Rippling cut AI token costs by routing work. Now it wants to score employee ROI

The token bill became a headcount-scale expense

Rippling reports that AI-token spending was growing 80 percent month over month after it gave employees broad access and encouraged experimentation. Its internal forecast put token costs on a path to equal 40 percent of the research-and-development headcount budget and approach 90 percent the following year. Concentration was extreme: 10 to 15 percent of employees reportedly generated about 60 percent of spend.

The company says the expensive default was often the problem. Employees used the newest frontier model even when a cheaper model could do the task. Rippling capped tool budgets, centralized vendor data, built a gateway to route prompts, and trained internal AI captains to spread more efficient practices.

Routing appears to have changed cost more than use

Rippling says its projected token burden fell to roughly 10 to 15 percent of the research-and-development headcount budget. TechCrunch reports that July usage reached 600 billion tokens, close to the April peak of 605 billion, while July cost was 37 percent of April cost. The company attributes the difference to model routing and spending controls.

These are company claims associated with a product launch, and independent customers may see different savings. The operational lesson is still credible: model choice, pricing, request routing, and workload design can matter more than raw token volume. An organization buying AI without task-level cost data is not managing an infrastructure expense.

The dashboard can become the manager

AI Spend Console links usage to identity and organizational attributes. Rippling says it can compare spend with pull requests, lines of code, cycle time, performance ratings, and rework requested during code review. Managers can use permissioned dashboards to inspect teams and individuals.

Each metric is incomplete. More code can mean more value, more duplication, or more technical debt. A pull request can be easy or essential. Peer-requested rework may reflect quality, task difficulty, bias, or healthy review. Once a composite score affects access, evaluation, promotion, or dismissal, the system is no longer only a finance tool. It is an automated workplace decision system.

Separate efficiency from employee worth

Companies should govern AI spend, route tasks efficiently, and stop vendors from turning default settings into runaway costs. They should not assume that the easiest available output metric is a reliable measure of a person's contribution. The employee must be able to see the data, understand its purpose, correct errors, and add context before a score carries consequences.

The best result is not maximum token use or minimum token use. It is valuable work at a justified cost, measured with enough humility to recognize what the dashboard cannot see.

• Use spend data first for infrastructure optimization, not individual discipline.
• Validate every productivity proxy against work quality and long-term outcomes.
• Disclose employee-level monitoring and limit data to a defined purpose.
• Require human context and correction before scores affect employment decisions.