Companies Are Slamming the Brakes on AI Token Use

Overview

Companies are “slamming the brakes” on internal AI usage because AI—especially token-based and agentic workloads—is becoming too expensive to run at the scale businesses initially promised. The central tension is between early AI enthusiasm (including internal incentives to drive usage) and the CFO reality of ballooning inference and token costs.

Key Themes from the Coverage

1) AI usage is being throttled internally due to cost

• Multiple companies reportedly limit employees’ ability to use AI tools because token spending/inference costs are “going bananas.”
• The coverage (including references to reporting by 404 Media) points to named examples of the “real” cost behind AI adoption, such as Amazon, Adobe, Atlassian, Citi, and others.

2) “Token leaderboards” and incentives are being reversed

• A year earlier, executives and CEOs promoted AI adoption publicly to attract investors.
• Internally, some firms encouraged adoption through token leaderboards, recognition programs, and even bonuses tied to AI usage.
• In Q2, as costs spiked, CFOs pushed a major shift—moving from “letting every flower bloom” to a “lawnmower” approach (dramatically cutting back usage).

3) Monitoring and spend controls are increasing

• Meta is cited as adding AI Gateway to monitor employee AI token usage and trigger automated alerts for unusual spending spikes—indicating tighter governance rather than broader access.

4) AI agents are a major driver of runaway costs

• Speakers note that AI agents burn tokens continuously in the background, quickly exhausting budgets.
• This raises skepticism about “agentic” systems that run unattended without strict cost controls.
• An example is mentioned where a team allegedly burned through inference costs within months.

5) Outcomes-based pricing is replacing simple token/subscription models

• Vendor pricing is shifting from:

  • flat subscriptions, or
  • simple token allowances toward hybrid models combining token budgets + workflow counts.

• Customers increasingly push back, demanding measurable, quantifiable results, moving toward “pay for outcomes.”

6) More cost-saving infrastructure is expected in the second half of the year

Anticipated trends include:

• Local inference (running models on devices for near-zero marginal cost)
• Small/domain-specific models for common tasks
• Model routing (“routers”) that use expensive cloud models only when necessary, otherwise relying on local or smaller models

The underlying argument: stop “cutting down daisies with chainsaws” (overusing large cloud models for simple requests).

7) Re-evaluation of layoffs/hype vs real automation

• Gartner is referenced as suggesting many companies that claimed layoffs “because of AI” will later regret it and may need to rehire—because genuine automation/workflows weren’t ready.
• A nuance is offered around “jobs and AI”: some “reductions” may involve not filling open roles rather than true layoffs, and some work may shift instead of disappearing.

8) Product design analogy: make AI power more “civilian-friendly”

• One analogy compares AI adoption to consumer software pricing (e.g., MoviePass/Movie Unlimited)—arguing vendors must make AI usage more predictable and manageable.
• Another analogy compares AI tools to Excel vs Google Sheets:
  • start simple for entry-level users,
  • allow “power” to build for advanced users,
  • avoid forcing users into complex, cloud-heavy systems immediately.

Presenters / Contributors Mentioned

• Jason
• Richard Campbell
• Lisa (referenced only by first name)
• Gartner (organization mentioned, not a specific person)