"For a brief, dazzling moment, Chegg looked like the future of learning. The subscription-based academic help platform built a business on something simple: students had questions, and Chegg had answers. At its peak, it served tens of millions of users and was valued at nearly $12 billion. Then came ChatGPT, and the future arrived ahead of schedule — just not in the way Chegg had planned."

Almost overnight, students stopped paying for homework help. Why subscribe to Chegg when a free AI chatbot could answer the same questions in seconds? Traffic collapsed. Subscribers evaporated. By early 2025, non-subscriber traffic had fallen 49% year-over-year. In May 2025, Chegg laid off 22% of its workforce. By October 2025, it cut another 45% — more than half its remaining staff gone in under six months. The company that built a business answering questions had been silenced by the very technology it once tried to partner with.

Chegg’s story is not a cautionary tale about AI. It is the opening chapter of a much larger reckoning now playing out in boardrooms across the world — a story about who really pays when the AI bill arrives.

The Great Displacement: When Companies Fired the Humans

In 2022 and 2023, the narrative was intoxicating. AI had finally arrived at scale, and the economics seemed irresistible. Executives saw a once-in-a-generation chance to automate expensive, repetitive human labor and redirect that capital toward growth. The layoffs came swiftly and, in many cases, proudly.

Klarna, the Swedish buy-now-pay-later giant, became the poster child of the movement. The company reduced its workforce from roughly 7,000 employees in 2022 to around 3,000 — and then openly celebrated. In 2024, CEO Sebastian Siemiatkowski announced that Klarna’s AI agent was handling the workload of 700 full-time employees, covering customer queries and refund processing at unprecedented scale. The coverage was breathless. The future, it seemed, had arrived. Duolingo followed, terminating 10% of its contractor workforceand signalling a firm pivot to ‘AI-first’ operations. Google restructured entire divisions around AI-driven ad tools. Across industries, the message was the same: humans were expensive, AI was cheap, and efficiency was king.

The math looked obvious on a slide deck. It looked very different in practice.

The Price Tag Nobody Budgeted For

Here is the number the boardroom presentations left out: deploying generative AI across an enterprise can cost between $5 million and $20 million, according to Gartner. Building a custom model from scratch pushes that figure higher still — up to $20 million — with a further $6.5 million to customise it and embed the APIs into existing applications. Once deployed, companies can expect to spend more than $10,000 in recurring annual costs per user.

That is before the hidden infrastructure arrives. As one analysis of enterprise AI economics found, an AI agent operating continuously costs $50,000–$100,000 per year in compute alone — roughly equivalent to an entry-level human salary, before you account for the engineers to maintain it, the oversight staff to catch its errors, or the platform teams to keep it running. In some cases, the math has inverted entirely. Bryan Catanzaro, vice president of applied deep learning at Nvidia — one of the primary beneficiaries of the AI boom — made the admission plainly in April 2026: “For my team, the cost of compute is far beyond the costs of the employees.”

“For my team, the cost of compute is far beyond the costs of the employees.” — Bryan Catanzaro, VP Applied Deep Learning, Nvidia (April 2026)

Catanzaro’s confession landed with the weight of irony. Nvidia sells the chips powering the AI revolution. If Nvidia’s own engineers are finding that AI costs more than it saves, the implications for every other enterprise — operating without Nvidia’s economies of scale — are severe.

A 2024 MIT study quantified the problem more precisely: AI automation is economically viable in only 23% of roles that rely heavily on visual tasks. In the remaining 77% of cases, keeping a human on payroll remains the cheaper option. Meanwhile, enterprise AI spending hit an average of $85,521 per month in 2025 — a 36% increase year-over-year — while only 51% of organisations could confidently confirm whether those investments were delivering any return.

Uber’s chief technology officer, Praveen Neppalli Naga, perhaps put the executive experience most succinctly when he told The Information in April 2026 that after pivoting to AI coding tools, he was “back to the drawing board because the budget I thought I would need is blown away already.”

Pilot Purgatory: The $644 Billion Graveyard

For many companies, the AI investment never even reached the stage of disappointment. Instead, it stalled in what industry analysts now call “pilot purgatory” — a limbo of technically functional proofs-of-concept that consume budget, require dedicated personnel, and never reach production.

The numbers are damning. The share of companies abandoning the majority of their AI initiatives jumped from 17% in 2024 to 42% in 2025, according to McKinsey — a figure that more than doubled in a single year. Gartner had predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025; the reality exceeded the forecast. A landmark 2025 MIT study of enterprise AI deployment found that 95% of pilot programmes delivered no measurable P&L impact. Nineteen out of twenty projects produced nothing.

