there is no free lunch. The invoice always comes due and, in that case, the bill was already in the mail.

We recently launched an episode on Data Voyagers Podcast about FinOps and the cost of data. This is not an article about that. What I mean by cost is something much bigger than money, it involves jobs, business, the environment, our relationships, mental health, and of course, the future.

This is also not an article about whether AI is good or bad. That debate is roughly as useful as arguing about whether electricity was good or bad in 1895. What I want to do here is something more uncomfortable: open the full invoice. We can delay it, but we will have to pay it, sooner or later. Some people are already paying it.

The cost of AI is real and measurable. It does not fit neatly in a financial model because it is distributed across more stakeholders than any single financial model tracks. It lands on companies, workers, families, governments, the environment, and the political order. It does not arrive all at once either. It comes in installments, often addressed to someone who was never in the room when the original decision was made.

To read it properly, you need to read it across time. Short term, mid term, long term. The costs that are easiest to see right now are not the most expensive ones.

Part one: Short-term costs (0 to 2 years)

The savings slide gets the applause; the invoice arrives by mail slot.

The efficiency math nobody checks

In 2023, Klarna, the Swedish buy-now-pay-later company, announced that its AI assistant was handling the work of 700 customer service agents. The CEO was effusive. The press was enthusiastic. The savings were real. By 2025, Sebastian Siemiatkowski publicly admitted the transition had gone too far. Service quality on complex interactions had deteriorated. The projected savings were lower than expected. The company started rehiring into a hybrid model: AI for routine queries, humans for everything that actually requires judgment.

Salesforce went through a similar arc. The same applies to a long list of companies whose reversals received considerably less coverage than the original announcements. The pattern is not isolated. It is a trend.

None of this is surprising if you read the economics carefully. Daron Acemoglu, professor at MIT and one of the sharpest thinkers working on this problem, published a paper in 2024 called "The Simple Macroeconomics of AI." His finding was that only 20% of U.S. labor tasks are exposed to AI at all, and of those, only 23% can be profitably automated, with average labor cost savings per task around 27%. The resulting GDP gain projected over the next ten years sits between 1.1 and 1.6 percent. Not per year. Total.

That is the upside. The downside does not appear in the same model.

What typically gets missed decision-makers: integration expenses, quality monitoring, error correction, customer escalation management, and the organizational drag of running a half-human, half-automated system that was never quite designed to be either. Companies that cut hourly workers first, because those cost lines are the easiest to model, often discover too late that "easy to model" and "safe to cut" are not the same thing.

The worker who took the first hit

Automation rarely starts at the top of the org chart. It starts where the hourly rate is the easiest number to put in a spreadsheet.

Customer service. Data entry. Call center operations. Warehouse processing. Support Agents. The data on what is happening to these roles is not subtle. Research from SSRN's 2025 AI Job Displacement Analysis estimates an 80% automation risk for customer service representatives. Data entry: 7.5 million jobs projected eliminated by 2027. Roughly 40% of jobs globally show meaningful exposure to AI capabilities. In high-income countries, that number rises to 60%.

The detail buried in the positive projections of the SSRN Analysis is this: 77% of new AI-adjacent roles require a master's degree. The people being displaced are not, by and large, the people being hired to replace what they did. The ladder is being pulled up faster than anyone is climbing it.

We cannot train the people being displaced to fill the new jobs being created.

This is not inevitable. It is a choice: the compounded result of thousands of individual corporate decisions made without full accounting for the costs being transferred to workers, families, and governments downstream.

The environmental down payment

Every prompt has a receipt. Most people just never see it. And I am not talking about the Token financial cost.

AI systems generated between 32.6 and 79.7 million tons of CO2 in 2025, according to research published in the journal Patterns by Cell Press. Global data centers consumed 415 terawatt-hours of electricity in 2024. The water footprint of AI systems in 2025 sat somewhere between 312.5 and 764.6 billion liters. Google's single data center in Council Bluffs, Iowa consumed 1.3 billion gallons of potable water in 2024 alone.

