Rippling cut its AI token bill from 40% of engineering payroll to 15%, without cutting AI use

One Rippling engineer used $50,000 worth of AI tokens in a single month. The company found that out only after its chief financial officer, Adam Swiecicki, put a projection in front of the executive team in March: Rippling was on course to spend 40% of its research and development payroll on tokens, millions of dollars, with the bill growing 80% month over month. Left alone for a year, chief product officer Matt MacInnis told TechCrunch, the token bill would have come to nearly 90% of what that whole unit is paid.

A software licence priced per person has a number a finance team can multiply out in January. Tokens have no such number. The bill follows what each employee asks the model to do that day, and at Rippling 10% to 15% of the people using the tools accounted for about 60% of the cost. Much of it went on routine work sent to the most expensive models available.

Uber's chief technology officer said this week that his company had spent its entire 2026 budget for Claude Code by April.

Rippling built a tool that tracks tokens by employee, team and role and puts that cost beside what the person produced, pull requests and code reviews included. It ran its own comparisons between models and moved routine work to cheaper ones. The tool is now a product, AI Spend Console.

Token spending fell from 40% of that payroll to about 15%, and the work did not fall with it. Rippling used 605 billion tokens in April and 600 billion in July. July's bill came to 37% of April's.

Brian Chesky told investors on Airbnb's second-quarter call that AI now writes about 60% of the company's code. The time from an idea to a live feature has fallen by as much as 60%, and Airbnb has shipped nearly 80% more features and improvements this year than in the same period last year. The engineers did not leave; they got through more.

Airbnb's support assistant now closes about 45% of contacts with no person involved, the questions about a refund or a booking date that used to reach an agent.

Humanitarian agencies are sorting their work the same way under harder conditions. Global humanitarian funding fell by more than 30% in a single year, and GiveDirectly, Mercy Corps and the International Rescue Committee have put AI on damage assessment after disasters, on the report writing that consumes program staff, and on the questions refugees ask over and over, all of it under tight rules on what the tools may decide.

Airbnb and the aid groups handed over the same kinds of work: the drafting, the digging, the first line of questions. The one thing an Airbnb guest would actually touch, a search box where you type what you want instead of clicking filters, is still a test.

Employment judges in England, Wales and Scotland have been told to be stricter about interim relief, the order that keeps a dismissed employee on full pay while a tribunal case runs. Applications for it have gone from about 20 a year to 20 a month at each office.

Judges connect much of that rise to claimants drafting with chatbots. A chatbot asked how to make an employer keep paying finds interim relief in the law and writes the application. It does not weigh the high legal threshold behind the order, which tribunals grant rarely.

Figures published with the guidance show 64,000 single cases open at the end of March, up from 45,000 a year earlier. Courts in Australia and the United States report the same pattern of chatbot-drafted filings.

The guidance, from Judge Barry Clarke, president of the employment tribunals in England and Wales, and Judge Susan Walker, his counterpart in Scotland, changed no law. It took effect on June 22 and tells judges to hold the existing threshold and to deal with weak applications earlier.

Among people who use AI at work every day, 44% report a real gain in productivity. Another 26% say they lose time.

Education and job type separate the two groups. In a survey of 4,595 unionized workers in Quebec, run by the Obvia research group with 11 unions representing more than 1.4 million members, the gains in output and wellbeing went mostly to highly educated professionals. Workers with less education and those in industrial jobs reported the opposite from the same technology: more monitoring, heavier workloads, and fear for their jobs.

Only about a quarter of the workplaces in that survey had any written rules for AI use.

Cloudflare gave its own staff an internal AI workspace in May. By August thousands of employees, most of them not engineers, were using it daily to draft documents, build slide decks and put together small internal tools. On August 5 the company released it as open source, under the name Cloudflare OS, for any organization to run on its own.

Someone describes an app in plain English and the system builds it. Each app runs as a private, isolated copy that its owner can change by asking, and anything it wants to touch outside that box goes through a component Cloudflare calls a Gatekeeper, which authorizes the access, logs what was done, and holds actions back for a person to approve.

Self-hosting it requires a paid Cloudflare Workers plan and a key from an AI provider.