How much AI should one employee cost?

In July I mentioned how one CEO I talked to spent over a million euros on Claude last year. The CEO didn’t have any proof that it was helping the business other than, “We feel good about it.”

This week the trend on token maxxing vs. efficiency—routers, limits, and measurement—continued to pick up speed. (See Vercel’s AI Gateway traffic data.)

Rippling, for example, made drastic changes to bring its AI bill back in line after the forecast reached the equivalent of 40% of its total R&D headcount budget. The most interesting story with Rippling, though, is how it is thinking about measuring efficiency on a per-employee basis. More below.

Other stories include peak/off-peak variable pricing from DeepSeek, tying AI to real-world infrastructure, and contrasting takes on job displacement: IKEA retrained its team, while Ramp shut down its public AI sales-development program and Replit CEO Amjad Masad argues that layoffs are “100%” coming. That interview also includes stories about Replit’s engineering speed and AI adoption that are worth listening to.

TechCrunch says Rippling's AI bill was growing 80% month over month, with 10% to 15% of employees driving 60% of the spend. In response, Rippling negotiated caps with Cursor, OpenAI, and Anthropic, built its own AI gateway, and sent suitable work to cheaper models. By July, its forecast had fallen from 40% to about 15% of its R&D headcount budget.

Rippling's new measurement and tracking console puts employee spend beside proposed code changes, performance ratings, and whether peers send the work back.

The peer-requested rework tracking is interesting. Wikipedia editors report being overwhelmed by AI-generated edits, and curl now has specific rules for AI-assisted contributions. Companies need policies and enforcement to make sure people own their work. Otherwise it’s too easy to write a short prompt and throw a long-ass piece of AI slop over to the team to review.

Vercel's gateway data shows token volume rising 59% while spend rose 37% and average price per token fell 13.6%. Vercel attributes the entire price drop to customers switching models.

For a celebration of AI’s potential without disrupting every job, look to our favorite flat-pack retailer.

IKEA reported that it rolled out its “Billie” AI to handle call-center work, and by 2023 Billie was resolving 47% of customer enquiries. Fortune now reports that Billie assists 74% of users.

So what happened to all the people? IKEA now operates resolution and sales teams from 24 remote-sales centers. And they’re the fastest-growing channel in the business, growing 15% to 20% annually and producing €1.25 billion last fiscal year.

DeepSeek is announcing major price changes. First, DeepSeek V4 Pro will change what it charges for cached input, meaning text it has already processed. The price for one million cached-input tokens will move from less than half a cent to 2.2 cents off-peak and 4.4 cents at peak.

The same cached input costs twice as much depending on when it runs, which makes me think this is like the modern version of off-shoring. Send the work into the cheaper UTC window, then come back later. Appealing if it works.

The other thing that’s interesting here is how it underscores that delivering AI tokens has a real marginal cost. It’s easy to think about AI as this ethereal set of bits and bobs, but each request still depends on machines, electricity, and grid capacity somewhere.

NVIDIA's pre-alpha Switchyard project is an open-source traffic controller for AI requests. It can choose models by task, price, and response time, run some work on local machines, send harder jobs to outside providers, and coordinate the work across the clients, models, and providers it supports. It’s an important step toward using different models and providers in one workflow.

That whole open-source vs. closed thing is going to keep dominating news cycles for macroeconomics, the fate of frontier labs, and geopolitics alike.