That last number deserves a pause. It is so important that I will break it down. One facility. One year. 1.3 billion gallons of drinking water. By many estimates, that is enough for over 40,000 people in a whole year. In a county with a population of roughly 63,000 people. These are people who now need to pay more for something they had it for almost nothing. No new job created for the community. A long list of problems broght.

These are not projections. They are last year's numbers. And the demand is accelerating.

The people paying this cost are not the companies running the data centers. Quite the opposite, it is common for them to receive tax benefits, instead of paying the bills. The actual people taking the cost are the communities near the cooling towers, the agricultural regions competing for aquifer access, and the countries whose grids are straining to keep pace. The cost is real. It is just distributed to people who were not in the room when the infrastructure decisions were made.

Part two: Mid-term costs (2 to 5 years)

The installments come due, as a storm.

The talent you cannot buy back

BT Group announced plans to cut between 40,000 and 55,000 jobs by the end of the decade. CEO Allison Kirkby, who took over in 2024, suggested the cuts could go even deeper depending on what the company learns from AI deployment. IBM has been trimming thousands from HR, accounting, and back-office functions while simultaneously hiring in software development, AI engineering, and cloud services. The message across large organizations is consistent: automate the old, hire for the new.

What this framing ignores: institutional knowledge does not transfer like a file. When experienced workers leave an organization, they take with them years of context, exception-handling capacity, and judgment about when the process is right and when the process is wrong. Knowledge that was never written down because it did not need to be. It lived in the people. A company culture is rarely documented.

The hybrid (or augmented, like we used to call) model that companies like Klarna are now moving toward is essentially what they had before. Ideally, we have humans and machines working together, each one doing its half. However, they fired the human half. They are now surprised that the machine half cannot carry the weight of both. I wish I could say this was a niche problem. An exception. It is not.

Michael Sandel, in "What Money Can't Buy" (2012), makes an argument that applies here with uncomfortable precision:

Markets do not merely allocate resources, they express values.

When a company calculates that a decade of institutional knowledge is worth less than a quarterly cost reduction, it is not just making a financial decision. It is making a statement about what it values. Those statements have consequences that outlast the quarter they were made in. They communicate load and clear what that company believes, and we, the consumers, are always listening.

The social safety net nobody budgeted for

Governments are, in many cases, funding both sides of the same transaction. History has demonstrated that this is necessary for great accomplishments: from exploring the other side of the globe to traveling to other planets.

On one side: tax incentives, subsidized infrastructure, and public research funding that makes AI adoption cheaper for private companies. On the other: unemployment benefits, retraining programs, and income support for the workers that adoption displaces. The Congressional Budget Office projected in December 2024 that AI-driven displacement could increase federal income-support spending over an extended period, with labor reallocation stretching years rather than months.

By March 2026, economists and investors were still pitching Washington on the design of an AI-driven job-loss safety net. The policy response is running well behind the displacement curve. This is not a criticism of government specifically. It is a structural reality: the speed of technological deployment and the speed of democratic policymaking are not matched, and the gap between them is where displaced workers spend years of their lives waiting for a system that is still being designed.

Research from the Brookings Institution on U.S. workers' capacity to adapt to AI-driven displacement adds a dimension the optimistic projections tend to skip. The adaptation capacity of affected populations is low, slow, and deeply unequal, concentrated in communities that already had fewer economic options to begin with. In other words:

The people most affected by AI are the ones least capable to adapt to this new reality.

The family and mental health cost

Hourly work is not just income. It is not all about the money or survival.

The way job displacement gets discussed in economic models tends to reduce it to a wage replacement problem. Research on unemployment's psychological effects consistently shows impacts far beyond the paycheck: loss of purpose, social isolation, elevated rates of anxiety and depression, and health deterioration that compounds over time. When displacement is fast, concentrated in specific communities, and followed by years before equivalent employment becomes available, these are not temporary disruptions. They become chronic conditions. They shape generations.

