I was at a panel recently where a product manager and project manager shared that their teams were seeing 15% productivity gains from AI tooling. The room nodded. Nobody pushed back.
I was surprised. These were sharp people. But the number floated there, unexamined.
Nobody asked: what did those gains cost in application stability? More features shipped faster means more surface area, more incidents, more recovery time. Nobody asked about the OpEx increase from rolling out these tools.
A VC analyst recently documented his personal inference spend hitting $100k annualized in two quarters. An outlier — but he was asking the right question: what is this actually costing me?
Most teams aren’t asking that. Enterprise seat licensing looks modest on paper — until your engineers start running agent pipelines. Automated code reviews, commits, PR generation. The moment that happens, you’re no longer paying for seats. You’re paying for consumption. Token limits get hit. Overages accumulate. In my experience, consumption grows faster than productivity does.
Stack that against application instability and incident recovery time, and the fully loaded cost of that 15% productivity gain looks very different from what was reported on that panel.
Most teams can’t answer what it actually cost. They’re quoting the headline because that’s what circulates.
Net productivity means accounting for everything you spent — and everything you risked — to get the number you’re reporting.