Graham McNicoll

If your finance team is currently drafting a memo about capping engineering AI spend, I am going to save you a quarter of productivity loss before you send it. Token costs are exploding, and the reflex from finance is predictable. Ask the engineers to tone down their token usage. I think this is a mistake. Token usage corresponds to productivity. You would never walk into your engineering team and tell them to be less productive. Asking them to throttle their AI usage is doing exactly that, one layer removed. There is a reasonable version of this conversation. Teach people how to get to the answer in fewer attempts. Teach people to know which models to use for which tasks. That is a skills investment. The unreasonable version is a blanket directive to use less AI. The math does not work. If an engineer gets to a solution in three hours with heavy token use, versus 5 days with light token use, you are spending many hours of engineering time to save a handful of dollars in tokens. The better question is not how much you are spending on AI. How much progress is your product making? Are the things you engineers ship with AI actually moving the metrics you care about? Token spend is the input. Validated product improvement is the output. If you are only measuring one of those, you are thinking about it wrong. If your team is shipping AI-generated code and nobody is measuring whether any of it worked, send me a DM. Happy to talk through what that report looks like.