The two-tier dynamic now taking shape is not random. Those who work with AI (mostly graduate-educated, urban, in sectors already doing well) and those displaced by it (mostly hourly, without advanced degrees, in sectors that cannot absorb the transition quickly) are not scattered evenly across the population. They cluster along the demographic and geographic lines along which political radicalization also tends to run. That is not a coincidence worth ignoring.

Even with all the money in the World and all the government incentives, we need to find something else to replace the job of those who are not being invited to this Brave New World we are building. If we don't, the consequences in communities, or even countries, will be devastating.

Part three: Long-term costs (5 to 10+ years)

Price tag pending; the lake is already paying.

The Acemoglu question

Daron Acemoglu and Simon Johnson published "Power and Progress" in 2023. The core argument is not a prediction about AI specifically. It is an observation rooted in economic history: gains from automation have rarely, on their own, distributed broadly. The industrial revolution created enormous wealth. It also concentrated that wealth for decades before labor movements, regulation, and sustained political pressure forced a rebalancing. The same thing happened with electrification. With computing. In each case, gains flowed upward first and broadly later, and the "later" part required active intervention. It did not happen on its own. Every right was fight for.

AI is moving faster than any prior technological wave. The question is whether the rebalancing mechanism exists this time and whether it can operate at the required speed. Nothing in the current policy landscape suggests it can. Nothing in the current corporate incentive structure suggests it will try if not provoked. We are too busy hatting the other side of the aisle for XVII Century topics instead of discussing it together our present and future.

Acemoglu's 2024 GDP projection, 1.1 to 1.6% over ten years, is modest by most accounts. I recommend reading it against the full ledger: the costs absorbed by:

Workers who cannot transition fast enough, by governments funding displacement without receiving the tax base that employment would have generated, by communities and ecosystems bearing the infrastructure burden of an industry that does not pay the full cost of its operations.

That is the actual math. The number looks different when you include the other side of the balance sheet.

What Arendt saw coming

Let me mention someone that I often do when talking about the anthropological and philosophical side of AI development. Hannah Arendt published "The Human Condition" in 1958, book that I referenced many times before in multiple articles. She was writing about automation in the early industrial age, about the political implications of machines doing what people used to do. She did not know about large language models. She did not need to. The technology is new, but we are not.

Arendt distinguished three modes of human activity. Labor: the cyclical, biological work of maintaining life. Work: the production of lasting things through skill and craft, things that outlive the maker. Action: the capacity to begin something genuinely new, which she considered the seat of freedom and political life. Her concern, articulated sixty-seven years before the current wave, was that:

Automation's deepest risk was not unemployment. It was the erosion of meaning.

The pattern of AI displacement follows her taxonomy with uncomfortable precision. Labor goes first: repetitive, cyclical, lowest-margin tasks. This is already happening at scale. Work is next: documents, analyses, designs, code, decisions. Already underway. What remains, in Arendt's framework, is action. But action requires conditions. Economic participation. Social standing. The lived experience of being needed by something larger than yourself. I brifly discussed this in my article about AI Art.

A society where a shrinking minority initiates and a growing majority simply subsists is not, in Arendt's terms, a free society. It is a management problem that has learned to dress itself up as progress. A 2025 paper in the Journal of Business Ethics, "Rethinking Automation and the Future of Work with Hannah Arendt," puts the same argument in contemporary academic language and reaches the same conclusion.

The political invoice

UBI: Universal Basic Income. The technical elite's preferred answer to the question of what happens to the workers that AI displaces. The idea is clean: AI generates enormous productivity gains, those gains fund a basic income for everyone, the distribution problem solves itself. Elegant on a whiteboard. Easy to defend on an academic room. Also, not yet real.

