It turns out that “use AI for everything” is excellent advice — right up until the moment the bill arrives. Barely months after companies enthusiastically urged staff to embrace generative AI, a quieter, more accountable memo has gone out: slow down, we’re bleeding tokens. What began as a productivity revolution is bumping into a rather old-fashioned constraint — money.
The culprit is not a lack of enthusiasm. If anything, that’s the problem. Tokens — the basic units of text that AI models chew through to produce their outputs — are being consumed at a rate that has caught finance teams off guard. Throw in agentic workflows (AI systems that autonomously execute multi-step tasks), software licence fees and infrastructure costs, and the bill looks rather less like a modest IT upgrade and rather more like a second headquarters.
The corporate confessions have been remarkable in their candour. Uber reportedly exhausted its annual AI budget in four months before capping consumption at $1,500 per employee per month for agentic coding tools. Microsoft cancelled licences for Anthropic’s Claude Code for some internal users, redirecting developers to GitHub Copilot CLI instead.
Amazon, perhaps most entertainingly, had to withdraw an internal AI token leaderboard after employees began gaming it — apparently treating “most tokens used” as a performance metric worth optimising for. It was not.
“Enterprises are applying significantly tighter financial scrutiny to AI investments as compute intensity rises and ROI remains uncertain,” said Biswajeet Mahapatra, principal analyst at Forrester. The arithmetic is sobering: generative AI models can require up to 100 times more computing power than traditional AI systems, making them considerably more expensive to run at scale — and considerably less forgiving of casual, unconstrained use.
Gartner analyst Anushree Verma has done the maths on where this trajectory leads. By 2027, she estimates, 80% of enterprises will face significant budget overruns from consumption-based AI coding tools. In many cases, software tool costs will exceed the human costs of software development. Read that again.
The recommendations from analysts are sensible, if unglamorous: align licences with actual employee needs, monitor token consumption, introduce model selection controls, and choose vendors that don’t encourage unnecessary back-and-forth. In other words, the discipline that enterprises apply to every other major technology investment has arrived, fashionably late, at the AI party.
Not everyone is resorting to strict rationing, however. Cybersecurity firm Palo Alto Networks, where over 6,000 engineers use AI coding tools daily, has opted for team-level budgets and a centralised governance framework rather than per-employee quotas.
“We don’t believe in micro-managing individual employee token quotas, which risks stifling the very innovation we’re trying to unlock,” said Swapna Bapat, the company’s vice-president and managing director for India and SAARC. It is a reminder that oversight and restriction are not the same thing.
The spending reckoning is also prompting a more fundamental question: what, exactly, are companies buying? Alex Bouaziz, co-founder and CEO of HR technology firm Deel, argues that the framing of AI-versus-human is a false one. At Deel, AI agents are handling payroll queries, onboarding paperwork and compliance processes — the administrative sediment that tends to accumulate at the bottom of every job description — freeing employees for higher-value work. The goal, he says, is not replacement but amplification.
“The sharper point is that the real measure of AI cost today is tokenomics: what an agent consumes, not what it’s licensed at,” Adam Auerbach, VP, Head of Applied AI for North America at EPAM noted. As per an AI index by Stanford, per-token prices are collapsing yet enterprise bills aren’t, because agents and reasoning models run in iterative loops that consume tokens at a completely different scale.
Stanford’s Digital Economy Lab found agentic tasks burn roughly 1,000 times more tokens than a simple query, and the same task can vary in cost up to multi-fold, with models unable to predict their own spend. “That unpredictability is precisely what enterprises are struggling with: you can’t budget or price an agent reliably when the bill only becomes visible after the work is done. So, the discipline shifts from negotiating a rate to governing consumption and scoping use cases tightly,” he went on.
Vijay Gopalakrishnan, Partner at Deloitte India, offers a useful lens: the use cases most worth pursuing are those tied to a company’s core strategic pain points, not just wherever AI happens to be technically capable. Successful deployments, he notes, depend on monitoring both performance and costs — simultaneously.
Which brings us to what may be the defining insight of this moment. Vishal Sikka, former Infosys CEO and founder of Vianai, captured it precisely in a post on X: “Token costs are becoming a real issue.”
What is changing, he wrote, is not enterprise appetite for AI, but the assumption that more usage automatically equals more value. The next phase of the market, Sikka believes, will see three things happen at once — higher adoption, cheaper models, and tighter scrutiny of how AI is actually being used. Cheap intelligence, it turns out, still needs to be used wisely.
