Representative Image / Pexels

The story of artificial intelligence (AI) appeared plausible not so long ago: it will replace expensive human labour, reduce operating costs and achieve unheard-of levels of productivity. Behind it, investors rushed in. CEOs pushed it. Boards adored it.

The excitement of keynote speeches and billions of dollars being poured into data centers aside, thousands of employees were being told their employment had been made redundant by algorithms.

This loss of manpower was severe. Challenger, Gray & Christmas says that of the expected loss of 1.17 million jobs in 2025, about 55,000 will be directly attributed to AI. Microsoft eliminated over 15,000 jobs in many rounds, and Salesforce cut 4,000 customer-support jobs.

Also Read: When AI starts to feel like a new religion

Marc Benioff, the chief executive, said that AI currently does 50% of the work at the company. Workday shed 8.5% of its workforce in a move to focus on AI projects, and Amazon announced plans to reduce 14,000 corporate jobs.

The financial argument was simple: why continue paying for human labour when artificial intelligence could do the same for less? Except, it turns out, not to be cheaper at all.

“Uber burned through its entire 2026 AI budget in four months — and its COO admitted he can't pin that spending to a single useful consumer feature.

Tokenmaxxing When the Bill Comes Due

One of the signature mistakes of the AI age is the idea of “tokenmaxxing” — the crazy dash to burn AI credits at the highest rate, assuming that more consumption means more value. The results increasingly look like a case study in executive overconfidence.

Uber burned through their entire 2026 budget for Claude Code and Cursor in only four months, with 70% of committed code being AI-generated and 95% of developers using AI tools weekly. The company’s COO, Andrew Macdonald, said that there was no clear correlation between the company’s surge in token use and major product developments. As one expert put it, “The economics are brutal.”

Starbucks definitely fell off. An AI-powered inventory system meant to automate counts of milk and beverage supplies was mothballed after it regularly miscounted and misidentified commodities. To be a company that prides itself on operational stability and not have the technology to precisely count inventory.

Even more alarmingly, Axios reported that an unnamed corporation had not put usage limits in place and burned $500 million in AI credits in one month. No limitations on spending and half a billion dollars disappeared.

Meanwhile, Microsoft has quietly dialed back its promises on AI infrastructure, including the alleged cancellation of data-center leases worth as much as two gigawatts. The company also stopped diverting developers away from Anthropic’s Claude Code and toward its own proprietary Copilot CLI tool at the same time its fiscal year was coming to a close.

Hype Architects Take a Breather

What makes this reckoning so stunning is who is now recanting. The CEOs who sold the world on the transformative possibilities of AI are quietly — and, at times, brazenly — rethinking their projections.

OpenAI CEO Sam Altman, who once said AI would “probably replace most of the jobs people do today” and that entire job categories will be “totally, totally gone,” says he is “delighted to be wrong.” At a Commonwealth Bank of Australia summit in Sydney, Altman said his estimate of the economic impact of AI had been “pretty wrong”. “I thought there would have been more impact by now on the elimination of entry-level white-collar jobs than there has been,” he said. "I think I understand more why it hasn't.

Anthropic CEO Dario Amodei has also walked back his previous claim that AI may take 50 percent of white-collar employment, instead suggesting that automation would in fact increase the amount of labor people undertake. David Solomon, CEO of Goldman Sachs, said since late 2025, “You had a lot of fear around the economic environment, which was overblown. You look at 100 years of American economic history.

Yale Budget Lab research confirms what the statistics has always silently suggested: unemployment for those in high-AI-exposure industries has remained pretty much flat until March 2026. At least for now, A.I. is replacing responsibilities within jobs, not the jobs themselves. “AI was a convenient excuse for some companies to make layoffs that were going to happen anyway. Sam Altman himself termed it “AI washing.”

But behind the CEO’s mea culpas lies the disturbing fact that the layoffs still happened. The question of whether AI “caused” those cutbacks becomes essentially easy to the worker updating their résumé at midnight when Oracle quickly fired at least 10,000 people after claiming great results and funneled the payroll directly into AI data-center expansion.

Altman himself revealed what experts long have suspected: Some organizations are “AI washing” – employing the technology as a convenient cover for layoffs that would have happened anyhow.

The Economy Was Never Healthy

While the gold rush was going on, nobody wanted to accept the basic math of AI adoption was flawed from the outset. For the first time in the history of business, enterprise technology costs about as much as humans, and CFOs are making the comparison openly. Within weeks, annual AI funding dried up.

There is a glimmer of hope from a Gartner study that says that by 2030, the cost of inference for generative AI models will likely fall to around 10% of its 2025 levels. But that promise comes with a grim dose of realities from the same studies. The kind of autonomous systems corporations are betting on, agentic AI, requires many more tokens per task than typical models.

