The End of Tokenmaxxing: Rippling's Wake-Up Call on AI Unit Economics
AI-generated image · Bay Street Wire
After discovering AI token costs were on track to rival nearly half of its R&D payroll, Rippling is pivoting from raw compute growth to a brutal audit of employee productivity.
For the first half of 2026, the enterprise AI playbook was simple: 'tokenmaxxing.' The goal was raw compute and maximum integration. But as the novelty fades, the industry is hitting a hangover where the focus has shifted from what AI can do to what it actually costs per unit of output.
Nowhere is this shift more evident than at HR software provider Rippling. As TechCrunch first reported, the company recently unveiled 'AI Spend Console,' a tool designed specifically to track and contain runaway AI expenses. The product was born from a moment of internal crisis. In March, CFO Adam Swiecicki presented data showing that Rippling was on track to spend as much on AI tokens as 40% of the total compensation paid to its R&D headcount budget.
Chief Product Officer Matt MacInnis told TechCrunch that the executive team was 'incredulous' to find spending growing at a rate of 80% month-over-month. Had the trend continued, the company projected that AI token costs would eventually consume 90% of the R&D unit's employee spend. The root cause was a lack of discipline: employees were defaulting to the most expensive frontier models for every task, regardless of complexity.
**Opinion:** *The Rippling experience exposes a fundamental misalignment in the current AI ecosystem. As MacInnis noted, inference providers like OpenAI and Anthropic have no incentive to help customers control costs; their business models thrive on runaway expenses. The 'AI hangover' is the realization that the burden of efficiency falls entirely on the enterprise, not the provider.*
When Rippling audited its usage, the results were stark. The company's blog post revealed that roughly 10% to 15% of employees were responsible for about 60% of all AI spending, with one single engineer costing the company $50,000 a month. More concerning was the quality of the output; the new AI Spend Console is designed to identify 'AI slop' by flagging engineers with high spend whose peers frequently request they redo work during code reviews.
To solve this, Rippling moved toward a multi-model strategy. CEO Parker Conrad noted that internal benchmarks showed SpaceX's Grok as an all-around leader, but Z.ai's GLM 5.2—a model also championed by Databricks—provided nearly identical performance for coding tasks at 85% of the cost.
By implementing an AI gateway to route prompts to the most cost-effective model, Rippling managed to maintain its volume while slashing costs. MacInnis shared that while internal usage hit 600 billion tokens in July—similar to the peak in April—the cost of that July spend was only 37% of the April cost. Through these optimizations, Rippling reduced its token expenditures from 40% of its headcount budget to roughly 15%.
However, the company warns that technical routing isn't a total cure. Rippling is now attempting to link token consumption in G&A and customer-facing functions to tangible productivity metrics, such as the speed of customer onboarding. MacInnis warned that if this productivity cannot be measured, the era of open AI access may end, suggesting that AI access might no longer be a universal utility like email or Slack.