Every time I describe how I run an AI project manager, a team of AI department specialists, a meal planner that fetches grocery deals on the weekends, I get the same question, usually with a wince… doesn’t that burn all your tokens?
Fair question! It would if I ran it the way many people run AI. They paste everything in, let the model figure it out, repeat weekly. I did some of that early on but after the potential bill and the slowdowns, I researched and picked up a different set of habits. Here they are, roughly in order of how much they save.
1. If a script can do it, we let a script do it.
Tokens burn when a model reads and shuffles data. So anything deterministic like fetching, filtering, formatting, gets moved out of the AI entirely. On the personal level, my meal planner used to have Claude pull three stores’ weekly ads and two coupon sites and merge them (that’s 30–50k tokens, every week). Now a 200-line Python script does the identical job on a schedule. Weekly token cost there is now next to zero. This is the single biggest lever, and most people never pull it because the AI can do the work. Can is not should.
2. We filter data before the model ever sees it.
That same script used to pass along everything the ad data contained, which is exactly how a hot tub could up in my grocery deals. Now it whitelists categories and the file the AI reads dropped by nearly half. The cheapest tokens are the ones the model never reads.
3. We put routine agents on a cheaper model.
Our department agents run on a model tier roughly 5x cheaper than the one I talk to. Routine department work doesn’t need the most expensive brain in the building, save that for Judgment calls. You don’t put your most senior person on data entry (hopefully).
4. We give every file a “load when” rule.
Each of our systems has a table in its instructions (this file, this purpose, load it only when doing X). The instruction that saves the most money in my setup is four words long: load only what you need.
5. We use Google Drive for Desktop (any local sync will work).
Our agents save deliverables as real files into client folders that sync automatically. The alternative is pushing documents through an API, encoding files into the conversation that burns tokens on plumbing and fails more often.
6. Don’t let the same model re-read its own work.
A surprising amount of waste is the model double-checking things that can’t have failed. My commands say it explicitly: read the one-line confirmation, not the whole file.
7. We batch our questions.
Every round trip resends the whole conversation. My meal planner used to stop three separate times with three separate questions. Now it asks one combined question.
8. We cap and schedule the autonomous stuff.
My overnight dispatcher drafts work while I sleep and is capped at two tasks a night, one at a time. Autonomy without a budget isn’t automation.
9. We audit what loads on startup.
I personally archived 22 skills I wasn’t using and demoted two tools to on-demand. It’s the AI equivalent of unsubscribing.
Spend tokens on judgment, spend code on everything else. That’s why an eight-agent setup costs less than people assume a chatbot habit does.