Rippling Was About to Spend 40% of Its Engineering Payroll on Tokens

Key Takeaways

• Rippling's AI token spend was growing 80% month over month, putting it on a forecast path to spend 40% of its R&D headcount budget on tokens, and roughly 90% the year after.

• Roughly 10 to 15% of employees drove about 60% of total AI spend. One engineer was spending $50,000 a month.

• The most useful finding is the one getting the least attention: the productivity gain came from initial adoption, not from spending more on frontier models. The curve is not linear.

• Rippling brought the forecast down to 10 to 15% of headcount budget while token volume stayed flat, roughly 600 billion tokens in July against a 605 billion peak, at 37% of April's cost. That is routing, not rationing.

• The governance layer everyone is quoting is not shipped. Rippling's own product section says spend control is coming "soon" and the AI Gateway sits behind a waitlist.

• Scoring individual engineers on lines of code and PR volume is the part I would not copy. Measure the workflow, cap the spend, and leave the leaderboard alone.

A finance team walks into a March exec meeting with a number, and the number is that your company is on track to spend as much on AI tokens as it pays 40% of its entire engineering organisation. Not 40% of the tooling budget. 40% of the payroll for the people.

That happened at Rippling this year, and on August 6 the company published the whole postmortem alongside the product it built to fix the problem. TechCrunch picked it up the next day. Both pieces are worth reading. Neither one, I think, lands on the finding that actually matters.

I've shipped 126 production systems and every autonomous one runs behind a hard cost ceiling, because I learned early that an agent with an open budget will find a way to use it. So this one is personal. Here is what the numbers say, where the two accounts quietly disagree, and the one design decision in Rippling's fix that I would not copy into a client's stack.

What exactly did Rippling find when it audited its AI spend?

Four things, and they are more specific than most vendor postmortems get.

Spending was concentrated: roughly 10 to 15% of employees drove about 60% of the total bill, with one engineer at $50,000 a month. Defaults were expensive, because the newest models were set to fast mode and nobody had ever gone in and decided otherwise. The productivity signal was real but non-linear. And the trajectory was, in the company's own word, unsustainable, with token spend compounding at 80% month over month.

Rippling's own launch ad features its CFO on a stool while employees feed cash into a shredder. The company is not being subtle about the size of the number that started this.

That third finding is the one I would put on a wall. Engineers using Cursor heavily shipped more pull requests, but the biggest gains came from adopting the tools at all, not from buying more expensive inference on top. If that holds outside Rippling, and my own experience says it does, then most of the marginal dollar in a runaway AI bill is buying nothing.

Why did 10 to 15% of employees drive 60% of the bill?

Because nobody set a default and every incentive pointed up. Rippling turned AI tools on for everyone with no limits, encouraged experimentation, ran hackweeks. That is a reasonable way to start. It is a terrible way to continue.

The mechanism is worth naming precisely, because Rippling names it well: when spend is unlimited, picking the cheapest model that can do the job is pure cognitive load with no payoff. So people default to the most expensive model for everything, including tasks a small model would finish for a fraction of a cent. The company's line is that constraints made engineers smarter, not less capable. Once there was a budget, people actually learned which model was good at what.

I ran into the same thing at a much smaller scale, and the fix was never a policy document. It was a hard ceiling. One of my own tools, a backlink tracker that calls a paid scraping API on a schedule, has a $15 a month cap, a $2 a day circuit breaker across every paid service it touches, and a kill file on disk that stops it dead. It fails closed if it cannot read its own spend ledger, because a budget check that passes when it breaks is not a budget check. It currently runs at about $5 a month. The caps did not slow anything down. They forced a routing decision that should have been made on day one.

This is the same failure pattern I wrote about when the Army promised its staff unlimited AI tokens and ran dry in weeks. Different organisation, identical physics. "Unlimited" is not a pricing model, it is a deferred decision.

Does spending more on AI actually make engineers more productive?

Not proportionally, and that is the finding with the most money in it. Rippling's data says the biggest productivity improvements came from initial adoption rather than from heavier spend on frontier models. So the curve flattens fast.

Look at what happened after the intervention. Rippling cut its forecast from 40% of R&D headcount budget down to 10 to 15%, and it did that without cutting usage. The company spent a peak of 605 billion tokens in the month the CFO raised the alarm. In July, internal usage hit 600 billion tokens again. Same volume. But per the chief product officer, July's token spend cost 37% of April's.

Read that again, because it is the whole argument in one comparison. Roughly the same tokens, a bit over a third of the cost. Nothing was rationed. The work just stopped being routed to the most expensive model available by reflex.

That is a routing problem wearing a budget problem's clothes, and it is why I keep telling people that a price cut at the frontier will not move your agent bill on its own. Your bill is mostly a function of which model handles which task, and almost nobody has made that mapping explicit. If you want the decision framework rather than the anecdote, I went through it in detail in how I actually choose between OpenAI and Claude across 126 shipped systems.

