Rippling burned 40% of R&D payroll on AI tokens. Then it built a meter

Here is the number that should stop you: in March 2026, Rippling worked out it was on track to spend as much on AI tokens as it paid every engineer in its R&D division. Roughly 40% of the R&D headcount budget, growing 80% month over month. On August 7 the company shipped AI Spend Console, the tool it built to fix its own problem — and the story of how it got there is more useful than the product.

What actually happened

Per TechCrunch's reporting and Rippling's own announcement, the discovery came in a March meeting. Spend was compounding at 80% month over month. Left alone for a year, the token bill would have matched the entire R&D payroll.

The distribution was worse than the total. One engineer was spending $50,000 a month. Roughly 10–15% of employees drove about 60% of all AI spend. Token volume peaked at around 605 billion in April.

AI Spend Console is two things bolted together: a gateway that routes prompts to a cheaper model when the task doesn't need a frontier one, and a dashboard that ties spend per person and per team to output signals from GitHub and Salesforce — pull requests, code velocity, revenue contributed. It also flags people producing work that needs rework, which is the awkward half of the pitch.

The result Rippling reports: token spend down from ~40% of R&D headcount budget to ~15%. July usage hit roughly 600 billion tokens — about the April peak — at 37% of April's cost. Same work, a third of the bill. It ships with Rippling HR subscriptions and as a standalone product for companies on other HR systems.

CEO Matt MacInnis put the incentive plainly: the inference providers, he said, have every reason for this to be a runaway expense.

Why AI spend control matters for your business

You are not Rippling. You do not have a 605-billion-token month. You have a $400 Claude bill, a $180 Cursor bill, three ChatGPT Team seats, and an n8n workflow calling an API you set up in February and haven't looked at since. Nobody is adding those up, and no line item on your P&L says "tokens."

Three moves, in order.

Put every AI call behind one key you control. Not per-tool vendor billing. A gateway — LiteLLM, OpenRouter, or your own thin proxy — so spend, model choice, and rate limits are yours to change without touching application code.

Route by task, not by habit. Most of what people burn frontier tokens on is classification, extraction, and summarization. Those run fine on a cheap model. Rippling's 63% cost cut at flat volume is almost entirely this.

Tie spend to a shipped thing. Cost per resolved ticket, per drafted quote, per processed invoice. A number with no denominator is not a metric — it's just a bill you feel bad about.

Sources: TechCrunch, Rippling.