Let's get back a step because this is something I have been reading about for a few years now. Do we really need UBI? Short answer is yes. Why? Because the alternatives are worse. If work really does disappear faster than it returns, and the new work asks for credentials the displaced do not have and cannot get quickly, you are left with a large group of people who did nothing wrong and have no way to earn a living.

A society can handle that a few ways:

Option A: It can let them fall and pay for it later in crime, illness, addiction, and political anger, which is what the family ledger and the two-tier split were already describing. Just check what happened in 1929 stock market crash. Now, do nothing and see that effect times ten.

Option B: It can run them through the existing patchwork of means-tested benefits, which is slow, stigmatizing, and built for a downturn that ends rather than a shift that does not.

Option C: it can cut the link between survival and a paycheck and simply send the money.

That last option is what UBI is. Not a reward for being replaced, but a floor under people, so that losing your job to a machine does not also mean losing your home, your health, and your place in the world. The case for it does not rest on generosity. It rests on the rest of this invoice. Someone pays either way. UBI is just the argument for paying on purpose instead of by accident.

Research published in 2025 on AI-capability thresholds for rent-funded UBI notes carefully that the level of AI productivity required to fund a meaningful universal income at scale has not been demonstrated and may not arrive on the timeline that displacement is operating on. Economists and investors pitching Washington on AI job-loss safety nets in early 2026 were still designing the thing. The workers displaced now are not waiting for the design to be finished. They rarely have the luxury of waiting.

The structural issue runs deeper than timing. Democracy depends on economic participation in a way that goes beyond income. Political agency, social standing, and the capacity to organize and exert pressure all correlate with economic participation.

If you don't have the bare minimum to survive, you can't function as a citizen.

Power follows money. When gains concentrate, power concentrates. When power concentrates, the conditions for correcting that concentration become harder to build. This is not a novel pattern. It is the oldest one in economic history. AI is just running it at a speed that makes the adjustment window much smaller.

On the other hand, AI also brings a different kind of power: knowledge. More and more often, people rely on what the AI told them blindly. LLM responses sound so reasonable that it must be true, right? No fact-check, just pure believe. The models running behind the scenes are control by a very small number of big techs. Those few people have the power to decide what we think, what we know, what we don't know. If that power goes unchecked, unregulated, they can decide who we vote for, what causes we support and much more. Is that the future we want?

Research from PMC/NCBI published in 2025, examining AI, UBI, and power dynamics, uses the phrase "symbolic violence" to describe the tech elite's framing of UBI as a solution. The critique is pointed: proposing a solution that depends on your own continued success and distributes benefit on your own schedule is not generosity. It is the management of a problem you created. Should the ones who create a problem be the ones who proposes a solution? The answer is obvious: the solution is owned by all of us, as a society, not by a small group of CEOs and politicians.

The environmental horizon

In one upper-bound MIT estimate, data centers could reach 21% of global energy demand by 2030 once the full cost of delivering AI is counted. Mainstream forecasts from the IEA are far lower, around 3% of global electricity. Even the low end is a steep climb from where we are now.

Now, let's go to the next alarming data. One in four data center assets globally may face increased water scarcity by 2050, according to an MSCI analysis of 14,000 facilities. These are not speculative scenarios. They are capital expenditure plans already in motion. They are not in a sci-fi script; they are in a Jira board.

The grid is feeling it now. Communities near major data center clusters are dealing with power cost increases, water table pressures, and infrastructure strain they did not agree to absorb and are not being compensated for absorbing. The environmental cost of AI, like most environmental costs in market economies, is being externalized to the people with the least power to push back. There is a version of the climate conversation that focuses on individual carbon footprints. We try to save the planet by taking shorter baths and recycling but the real impact is elsewhere.

Back to Sarah

It is 2028 now. Sarah's company is doing reasonably well.

The AI platform handles the routine queries. Human agents, rehired at somewhat lower wages and with narrower responsibilities, handle the escalations. The arrangement is called a hybrid approach. It is also essentially what they had three years ago, before they fired forty people and spent eighteen months learning that the machine could not carry the weight alone.

