Microsoft exempted Copilot from a company-wide hiring freeze. Same week, AI developers measured 19% slower than they believe. Same week, $400 in tokens saved $500K/year.
Last week I shared the DORA data: 90% of devs use AI, delivery metrics are flat. This week, three data points explain why.
-
Microsoft froze hiring across Azure. Every team except Copilot. They are betting AI replaces headcount before it proves ROI.
-
90% of Claude Code's 20.8 million commits went to repos with under 2 stars. Developers believed they were 20% faster. Measured on specific tasks: 19% slower.
-
Reco.ai rewrote JSONata in Go with AI. Seven hours. $400 in tokens. 1000x speedup. $500K/year saved. (The spec was formal. The tests were exhaustive. The AI could verify its own work.)
DORA showed the gap last week. These three stories show why.
Give AI a formal spec and a test suite, it ships. Give it ambiguous requirements, it makes things worse. Two completely different outcomes hiding behind one average.
The average is lying to you. Measure by task type or you are measuring nothing.
PS: The article of Reco.ai is worth reading I believe: https://www.reco.ai/blog/we-rewrote-jsonata-with-ai