On August 7, 2026, Rippling shipped an AI Spend Console — a product to show companies exactly which models their teams burn tokens on, cap what they can spend, and route every request to the cheapest model that works (TechCrunch; Rippling). The reason it exists is the story: Rippling was the cautionary tale. Its own AI spend was growing about 80% month-over-month, at one point equal to roughly 40% of its R&D headcount budget, and on track to nearly match its entire R&D payroll within a year if nothing changed.

You can't buy an enterprise HR console as a team of one. But the growth rate that scared Rippling is the same one running under your OpenAI and Anthropic invoices — just smaller, for now. This is the drill for keeping it that way.

The number is the news: 80% a month#

Ignore the product for a second and stare at the exponent. 80% month-over-month is roughly 180x a year (1.8¹²). A spend that's a rounding error in January is a payroll-sized line item by December — not because any single day looked alarming, but because compounding doesn't announce itself. This is why AI bills blindside teams that watch their AWS spend like hawks: cloud grows with users, but token spend grows with usage per user — more agents, longer contexts, more retries, more background runs — and each of those quietly multiplies the others.

What it means: budget the exponent, not the balance. If your AI line is up 40–80% this month and you're treating it as noise, you're reading the same chart Rippling read too late. The fix isn't spending less today; it's instrumenting so the curve can't hide.

What the console actually enforces — and the cheap version of each#

Strip the enterprise packaging and the AI Spend Console does four concrete things. Each has a version a solo founder can stand up this week.

1. Attribution. Rippling's console breaks spend down by model, team, and person. Your version: tag every LLM call with a task or customer id and track cost per task, not per call. A per-call number tells you nothing; cost-per-resolved-ticket or cost-per-signup tells you whether a feature is even worth running. We laid out the mechanics in cost per task, not per call: attributing agent spend with Langfuse and OTel.

2. Policy and caps. Administrators can limit token spend and which models people can touch. Your version: hard monthly budget caps per API key, with an automatic kill-switch when a key crosses its limit. A runaway agent in a retry loop can spend a month's budget in a night; a cap turns that from an incident into a log line.

3. Routing. The console sends each request to the most cost-effective model. Your version: a model router that defaults cheap and escalates only on failure — small model first, big model on a low-confidence or failed result. The tradeoffs (and what you give up letting a router pick for you) are in what founders give up to Cursor's auto model routing.

4. Gateway control. The console's real edge over passive dashboards is that it acts — a gateway that shapes usage instead of reporting on it after the money's gone. Your version: one LLM proxy in front of every provider, so attribution, caps, and routing are enforced in a single place you actually control, not scattered across a dozen SDK calls. The rest of the cheap toolkit is in AI cost-control tools for founders.

The strategic read#

The reason Rippling turned its own near-miss into a product is that "who spent what on which model" is becoming a first-class operating metric — as normal to review as CAC or burn. That's the shift to internalize even if you never touch this console. The founders who come out ahead aren't the ones who spend the least on AI; they're the ones who can see what they spend, attribute it to outcomes, and cap the runaways — while the bill is still small enough that instrumenting it is an afternoon, not a reckoning. Rippling built the tool because it waited. The move for a team of one is to not need it.