AI budgets are replacing blanket access at major employers

JPMorgan put a $2,000 monthly ceiling on some engineers’ use of Anthropic’s Claude Code in September. The bank says it is tying those costs to productivity, quality, risk reduction and business impact—a sharper test than the open-ended access many companies used to drive AI adoption.

Goldman Sachs, Cisco and JPMorgan are among the employers replacing blanket permissions with a targeted system of budgets, usage tracking and limits on which employees can use which tools. They still want employees to use AI, but they no longer treat every model, task and user as equivalent.

The aim is not simply to cut spending. Companies are pouring ever-larger sums into AI while trying to make each deployment more secure and productive. Lari Hämäläinen, a senior partner in McKinsey & Company’s technology, media and telecommunications practice, says employers see usage picking up and therefore need guardrails.

Who needs frontier models?

Tokens are units of text and code processed by AI models. A token budget limits how much a person or team can consume; a usage dashboard aggregates that activity so managers can track adoption.

At Goldman Sachs, CEO and chairman David Solomon said this month that the bank is rapidly expanding the range of AI tools available to its 45,000 employees.

David Solomon told Bloomberg TV: “We need frontier models for certain things, particularly cyber.” For many other tasks, “there are a lot of things where we don’t need the frontier,” he said, to help employees be more productive. Frontier models are the most advanced AI systems.

At Cisco, Chief Product Officer Jeetu Patel said last month on the CXOTalk podcast that the company first gave coding teams months of unlimited tokens, comparing that approach to learning to ride a bicycle. Managers across his 32,000-person team will now receive token budgets.

Companies are reserving their most capable models for the work that needs them.

How did token use become a cost problem?

The turn toward limits followed a period in which employers reported explosive growth in employee AI consumption.

• Visa said it doubled its monthly token use from 1 trillion in February to nearly 2 trillion in March. The company also gave prizes to its AI power users.
• In March, Nvidia said it was trying to spend as much as $2 billion annually on tokens for its engineers. CEO Jensen Huang said he would be deeply alarmed if a $500,000 engineer used less than $250,000 worth.
• Disney’s dashboard, reported in April, showed that the company’s biggest AI users made hundreds of requests per day on average and consumed tens of millions of tokens.
• KPMG US said in May that it hoped a dashboard for its roughly 10,000-person advisory business would encourage more frequent and sophisticated AI use. It later stopped monitoring usage through that dashboard because employees had incorporated AI into everyday work.

Those dashboards made activity visible and often encouraged more use. But KPMG’s decision to stop monitoring shows that adoption can outrun the need for a scoreboard.

Menlo Ventures’ new study found that enterprise AI deployment spending reached $177 billion by September, up from $33 billion in all of 2025.

Only 13% of roughly 2,000 technology executives polled by Gartner from May through July said their organizations had gained significant value from AI.

Alex Sobol, cofounder of the Millennium Alliance, an invite-only community for C-suite executives, argues that workers have little incentive to curb usage without limits. They will use AI in every way they can, he says.

Rapid token growth turned AI from an open experiment into a budget line.

Can limits preserve productivity?

Governance now also functions as a security control.

A 2025 KPMG survey of more than 1,000 US workers found that 44% were using AI at work without authorization or in inappropriate ways. Another 46% had uploaded sensitive company information or intellectual property to public AI platforms.

Rippling, a workforce-management software startup, set monthly caps on AI tools in August after projecting that token spending could reach 40% of its R&D headcount budget. The company said those caps and more efficient model routing lowered the projection to between 10% and 15% while token use stayed near its previous peak.

Matt MacInnis, Rippling’s chief product officer, put the intent plainly: “The point isn’t to cut AI costs. We’re actually fine with cost growth, provided it’s productive.”

Taken together, the examples favor conditional governance rather than a universal cap. Cisco let engineers build fluency before restricting use; JPMorgan links cost to outcomes; Rippling combines limits with model routing, which means directing work to different models.

The approach fails when token volume becomes a proxy for value or broad access outpaces controls on sensitive data.

For product managers, the next step is to classify jobs by required model, set budgets by role and pair usage data with quality, risk and business-impact measures.

The practical takeaway is simple: manage AI access as part of the product, and let measured outcomes—not raw usage—set the limits.

For product teams, that means the model choice, token budget and risk controls will increasingly be managed as part of each AI workflow, rather than left to employees and finance after launch.