Key takeaways
- Rippling reduced AI token spending from 40% to 15% of its R&D budget in eight months by routing requests to cheaper models instead of stopping usage.
- GLM 5.2 costs 85% less than frontier models while delivering nearly identical performance, making model selection the primary lever for cost control.
- Early analysis showed 10-15% of employees drove 60% of AI spending, with some engineers spending $50,000 monthly on tokens using expensive defaults.
- Measuring AI productivity remains unsolved outside engineering; without clear productivity signals, companies may restrict broader employee access to AI tools.
HR software company Rippling this week launched AI Spend Console, a tool designed to curb employee AI token spending without eliminating access to generative models entirely. The product identifies which employees, teams, and departments burn the most AI budget and correlates their spending patterns with actual productivity — measuring, for instance, whether high-spending engineers receive more code-review requests to redo their work.
The crisis that sparked a solution
Rippling’s journey to building this product began with a shock. In March, CFO Adam Swiecicki presented executive leadership with an analysis that revealed the company was on track to spend 40% of its entire R&D headcount budget on AI tokens — the same dollar amount the company paid salaries to engineers across the unit. That figure represented millions of dollars.
The spending trajectory was unsustainable. Month-over-month growth hit 80%, meaning if the trend persisted, Rippling would eventually spend 90% of its R&D compensation budget on tokens alone. The discovery triggered what Chief Product Officer Matt MacInnis called an “urgent” internal project to understand what the company was getting for that expenditure.
How widely was the spending distributed?
When Rippling dug into the numbers, the distribution was stark. Roughly 10 to 15% of the company’s employees accounted for approximately 60% of total AI spending. At the extreme end, one individual engineer was spending $50,000 per month on tokens. This concentration suggested that either a small subset of workers had discovered vastly more valuable use cases, or the company lacked visibility into and controls over token consumption.
The root problem: expensive defaults
Rippling’s analysis identified the core issue — most employees were defaulting to the newest, most expensive frontier models from providers like OpenAI, Anthropic, and Cursor for nearly all tasks, regardless of complexity. A simple text rewrite or grammar check was running on the same large models deployed for complex code generation or research.
MacInnis attributed part of the blame to the incentive structure embedded in the AI industry itself. “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. They don’t provide you with great usage insight, and they don’t collaborate with one another,” he said.
The cost advantage of alternative models
Rippling founder and CEO Parker Conrad had recently completed internal benchmarking that revealed an important finding: SpaceX’s Grok performed as the strongest overall model across tests, but Z.ai’s GLM 5.2 delivered nearly identical performance at 85% lower cost. That price gap meant that routing simpler tasks to GLM 5.2 instead of frontier models could dramatically reduce expenses while maintaining output quality.
Other tech companies had reached similar conclusions by mid-2026. GLM 5.2 had become a favored choice among engineering teams for coding tasks, with Databricks also championing its adoption.
Building an intelligent AI gateway
To solve the problem, Rippling constructed its own AI gateway — infrastructure that intercepts token requests and automatically routes them to the most cost-effective model suitable for each task. Rather than stopping employees from using AI, the company instead channeled them toward smarter choices.
The AI Spend Console dashboard presents the routing intelligence visually, scoring activities according to prompts per day, work output (measured in lines of code and pull requests submitted), and associated spend. This combination reveals which employees have high token consumption but low productivity — what the company described as workers whose peers “frequently ask them to redo work in code reviews.”
Results that validated the approach
The impact was immediate. By July, Rippling’s token consumption had climbed back to 600 billion tokens, nearly matching the peak month in April when the CFO issued his warning — 605 billion tokens. However, the cost structure had fundamentally changed. The July token bill represented just 37% of April’s cost for nearly identical usage volume.
Overall, Rippling reduced token spending from 40% of its R&D headcount budget down to approximately 15%, a reduction of roughly 63%. MacInnis joked that the system prevented the sales team from “doing grammar updates using Fable,” underscoring how the gateway actively steered simple tasks away from expensive frontier models.
Beyond engineering: measuring productivity across functions
Rippling’s experience also surfaced a harder problem that extends beyond token routing. The company identified high-performing employees using AI effectively and designated them as “AI captains” tasked with helping the broader organization adopt tools wisely. But this informal approach had limits.
Software engineers, by nature, have clearer productivity metrics — pull requests, lines of code, code review feedback. Other parts of the organization proved much harder to measure. Rippling is now working with customer onboarding teams to automate mailing data and data-reconciliation tasks, but measuring the productivity gains requires different metrics: more customers onboarded, faster processing, fewer errors.
The measurement challenge
MacInnis emphasized this constraint when discussing AI access beyond the engineering function. “We have to be able to link token consumption in G&A functions and in customer-facing functions back to productivity. If we can’t do that, all bets are off on any of this stuff being available to the broader employee base,” he said.
The implication is stark: without clear productivity signals, companies may choose to restrict AI access rather than leave spending unchecked. Rippling’s own experience suggests that enterprise AI adoption may not follow the frictionless adoption curve of tools like email or Slack if companies cannot demonstrably tie token consumption to business outcomes.
Product availability and integration
Rippling is offering AI Spend Console as an included feature for existing HR subscribers, though additional charges apply based on AI usage. The tool can also be purchased as a standalone product and integrated into other HR systems of record. The gateway component, however, remains proprietary to Rippling’s platform — companies using alternative gateways can still purchase the dashboard and spending analytics, but would need to adopt Rippling’s gateway to access the automatic routing and cost-control features.
Frequently Asked Questions
How much was Rippling spending on AI tokens before controlling costs?
Rippling was on track to burn 40% of its R&D headcount budget on AI tokens, representing millions of dollars. One engineer alone was spending $50,000 per month, and overall spending was growing 80% month-over-month.
How did Rippling reduce AI costs?
Rippling built an AI gateway that automatically routes requests to the most cost-effective model for each task. It discovered GLM 5.2 costs 85% less than frontier models like those from OpenAI and Anthropic while delivering nearly identical performance.
What does the AI Spend Console measure?
The dashboard tracks AI spending by employees, teams, and roles while correlating token consumption with productivity metrics like lines of code and pull requests, identifying high-spending workers whose output quality may be questioned by peers.