What is AI spend management, and what does the tooling actually do?

AI spend management is the practice of tying every token your company burns to a person, a team, and a business outcome, then governing which models they are allowed to reach. It is FinOps with an identity join bolted on.

Rippling's implementation has four parts, and the sequence matters more than the product.

First, ingestion: pull spend from Cursor, OpenAI and Anthropic into one place, then automatically map external identifiers to employee records, so a GitHub username resolves to a work email and from there to a team, role and department. Second, dashboards on top of that joined data: weekly spend, model usage by team, spend against merged PRs, spend against code rework. Third, an internal gateway that all LLM traffic passes through, which is where limits and model access policies get enforced and every transaction gets logged. Fourth, a human layer, 20 nominated "AI captains" who own impact, adoption and governance for their org.

The product page states the promise in the present tense: see costs, tie token usage to business outcomes, and govern the use of approved LLMs. Hold that thought.

The gateway is the piece worth stealing regardless of whether you ever touch this product. A routing layer between your people and your models gives you one place to set limits, one audit log, and the ability to add or swap a model without every team retooling. It also breaks vendor lock-in, which Rippling is refreshingly blunt about wanting.

Layer	Ungoverned setup	What Rippling built
Cost visibility	Finance manually collates vendor dashboards	Vendor spend ingested and joined to employee records
Attribution	Total spend only, no team or role breakdown	Spend by team, role, department and individual
Model choice	Newest model on fast mode, by default	Requests routed per task through an internal gateway
Limits	None, then a blunt per-tool dollar cap	Policy-based access tied to employee attributes
Outcome link	None	Spend mapped to merged PRs, rework and ratings

Where do the two accounts disagree?

In three places, and one of them changes what you should conclude.

The first is small but worth getting right. TechCrunch writes that Rippling "dropped its token spend from 40% of its headcount budget to about 15%." Rippling's own post says it went "from a forecast of spending 40% of our headcount budget on tokens to 10 to 15%." The 40% was a projection, not a bill the company actually paid. That distinction matters if you are about to quote the number in your own board deck.

The second is the endpoint: "about 15%" against "10 to 15%". Minor, but it rounds in the vendor's less flattering direction, which is at least an honest direction for a journalist to round.

The third is the one that matters. TechCrunch describes the gateway as built and shipped: "So it built its own AI gateway that is also part of this product." Rippling's own page is more careful. The product section says CFOs and CTOs can see spend, connect it to outcomes, "and, soon, govern the use of approved LLMs." The gateway carries an asterisk. At the bottom of the post there is a link to join the AI Gateway waitlist.

What did everyone miss about the AI Spend Console launch?

That the control half of "see and control your AI spend" is not generally available yet. Rippling's own launch post opens by saying the product "governs the use of approved LLMs." Further down, the same post says CFOs and CTOs get visibility, outcome mapping, "and, soon, govern the use of approved LLMs." Present tense in the intro, soon in the product section, one article.

I do not think that is a scandal. Shipping visibility first and enforcement second is a defensible order, and it is honestly how I would sequence it too, because you cannot write a sane routing policy before you know who is spending what. But it does change the buying decision. What you can purchase today, on a 30 day trial with no Rippling subscription required, is a very good attribution dashboard. The chief product officer credits the drop to 37% of April, a 63% cut, to routing requests to more effective models, and routing is precisely what the waitlisted gateway does.

Two smaller things worth pulling out of the launch, one from each account. Rippling's post introduces the AI Scorecard as highlighting "four key metrics" and then lists five. And the quote the chief product officer gave TechCrunch about vendor incentives is sharper than the attention it got: the inference providers "have absolutely no incentives to help you control your spend... They don't provide you with great usage insight, and they don't collaborate with one another." That is a vendor of AI tooling saying the quiet part about the layer underneath it, and it is the strongest argument in the whole launch for owning your own gateway.

Should you score individual employees on AI spend?

This is where I get off the bus. Measure the spend, absolutely. Attribute it to teams so you can find the outlier workflows. But Rippling's AI Scorecard scores individual engineers across adoption, usage, productivity, cycle time and efficiency, and the productivity component is explicitly "output (PRs, lines of code)."

The scorecard in Rippling's own words. Productivity is scored on pull requests and lines of code, which is the metric the industry spent twenty years learning not to trust.

Lines of code as a productivity measure has been discredited for decades, and generative models make it worse, not better, because producing more code is now nearly free. A per-employee leaderboard built on that number rewards exactly the behaviour the tool was built to stop. Rippling clearly knows this, which is why rework rate sits alongside it. In Rippling's own post that capability appears as one of the questions finance could not answer: "Which engineers have high AI spend, whose peers frequently ask them to redo work in code reviews?" TechCrunch recast the same line as a product promise. It is the promise, not the question, that ends up on a manager's screen. That is a real signal. It is also, pointed at a named individual and connected to performance ratings, a management instrument, not an efficiency one.