Goldman Sachs estimates demand for the tokens might be increased as much as 24 times by AI agents. Also, carriers are moving toward usage-based pricing, so even though prices are projected to fall, businesses won’t see any savings. “We envision a future where intelligence is a utility, like electricity or water, and people purchase it from us on a meter,” Altman said.

On a metre. " That doesn’t imply prices will come down. That promises infinite scalable billing.

How AI Really Should Be Used in Industry – A Sanity Road Map

That’s not to say that AI is a fraud, or that its revolutionary potential is unfounded. It demonstrates that the industry committed a category error in considering a powerful but unproven technology as a finished good ready for extensive industrial deployment. There is a path forward, but it demands self-control, humility, and a complete rejection of the tokenmaxxing attitude.

Begin with problems, not technology

“The biggest mistake organizations made was to buy AI and then find problems it could solve, not finding actual operational pain points and determining if AI was the right answer. Every deployment should begin with a particular, quantifiable business challenge, like “our average resolution time for Tier-1 support tickets is 4.2 hours; can AI reduce that to under 1 hour, and at what cost per ticket?” instead of “we should use AI for customer service.” If the business case looks bad on paper, it will probably look worse in production.

AI Budgets Should Be Treated As Infrastructure, Not Tests

Uber isn’t a victim of a technological glitch. It’s a government failure. AI expenditures, like any other big infrastructure expenditure, must be subject to stringent controls. Departmental limitations, usage dashboards, hard caps, compulsory ROI reviews every ninety days, etc. Companies that had based their risk management strategies on stable, predictable AI costs are suddenly learning they are anything but (often at a high cost). Usage-based billing requires usage-based governance.

Preserve Human Knowledge Through Times of Change

The damage done to institutional knowledge by firms that have terminated employees and then gently tried to rehire them often as contractors at cheaper pay is worth far more than their balance sheets. Salesforce eliminated its customer service employees and discovered that AI was not equipped to deal with the complexity of real-life customer issues. And there was no one remaining who knew how. Before cutting people, industries need to embrace purposeful overlap, running AI and human teams side by side, and exploiting the transition stage to document, train and validate. The human will not go until the machine truly masters the task—not just approximates it.

Design for Auditability Instead of Automation

Starbucks’ AI inventory system did not fail because the model powering the system was bad. It failed because there was no major human checkpoint in the loop. The milk went out because the machine counted it wrong and no one noticed. In an enterprise setting, AI needs to have explicit audit layers—places where a human being looks at, checks and approves the AI’s output before it touches operations. The idea is not to take humans out of the loop, but to change their behavior in it.

Ask suppliers for contracts linked to outcomes

Selling promise instead of performance has been a profitable approach for the AI vendor group. Businesses need to adjust the terms of engagement. AI contracts should incorporate outcome clauses requiring renegotiation or termination if the tool does not demonstrably improve the target measure by a set percentage within a predetermined timeframe. Procurement teams are already rethinking their approach to AI, as Fortune 500 clients’ cost reckoning, according to Arvind Jain, CEO of Glean. This pressure is good and should be codified.

Before scaling, build internal AI literacy

One of the most overlooked costs of adopting AI is not the tokens themselves but the tokens wasted by personnel who don't know how to use these tools effectively. When engineers utilize AI agents to generate code at scale, the risk to the business increases faster than it can be managed, and reviewers are unable to assess the quality of the output. The big rollout needs to happen first, not after investments in AI literacy such as rapid engineering, output review, and knowing when not to use AI.

Build Reasonable Paths Between Models

“Intelligent companies are increasingly spreading their work across multiple AI models to save money,” says Matan Grinberg, CEO of Factory AI. “They reserve the expensive frontier models for the work that truly needs them and use cheaper and more efficient models for everything else.” Not every job requires a frontier model. Many can be solved with smaller, faster, cheaper solutions without sacrificing much quality. AI infrastructure needs to be infused with routing intelligence to survive at scale.

The Reckoning Starts

Another version of this story has a sad ending – companies that cut their workforces, spent billions on AI infrastructure that hasn’t justified the expense, and are now caught between investor expectations, vendor lock-in, and the quiet shame of rehiring the people they let go. That’s already happening in that version.

But there's another way to look at it, which is that the industry should see it as a lesson to learn, rather than a crisis to manage. AI is not a cure, it cannot replace human judgment and it is not going to pay for itself overnight. Used with humility and discipline, this powerful mix of skills may change the way organizations work fundamentally. The companies that move fastest won’t be the companies that drive the adoption of AI over the next ten years. They moved with the utmost determination.

So here’s the bill. Now comes the hard work of making something that genuinely warrants it.

 

Manpreet Singh is an economist and Assistant Vice President, GENPACT. Badri Narayanan Gopalakrishnan is a senior economist and affiliate faculty member at the University of Washington Seattle.

 

(The views and opinions expressed in this article are those of the author and do not necessarily reflect the official policy or position of New India Abroad.)

Discover more at New India Abroad.