The most expensive AI token in my lab was spent on a job that did not need AI at all.

Man and Machine: The Most Expensive Token Is the One That Did No Useful Work

An August audit found a routine that was supposed to be deterministic. The script already knew what to collect, how to compare it, and what counted as complete. But the wrapper around it was still making premium-model calls. The job finished, but the expensive intelligence had added nothing.

That mistake gave me a buying rule for the lab. Subscriptions are best when I am sitting with the model, writing, coding, testing ideas, or making a judgement. APIs are better for controlled automation, where the system needs a narrow task, a retry rule, and a receipt. Local models belong on bounded private work where they have been tested and are good enough. Some jobs should use no model at all.

The mistake is to begin with the strongest model and then find work for it. I now begin with the job. Does it need judgement, structure, privacy, speed, or simply a script? Then I choose the smallest route that can do the work well and prove that it finished.

My agents can inspect usage, compare routes, and show me where capacity disappeared. They can also recommend a cheaper or more private path. I still decide whether the result earned the cost, and whether saving money would quietly weaken the work.

The best personal AI system is not the one with the most intelligence available. It is the one that knows when intelligence is needed, when it is not, and leaves me with work I can inspect and trust instead of more machinery to babysit.

-dg