Rippling, the HR software company, shipped a product this week called AI Spend Console — and the story behind it is the best public case study of AI cost overrun you'll read this year. In March 2026, an exec meeting almost went sideways: Rippling discovered it was on track to burn 40% of its entire R&D headcount budget on AI tokens, with spend growing 80% month-over-month.
The fix took months and a product team, but the headline is simple: Rippling cut AI token spend from about 40% of headcount budget to roughly 15% — while keeping AI usage at a peak of roughly 600 billion tokens per month. It didn't reduce usage. It routed it, capped it, and made AI leaders responsible for it. This article walks through exactly what happened, why this matters to every company using ChatGPT, Claude, or Cursor, and a 5-step playbook you can copy this week.
How Rippling discovered the AI money pit
Rippling went all-in on tokenmaxxing at the start of 2026 — the term for burning office-wide, frontier-model tokens on everything from code to email. Chief Product Officer Matt MacInnis still recalls the March executive meeting when CFO Adam Swiecicki presented the number that shocked everyone: Rippling was on track to spend 40% of its R&D headcount budget on AI tokens.
That number works out to spending as much on tokens as on 40% of all the compensation it paid to R&D employees. Millions of dollars. And spend was growing 80% month-over-month — if the trend continued, the next year would bring a ~90% token bill relative to R&D salaries.
The result: management immediately launched an "urgent" project to understand the spend and what they were getting for it. The launch ad for the new product literally shows CFO Adam Swiecicki on a stool while employees feed wads of cash into a paper shredder.
What the audit found: 10-15% of people, 60% of spend
When Rippling analyzed its data, it found a pattern that should worry every finance team: roughly 10-15% of its employees were driving about 60% of total AI spend. One engineer alone was spending $50,000 a month, they said. Employees defaulted to the newest, most expensive frontier models for every task — including grammar edits that any cheap model could do.
The company's blog post promises the tool will show "which engineers have high AI spend whose peers frequently ask them to redo work in code reviews" — meaning spend is high, but the output is so weak colleagues redo work. That's the definition of AI waste: cost with negative productivity.
Why AI providers won't fix your spend problem for you
MacInnis was blunt about why Rippling had to build the tool in-house: "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."
That's the part every CFO needs to hear: your enterprise AI vendor bills per token and has no incentive to reduce your billing. Per-user dashboards and cost controls are new and fragmented. The spend problem is genuinely yours.
The fix: caps, gates, and "AI captains"
Rippling didn't cut usage. It:
- Negotiated a max spending cap with each tool: Cursor, OpenAI, and Anthropic.
- Built its own AI gateway that routes every prompt to the best, most cost-effective model for the task.
- Ran dashboards (once called "leaderboards") scoring prompts per day, work output (lines of code / pull requests), and spend.
- Turned its best power users into "AI captains" who assist the rest of the company.
The gateway is the key. Rippling's own benchmarks, shared by CEO Parker Conrad, found SpaceX's Grok was the all-around leader — but "GLM 5.2 is 85% cheaper but [has] nearly identical performance" for coding tasks. Z.ai's GLM 5.2 has become a favorite open-weight model for coding; Databricks has been championing it.
Key Statistics
- 40% of R&D headcount budget was the token burn rate Rippling uncovered (TechCrunch, Aug 7 2026)
- AI token spend was growing 80% month-over-month pre-fix
- 10-15% of employees drove ~60% of AI spend; one engineer spent $50,000/month
- Post-fix: token spend fell to ~15% of headcount budget; July usage hit 600 billion tokens again — 37% of April's cost
- Rippling says GLM 5.2 is 85% cheaper with nearly identical coding results in its own benchmarks
| What happened | March 2026 (scary) | After the fix |
|---|---|---|
| AI token spend vs R&D headcount | ~40% | ~15% |
| Token volume | 605B peak (warning month) | ~600B/month (July) |
| Cost of the same volume | 100% (April baseline) | 37% of April |
| Where models were routed | Frontier model for everything | Cheap models for routine tasks |
The 5-step playbook to fix AI spend this week
Here's the short version you can send to your CTO/CFO today:
- Get a usage report per seat. OpenAI, Anthropic, and Cursor admin panels export per-user usage. Rank users by spend.
- Set a monthly cap per user and team. Most enterprise plans support this in their billing admin.
- Add a routing layer. LiteLLM or Portkey can send cheap tasks to cheap models and keep frontier models for complex ones.
- Benchmark your own workload. Rippling ran its own benchmarks and found a cheaper model for coding. Do the same with 5-10 real prompts from your codebase.
- Re-baseline quarterly. Token prices and model rankings change monthly — rerun step 4 every quarter and retune the router.
The savings are no joke: Rippling itself cut spend from ~40% of budget to ~15%, while usage stayed high. Cost per useful prompt is the metric that matters — not "did people use it."
Frequently Asked Questions
Q: What does "tokenmaxxing" mean?
A: The practice of pouring the most expensive frontier tokens into everyday work — email edits, trivial code reviews, quick searches. It's the modern autopilot of AI: costs rise silently while the tool tries to be helpful for everything.
Q: How much does AI Spend Console cost?
A: It's included for Rippling enterprise HR subscribers, with additional AI-usage-based costs, and can be bought standalone to integrate with another HR system. Full pricing hasn't been publicly detailed.
Q: Should I stop using AI once I see the cost?
A: No. Rippling didn't cut usage — it cut waste. Stop the frontier-model-on-trivial-tasks habit, keep the productive paths, and scale usage where the ROI ledger shows positive.
Q: Do budget caps actually hold against AI spend?
A: Yes, if enforced per user. Rippling combined caps with routing: caps stop drift while routing keeps work flowing. A cap without routing can also cut good usage — do both.
Q: How do I know which models are cheap enough?
A: Run internal benchmarks on your real workloads. GLM 5.2 became a star because companies like Rippling and Databricks tested it on coding prompts and found near-frontier quality at a fraction of the cost.
Q: What routing tools work?
A: Start with LiteLLM (open-source) or Portkey. Both do per-model routing, fallback, and token logging.
Bottom line
Rippling's near-40% blow-up is the rite of passage of enterprise AI — every company that jumped onto Copilot, Claude, or Cursor in 2026 has a version of this bill. The fix is measurement and routing, not abstinence. If your CFO doesn't have an AI token dashboard yet, bookmark this: treat every token as a dollar, cap per seat, and route by task. And when a vendor says they will manage your usage insights for you — remember MacInnis's quote, and build the visibility yourself.
Some links in this article may be affiliated links.