Rippling built a tool to tame its runaway AI bill

Rippling’s AI bill was growing so quickly that, by March, the HR software company was on track to spend an amount equal to 40% of its entire R&D headcount budget on AI tokens. Spending was rising 80% month over month. If that pace had continued, AI tokens would have consumed nearly 90% of the cost of its R&D employees within a year, according to TechCrunch.

The result is AI Spend Console, a new product Rippling says can track not only how much AI employees use, but whether that spending produces useful work. The tool maps costs across individual employees, teams, and roles, then compares usage with output such as lines of code and pull requests.

“Roughly 10–15% of our employees were driving about 60% of total AI spend. One engineer was spending $50,000 a month.”

How Rippling cut token spending

The product grew out of Rippling’s internal response to what the company calls “tokenmaxxing”: aggressively using AI tools without much attention to cost or results. CFO Adam Swiecicki presented the spending figures to executives in March, prompting what Chief Product Officer Matt MacInnis described as an “urgent” effort to understand where the money was going.

Recommended reading

AI infrastructure planning can no longer wait

Rippling did not ban AI tools. Instead, it negotiated maximum spending caps with Cursor, OpenAI, and Anthropic, and built an internal AI gateway to route requests between models. The gateway is included in AI Spend Console, although companies can continue using a different gateway if they do not need Rippling’s spending-governance features.

The routing approach addresses a basic purchasing problem: employees often default to the newest and most expensive frontier model, even when a cheaper system can handle the task. Rippling CEO Parker Conrad said the company’s internal benchmarks found SpaceX’s Grok to be the strongest all-around model, while Z.ai’s GLM 5.2 cost 85% less with nearly identical performance for Rippling’s use cases. The company has since used a mix of models at different price points, including open-weight models.

MacInnis criticized the incentives of inference providers:

“The truth is that the inference providers, like Anthropic and OpenAI, have absolutely no incentives to help you control your spend. They have every incentive for it to be a runaway expense, and that’s exactly what they do.”

Rippling says the strategy reduced AI spending from 40% of its headcount budget to about 15% without reducing usage. The company consumed a peak of 605 billion tokens in the month when its CFO raised the alarm. In July, usage reached 600 billion tokens again, but the cost was only 37% of April’s cost, MacInnis said, because more requests were routed to cheaper or more effective models.

Measuring productivity, not just usage

AI Spend Console’s dashboards combine metrics such as prompts per day, spending, and work output. One example in Rippling’s product description identifies engineers with high AI costs whose peers frequently ask them to redo work during code reviews. That is a more pointed measure than a simple usage leaderboard: it attempts to connect model consumption to the quality of the resulting work.

Rippling has also designated effective users as “AI captains” who help other employees use the tools. Engineering remains the primary adoption area, but the company is testing AI for customer-onboarding teams, including mailing-data automation and data reconciliation. In those roles, the intended productivity measure would be the number of customers onboarded.

That broader rollout exposes the product’s unresolved challenge. Rippling says it must connect token consumption in general and administrative and customer-facing functions to measurable productivity before AI access can be expanded across the employee base. The company did not provide a methodology for its internal model benchmarks, and the reporting does not independently verify either the performance comparison or the claimed savings.

AI Spend Console is included with Rippling’s HR subscriptions, but it carries additional AI usage-based costs. It is also available as a standalone product that can integrate with another HR system of record. Rippling did not disclose standalone pricing, leaving the cost of adopting the tool unclear for companies outside its existing customer base.

The numbers make the case for the product’s core mechanism: model routing and spending controls produced a large cost reduction while token volume stayed almost flat. But the harder promise—proving that AI makes non-engineering employees more productive—has not yet been demonstrated. For now, Rippling has a credible answer to runaway inference costs, not a finished system for measuring AI’s return on every employee.