The honest version of the risk: the same dashboard that finds a $50,000 a month outlier can find an engineer whose PR count dipped during a quarter spent on a migration nobody can see in GitHub. One of those is a cost finding. The other is a career event. The tool cannot tell them apart, and the CFO reading it will not either.

What I tell clients is simple. Cap at the workflow, report at the team, and never build the individual leaderboard, because the moment a number is attached to a person's name it stops measuring behaviour and starts producing it. That is not a soft principle, it is the same reason I scope agent projects around a process rather than a headcount.

What should you actually do this month?

You do not need to buy anything to get most of the benefit here. Four moves, in order.

Go and get the number first. Pull last month's actual spend from every AI vendor you use and divide it by whatever unit of work it is supposed to produce. Most teams I talk to have never done this once, and the answer is usually either reassuring or genuinely alarming, with very little in between.

Then set a hard cap per tool per month, exactly as Rippling did as a stopgap. It is blunt, it will annoy someone, and it buys you the time to design something better. Layer a daily circuit breaker underneath the monthly one, and make the system fail closed when it cannot read its own spend ledger, because a budget check that silently passes when it breaks is not a budget check.

Third, make the routing explicit. Write down which class of task gets which model. Cheap model for classification, formatting, extraction and retries. Frontier model for the reasoning step that actually needs it. This one decision is where the 63% came from at Rippling, and it costs nothing but an afternoon.

Fourth, only then think about a gateway. If all your traffic already flows through one service, you may already have the choke point you need. If it is scattered across a dozen SDK calls in a dozen repos, the gateway is worth building before the dashboard, not after. For a broader map of where this sits in the stack, my honest guide to what agentic AI actually is covers the vocabulary without the vendor gloss, and the 2026 updates cheat sheet tracks what has changed since.

If you would rather find out where your own setup leaks before you spend another month guessing, the AI readiness assessment walks the same ground in about five minutes and tells you which layer to fix first.

What this launch really signals

Early 2026 was the tokenmaxxing era, and it is over. The interesting shift is not that companies are cutting AI budgets, because Rippling did not cut usage at all. It is that AI spend is graduating from an experiment line item into something with a cost centre, an owner, and a review cadence, the way cloud spend did around 2015.

The part I would watch, and the part TechCrunch flags at the very end, is the access question. If a company cannot connect token consumption to productivity in a function, the chief product officer's own framing is that "all bets are off" on giving that function access at all. Employee AI access stops being like email and starts being like a seat licence you have to justify. That is a much bigger story than one HR company's dashboard, and it is going to arrive at most companies through a finance review, not a technology decision.

How much was Rippling actually spending on AI tokens?

The company has not published a dollar figure. What it published is the ratio: token spend was on a forecast path to equal 40% of its R&D headcount budget, growing 80% month over month, with a peak volume of 605 billion tokens in a single month. TechCrunch describes the absolute number only as millions of dollars.

Is the AI Spend Console available now?

The visibility and attribution product is, on a 30 day free trial that does not require a Rippling subscription. The enforcement side is not fully there: Rippling's own post says governing approved LLMs is coming "soon" and directs readers to a waitlist for its AI Gateway.

Can I do AI spend management without buying a product?

Yes, and for most companies under a few hundred people that is the right call. A per-tool monthly cap, a daily circuit breaker, a written model-routing policy and a monthly spend-per-outcome number will get you most of the result. Buy tooling when the manual collation starts costing more than the licence.

What is an AI gateway and do I need one?

It is a routing layer that sits between your people or agents and the model providers, so all traffic passes one control point for limits, model access policy and logging. You need one when AI calls are scattered across many services and you cannot answer "who spent what on which model" without a spreadsheet.

Does using a more expensive model produce better work?

Sometimes, but not proportionally to the price. Rippling's internal data found the biggest productivity gains came from adopting AI tools at all rather than from spending more on frontier models, and it held token volume roughly flat while cutting cost to 37% of the earlier month by routing better.

Citation Capsule: Token spend growing 80% month over month on a path to 40% of R&D headcount budget; 10 to 15% of employees driving about 60% of spend; one engineer at $50,000 a month; 605 billion token peak against 600 billion in July at 37% of April's cost; 20 AI captains nominated. Rippling engineering blog, "From unchecked AI spend to complete control" (Aug 6, 2026) · TechCrunch, Julie Bort (Aug 7, 2026). The 63% figure is arithmetic on the chief product officer's statement to TechCrunch that July's token spend cost 37% of April's. That 37% appears in the TechCrunch interview rather than in Rippling's post, and neither source states 63% directly.