Refactoring Orgs

I like numbers. But I tend to distrust large numbers that are being multiplied and floated around. They have a certain smell to them.

A new paper from a team at the Elasticity Institute (https://elasticity.institute/rsi-paper.pdf) did what most AI advocacy has quietly avoided: put a number on the actual research productivity gain per unit of capability. The self-reported uplift from frontier labs sits around four times. The back-of-envelope calculation using measurable capability progress puts it at about nine percent per unit on the standard capabilities index. The gap between those two numbers is the gap between "the exponential is here" and "we are still on a normal curve."

That is not a small correction. Every business case for AI tooling written in the last eighteen months uses a productivity multiplier grounded in self-report. Some are five times. Some are ten. All of them survived executive review by looking like the four-times number was a floor, not a ceiling.

The nine percent is not disappointing. But it may not be what your CFO signed off on, or why you decided to invest millions into Agentic Coding tooling. The exposure lives in the delta. When the tooling budget was greenlit projecting a 4x productivity gain at last year's numbers, the ROI made sense. At the actual number that emerges when someone runs the math with published capabilities data, it may not hold. That is important ammunition for your next renewal conversation.

Let's face it: That four times multiplier was the story. It convinced a lot of people in high positions to heavily invest in AI tooling and tokenmaxxing. It also smelled fishy from the start. The nine percent is the ledger.

I still like numbers.