By Sophie Lindqvist

By September 2026, enterprise artificial intelligence (AI) deployment spending had reached $177 billion, compared with $33 billion for all of 2025. That surge is pushing Goldman Sachs, Cisco and JPMorgan toward token budgets, spending caps and tighter model access. The aim is not necessarily lower spending, but more accountable use and better returns.

After years of encouraging staff to use AI, major employers now distinguish tools needed for specialist work from those used for routine productivity. Some still support broad experimentation while adding ways to measure consumption and security exposure.

Lari Hämäläinen, a senior partner in McKinsey & Company's technology, media and telecommunications practice, said workplace AI use was picking up rapidly, making guardrails necessary. That shift followed early programs that rewarded experimentation before cost controls became the priority.

Goldman Sachs is expanding its range of AI tools for 45,000 employees, but CEO and chairman David Solomon says not everyone needs the most advanced models every day. Solomon told Bloomberg TV: "We need frontier models for certain things, particularly cyber." Frontier models, meaning the most advanced AI systems, suit those tasks; routine productivity work does not need that tier.

A token is a unit of text or code processed by an AI model. Cisco's chief product officer, Jeetu Patel, compared the earlier policy with learning to ride a bicycle: coding teams received months of unlimited tokens to become fluent before efficiency became the priority.

As use increased, managers across his 32,000-person team began receiving token budgets. The shift suggests Cisco sees measured allocation as a condition for scaling, not as a goal to reduce usage.

JPMorgan imposed a $2,000 monthly cap in September 2026 on some engineers' use of Anthropic's Claude Code. A spokesperson said the bank was tying costs to productivity, quality, risk reduction and business impact. The announcement cites four measures but does not provide results against them.

From adoption to allocation

Before tighter controls, some employers treated high consumption as evidence of successful adoption. Dashboards initially made usage visible and, at times, encouraged heavier use.

• Visa doubled its monthly token use from 1 trillion in February to nearly 2 trillion in March and gave prizes to its power users.
• Nvidia was trying to spend as much as $2 billion annually on tokens for engineers. Jensen Huang said he would be deeply alarmed if a $500,000 engineer used less than $250,000 worth.
• Disney's April dashboard showed its biggest users making hundreds of requests per day on average and consuming tens of millions of tokens.
• KPMG US said in May that a dashboard for roughly 10,000 advisory workers should encourage more frequent and sophisticated use. It later stopped monitoring usage through that dashboard once AI became part of daily work.

These figures measure appetite, not returns. Menlo Ventures' 2026 study put enterprise AI deployment spending at $177 billion by September, up from $33 billion in all of 2025. Only 13% of about 2,000 technology executives polled by Gartner from May through July said their organizations had gained significant value from AI.

Without limits, staff will use AI in every possible way they can, according to Alex Sobol, cofounder of the Millennium Alliance, an invite-only community for senior executives. He attributes rising bills partly to workers having little incentive to moderate their usage.

Controls also address a separate risk: improper use of broad access. A 2025 KPMG survey of more than 1,000 US workers found 44% used AI at work without authorization or inappropriately.

Another 46% had uploaded sensitive company information or intellectual property to public AI platforms. Those findings identify potential exposure, not a measured count of breaches.

Rippling set monthly caps in August 2026 after projecting that token spending could reach 40% of its research and development (R&D) headcount budget. Limits and more efficient model routing, which assigns work to suitable AI systems, lowered that projection to 10% to 15% while token use remained near its previous peak.

Chief product officer Matt MacInnis said: "The point isn't to cut AI costs." He added: "We're actually fine with cost growth, provided it's productive." Rippling reported a lower cost projection and sustained usage, but the disclosed figures cover projected cost and consumption, not comparative productivity.

What does this change for employees?

For employees, workplace AI is becoming governed by access, budgets and monitoring rather than remaining an unrestricted utility. Premium models may depend on job needs, individual usage may face a ceiling, and managers may receive consumption data.

For employers, the test is whether those controls improve useful output without increasing cost or exposure. The evidence remains mixed: the 13% value finding measures executive sentiment, while Cisco and JPMorgan cite governance objectives without comparative results. Budgets can ration consumption, but a cap alone does not prove productivity.

The strategic logic is straightforward: preserve AI adoption while preventing cost, waste and security exposure from scaling unchecked. The announcements show organizations moving from blanket access to differentiated permissions, budgets and reporting.

What they do not show is comparable returns on investment (ROI), consistent security outcomes or evidence that tighter limits improve output. That omission matters because token consumption alone cannot settle the productivity case.

Companies still want workers to use AI; the new question is which tools, tasks and spending levels they can justify.