After Rippling blew millions on AI in months, it built an employee ROI tool

Rippling didn't set out to build a cost-control product. It set out to tokenmaxx like every other venture-backed software company in early 2026, and the bill nearly sank the ship. Eight months later, the company is shipping AI Spend Console — a dashboard that maps exactly which employees, teams, and roles are lighting money on fire and whether the smoke smells like productivity or slop.

The origin story is the product. In March, CFO Adam Swiecicki walked into an executive meeting and dropped a number that made the room go quiet. Rippling was on pace to spend 40% of its entire R&D headcount budget on model inference. Not 40% of the AI budget. Forty percent of what it pays engineers. Millions of dollars, growing 80% month over month. Project that curve forward twelve months and the token tab hits 90% of R&D compensation. The company would effectively employ a shadow workforce made of API calls.

"We were incredulous," Chief Product Officer Matt MacInnis told TechCrunch. The marketing asset for the launch features Swiecicki on a stool watching employees feed cash into a paper shredder. Subtle it isn't.

The audit that followed revealed the usual suspects. Roughly 10 to 15% of the staff drove 60% of the spend. One engineer — one — ran a $50,000 monthly tab. The pattern was ugly but familiar: developers defaulted to the newest, most expensive frontier models for every task, from writing a regex to refactoring a service. Why wouldn't they? The inference providers — Anthropic, OpenAI, the rest — have zero incentive to surface usage data or suggest cheaper alternatives. Their business model is the runaway train. Rippling negotiated hard caps with Cursor, OpenAI, and Anthropic. The caps held, but the underlying disease remained: no visibility, no cross-vendor coordination, no pressure to right-size the model to the job.

AI Spend Console is Rippling's answer to that vacuum. It ingests usage across providers, tags spend to the human who triggered it, and then asks the uncomfortable question: did the output actually ship? The console flags engineers whose AI bills run hot while peers routinely ask them to redo work in code reviews. That signal — high spend, low trust — is the killer metric. It turns a finance problem into a management problem.

Here's where the editorial sharpening happens. The industry spent the first half of 2026 pretending that "more intelligence per token" was a roadmap. It isn't. It's a pricing tier. Rippling's own benchmarks, run for its own workloads, found SpaceX's Grok leading the pack — but also found GLM 5.2 delivering nearly identical results at 85% less cost. Eighty-five percent. That's not a rounding error. That's a structural indictment of the frontier-model default.

Enterprises are finally figuring this out. The smart ones now run a model router: a cheap open-weight model for the 80% of tasks that don't need reasoning firepower, a mid-tier for the rest, and a frontier model only when the benchmark says the gap matters. Chinese-origin open weights have entered the chat not for geopolitics but for unit economics. Rippling's tool makes that routing visible and auditable.

Skepticism is warranted on two fronts. First, the console measures what it can see — token counts, code-review friction, ticket velocity. It cannot measure the developer who uses AI to explore a design space and then writes the correct solution from scratch. That work looks like zero AI spend. The tool will punish explorers if managers treat the dashboard as gospel. Second, Rippling is now a vendor selling the cure it bought for its own sickness. The conflict is obvious. The product might be good; the incentive to upsell seat licenses is real.

But the bigger truth is uncomfortable for the AI labs. They built a business on opacity. They sold "intelligence" as a monolith and priced it like heroin. Rippling's console, and the wave of similar tooling coming behind it, shatters that opacity. When a CFO can see that 12% of engineers consume 60% of the AI budget while producing rework, the conversation shifts from "which model" to "which human." The labs will hate this. They should.

The tokenmaxxing era isn't over. It's just moving from the credit-card phase to the procurement phase. Rippling built the spreadsheet that makes procurement possible. Every company burning six or seven figures on inference now needs one. The ones that don't adopt it will keep feeding the shredder.