{"slug": "what-if-ai-became-more-expensive-than-the-workers-it-replaced", "title": "What if AI became more expensive than the workers it replaced?", "summary": "A developer's analysis suggests that the economic promise of AI replacing human workers is faltering, citing examples like Klarna rehiring customer service reps and Gartner's forecast that half of companies cutting such staff will rebuild teams by 2027. The piece highlights rising costs of AI usage, such as Uber burning its annual AI coding budget in four months, and argues that while per-token costs have plummeted, total spending escalates due to rapid adoption, leading to a 'tokenmaxxing' culture that confuses usage with productivity.", "body_md": "I write code with an AI assistant sitting right beside me most days, and it genuinely makes me faster. THres No argument there. But something has been flying around in my head all year: what if the core promise everyone built their AI strategy on, **“replace the human, save the money,”** turns out to be way shakier than the pitch decks suggest?\n\nAnd the truth is, AI is not bad. It’s just messier than advertised.\n\nSo, let’s look into it together…\n\nYou've probably heard some version of this. Back in 2024, Klarna announced that its AI agent was doing the work of 700 human customer service reps. Hiring froze. The support team shrank. It became the go-to example in every AI-hype newsletter, mine included, for how AI had already eaten a whole job category.\n\nThen, quietly, Klarna started hiring the reps back... wtf is happening😂\n\n...Turns out that's not a one-off. It's a pattern with numbers behind it now. Gartner's forecast: by 2027, half of the companies that cut customer service staff over AI will be rebuilding those same teams, usually under a new job title so nobody has to say \"we were wrong\" out loud. A separate Orgvue survey put the regret rate at 55% of employers who swapped people for AI. And a February 2026 survey of 600 HR professionals found two in three companies with AI-driven layoffs are already rehiring, more than half of them bringing back over half the roles they cut.\n\nNone of this means the AI was useless. One thing researchers keep flagging: the AI handles maybe 60% of a role competently, and the remaining 40% needs institutional memory, customer trust, and judgment calls that no training run captures. That last 40% is expensive to live without.\n\nHere's the part that stings. Abeg, who approved this math? Companies don't rehire at the old rate. Positions that paid $55k before the cuts are now going for $75k or more, because whoever fills the role has to double as an AI handler: managing it, auditing it, prompting it well. You don't just undo the layoff. You buy a more expensive version of the job, plus a subscription fee on top of that.\n\nAnd that subscription fee has its own way of spiraling.\n\nTake Uber. Its CTO admitted this year that the company burned through its entire 2026 AI coding budget in about four months. By March, 84% of Uber's engineers had adopted Claude Code, and something like 70% of shipped code was AI-generated. Massive usage. Whether that usage turned into proportional value is a lot murkier.\n\nThere's a name for the culture behind numbers like that: tokenmaxxing, engineers racing to burn as many tokens as possible because heavy usage gets mistaken for heavy output. At Meta, some engineers reportedly chase leaderboard bragging rights on an internal usage tracker. One exec's response to the whole trend was the sharpest I've read: a big budget doesn't prove a campaign worked, so a big token count shouldn't prove an engineer did either. High burn rate isn't the same thing as a high success rate.\n\nAn Nvidia engineering VP put it even more bluntly: for his team, the compute bill has completely outgrown what they used to spend on salaries.\n\nIf this post stopped here, it would just be another \"AI is a money pit\" take, and that's not honest either. Here's the part that complicates everything.\n\nPer-token cost, the actual price of running these models, has been crashing. Stanford's AI Index tracked something wild: running a GPT-3.5-level system got over 280 times cheaper between late 2022 and late 2024 alone. Hardware is getting roughly 30% cheaper every year, and energy efficiency is climbing about 40% annually. One 2026 economics paper tracking prices since 2020 puts the market-wide decline at close to 600-fold.\n\nSo how do \"tokens are 600 times cheaper\" and \"Uber burned its yearly budget in four months\" both sit true at once?\n\nTwo things, happening at the same time. Usage is scaling way faster than price is falling, so total spend keeps climbing even as the per-unit cost drops. And the deflation isn't evenly spread: that same 2026 paper found flagship reasoning models barely follow the falling curve at all, they carry a premium averaging more than 31 times what non-reasoning models cost. The cheap AI everyone cites is usually the small, fast, non-reasoning tier. The AI actually being deployed to do hard, worker-replacing work is a different and much pricier animal entirely.\n\nZoom out even further and MIT's 2026 NANDA study drops a number that should be pinned above every AI steering committee meeting: 95 out of 100 enterprise generative AI pilots show no measurable financial return. Not underperform. Zero. This is after $30 to 40 billion in enterprise AI spend.\n\nThe researchers didn't blame the models. They blamed the architecture around them, bolting a chatbot onto a document library and calling it transformation, instead of actually redesigning how the work gets done. The 5% that succeed are earning $3.70 back for every dollar spent, which tells you this isn't really an \"AI doesn't work\" story. It's an execution-gap story wearing an economics costume.\n\nIf I had to compress this whole thing into one line: AI itself didn't get expensive. Getting AI to safely, sustainably replace a human, without quietly rebuilding that human's role at a premium six months later, that's the expensive part. Compute is one line item. Rehiring, retraining, lost institutional memory, and the pilot that never made it past a demo, those are the real costs nobody put in the original slide.\n\nBuilding solo the way I do, using AI as leverage rather than as a replacement for anybody, I try to hold both truths at once: the tools genuinely are getting cheaper and more capable by the month, and \"just swap the team for AI\" almost never pencils out as cleanly as it sounds in a boardroom.\n\nSo if you run a team, or even just manage your own workflow, the right question to ask might not be “How much will AI save me?” It’s probably: “What’s the full cost of doing this without a human in the loop once the edge cases start showing up in month six?” 🙄\n\nAnd I’m genuinely curious: has anyone reading this watched a company or client go through the replace-then-quietly-rehire cycle up close?\n\nI seriously want to hear from you.", "url": "https://wpnews.pro/news/what-if-ai-became-more-expensive-than-the-workers-it-replaced", "canonical_source": "https://dev.to/ctrotech/what-if-ai-became-more-expensive-than-the-workers-it-replaced-gbf", "published_at": "2026-09-02 14:59:14+00:00", "updated_at": "2026-09-02 15:25:23.562082+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-products", "ai-infrastructure"], "entities": ["Klarna", "Gartner", "Orgvue", "Uber", "Meta", "Nvidia", "Stanford AI Index", "Claude Code"], "alternates": {"html": "https://wpnews.pro/news/what-if-ai-became-more-expensive-than-the-workers-it-replaced", "markdown": "https://wpnews.pro/news/what-if-ai-became-more-expensive-than-the-workers-it-replaced.md", "text": "https://wpnews.pro/news/what-if-ai-became-more-expensive-than-the-workers-it-replaced.txt", "jsonld": "https://wpnews.pro/news/what-if-ai-became-more-expensive-than-the-workers-it-replaced.jsonld"}}