One consulting firm estimated that pilot purgatory alone costs the average enterprise between $15 million and $25 million annually in wasted development, infrastructure spending, and opportunity costs. Most of those zombie projects never appear as failures in financial reporting. They linger on the books as “active initiatives” while quietly bleeding capital.

Gartner’s Distinguished VP Analyst Rita Sallam summarised the executive experience with characteristic precision: “After last year’s hype, executives are impatient to see returns on GenAI investments, yet organizations are struggling to prove and realize value. As the scope of initiatives widens, the financial burden of developing and deploying GenAI models is increasingly felt.”

The U-Turn: Rehiring the Humans Who Were Shown the Door

Klarna, the company that made global headlines by replacing 700 employees with a single AI system, is now hiring them back.

By mid-2025, the cracks in Klarna’s AI experiment had become impossible to ignore. Customer satisfaction had declined. Operational quality had slipped. The efficiency gains had been real, but the hidden costs — reputational, relational, and operational — had not been modelled. “We focused too much on efficiency and cost,” Siemiatkowski admitted. “The result was lower quality, and that’s not sustainable.” Klarna is now piloting what it calls an “Uber-style” model of flexible remote human agents — a description that inadvertently reveals the full circle the company has completed. It fired workers to save money. It spent more on AI. It is now rebuilding a human workforce, simply with worse job security than before.

Duolingo’s pivot was equally swift. Within a week of its “AI-first” announcement, CEO Luis von Ahn clarifiedthat he did “not see AI as replacing what our employees do — we are in fact continuing to hire at the same speed as before.”

Gartner now predicts that by 2027, 50% of companies that cut customer service staff due to AI will rehire people to perform similar functions — often under different job titles to preserve the fiction that the AI initiative succeeded. A February 2026 Forrester Research report found that 55% of employers already report regretting the layoffs they attributed to AI. Customer service leadership surveys show that 95% of executives now plan to keep human agents around strategically, rather than pursue the mass-automation model that looked so appealing in 2022.

The human return, when it comes, arrives rebranded. Customer Service Representatives become “Solution Consultants.” Support Agents are relaunched as “Trusted Advisors.” The semantics change. The underlying reality does not. The humans are back.

The Race Continues — But the Finish Line Has Moved

None of this means the AI race is slowing. If anything, the capital commitments are accelerating: big tech firms announced approximately $740 billion in AI-related capital expenditure in 2026 alone, a 69% jump from the prior year. Gartner projects worldwide IT spending will reach $6.31 trillion in 2026, driven substantially by AI infrastructure. The frontier models keep improving. The tools keep getting sharper.

What has changed is the story told around that investment. The early narrative — AI will replace human workers and deliver immediate, measurable cost savings — has collided with the empirical record. The revised narrative is quieter and more complicated: AI augments human work in specific, well-defined contexts; it requires substantial investment in infrastructure, oversight, and maintenance; and the ROI is neither automatic nor universal.

Investor Chamath Palihapitiya crystallised the new benchmark when he argued that AI agents need to be “at least twice as productive as another employee” to justify their total cost once token spend, infrastructure, and human supervision are included. Most enterprise pilots have not yet cleared that bar.

Chegg, meanwhile, is still fighting for survival. The company that was the first visible casualty of the AI disruption is now a symbol of something more nuanced than disruption — it is a symbol of what happens when a new technology arrives faster than the economics of adoption can be honestly calculated. ChatGPT did not destroy Chegg. The unexamined assumption that cheaper always means better did.

The Bill Always Comes Due

There is a pattern visible in the full arc of this story. Companies fired workers to cut costs. They spent more on AI than they saved. They are now rehiring, under different titles, at lower wages and with reduced security. The workers lose twice. The companies learn expensively. The AI vendors collect either way.

The technology is not the problem. The expectation was. Replacing human judgment wholesale — in customer service, in creative work, in complex professional contexts — turned out to be harder, more expensive, and less satisfying than a proof-of-concept pilot suggested. The gap between what AI can demonstrate and what it can sustain at enterprise scale, at real-world quality, at acceptable cost, remains wide.

The AI race is still going strong. The chips are still being built. The models are still improving. The capital is still flowing at historic scale. But somewhere between the boardroom slide and the quarterly earnings call, the honest conversation about what large-scale AI actually costs — in money, in quality, in human capital — is finally being had. The companies that have it sooner will be better positioned than the ones who wait for the bill to arrive.

It always does.

‍

Read - Does Your Startup Really Need an HR Department? The Bolt CEO's Case for Saying No — And What Founders Should Actually Do