The forty agents took between six months and two years to find equivalent work. The government covered their unemployment in the interim. Two of them are still in lower-paying roles. The river near the data center in Virginia is running a few degrees warmer. The customer who waited forty-seven minutes for a bot that could not help and a human who no longer felt empowered to make a decision without escalating three levels is probably not coming back.

The bill did not disappear. It was distributed. Quietly, efficiently, and without anyone's explicit consent.

The other side of the invoice

Before you close this tab, I want to be clear about something.

I am not against AI. I have been studying it for twenty years and talking about it for fifteen years, long before it became a boardroom keyword and a LinkedIn hype. My master's degree was in AI. I am a data scientist by training and by instinct, and apparently by body art: I have a loss function tattooed on my arm (the most committed career statement in the history of professional development, or a very specific kind of occupational hazard, possibly both).

A loss function is how a model learns. It measures the cost of being wrong. Without it, the model cannot improve. It just keeps making the same mistake with increasing confidence, which is an accurate description of several major AI deployment decisions made in the last three years.

What I am arguing is not that AI should stop. Let's be real: it will not stop. What I am arguing is that we apply the same principle to ourselves. Run an honest loss function on the decisions we are making. What did this automation actually cost? Who absorbed it? Is the answer informing the next decision, or are we running the same model again and expecting different outputs? And most importantly:

What adopting this technology the way we are doing tell the public about my values?

There is no free lunch in machine learning. Every optimization is a trade-off. Every gain in one metric costs something in another. The engineers in this field understand this instinctively. It is time for the business decisions to catch up.

The cost goes beyond the token. It always did. We just needed to read the full bill.

What now?

Reading the invoice is not the same as paying it down. So here is the part where I stop describing the problem and discuss what we can do about it. And that depends from which angle you look at.

If you are the employee:

You are the one the spreadsheet found first, or the one it finds next. "Learn AI" is not bad advice, but people throw it around like it costs nothing, and it does. What actually protects you is not knowing how to write a prompt. It is being the person who holds the judgment the machine cannot fake: the exceptions, the context, the moment when the process is wrong and someone has to say so out loud. Get closer to that work, not further from it. Write down what you know. The fact that nobody ever asked you to is exactly what makes it valuable. And watch your own employer with clear eyes. If the automation announcement arrived with a lot of confidence and very few numbers about quality, you already know how where story leads. In short, move to a more analytical and business-driven perspective of your role.

If you are the decision-maker:

You are holding the pen when the bill gets written, and most of it gets addressed to someone who is not in your room. Saving cost on the short-term is tempting. But pause a little. Run the loss function before you cut, not eighteen months after. Model the integration cost, the escalation cost, the quality cost, and the cost of the knowledge that walks out the door and does not come back. Ask for the customer satisfaction score even when nobody requested it. Especially then. And be honest about what "hybrid model" means when you arrive at it the expensive way. Klarna got there by firing people and rehiring them. You can get there by not firing them in the first place. One of those paths costs far less, and it is not the one that looks bold on a slide.

When you are the average citizen:

You did not deploy anything, but you are on the invoice too. It reaches you through your water table, your power bill, the tax dollars funding both the subsidy and the unemployment check, and the political weather that turns when a whole category of work quietly disappears. One person cannot do much here. A lot of people can. Ask where the data center gets its water and who signed off. Ask your representatives what the plan is for the displacement already happening, not the one economists are still sketching out. Vote like the answer matters, because it does. And push back on the tidy story that all of this is just progress showing up on time. Progress that only ever flows upward is not a law of physics. Somebody chose it, and choices can be fought.

None of this stops the technology, and I am not asking it to. I am asking that the people writing the bill and the people paying it stop being two different groups who never meet.

That gap is the real cost. The rest is just line items.

Author: Igor Alcantara

Note: This article was written by a human. AI was used for the illustrations and grammar/spelling check.