Thinking in Tokens: How to Run an AI-Native Operation

When I launched Regulus Consulting, I said I would run the firm the way I tell clients to run theirs. This is what that looks like day to day.

I’m a solo operator. I work like a small team. The difference is a handful of AI tools, a few rules about how I use them, and one unit of thinking that changed how I scope everything.

The token is the unit of work

A token is a chunk of text a model reads or writes. Every prompt costs tokens. Every answer costs tokens. Every document you hand over costs tokens.

Once you see work this way, you scope differently. You stop asking “can the model do this?” and start asking “what is the smallest, cleanest input that gets me a good output?” Cost is part of that. So is quality, because a model buried in irrelevant material performs worse than one given exactly what it needs.

The context window is a desk

Every model has a context window. Think of it as the desk the model works at. Only what’s on the desk exists for the task.

So I clear the desk. One task per session. Relevant material only. When a long job runs across stages, I carry forward a short summary of where we are instead of the whole transcript. A clean desk is cheaper, faster, and more accurate.

It’s also a confidentiality habit. If a document doesn’t need to be on the desk, it isn’t.

Commercial terms are the lock on the door

Here’s where my law practice shapes my consulting practice. I only run client work through products on commercial terms. API access and business tiers, where the contract says my inputs aren’t used to train the provider’s models. Google Workspace with Gemini on the same logic. Claude through the API on the same logic.

Consumer chat apps don’t get client material. Free tiers don’t get client material. The convenience is real and the terms are different.

Match the model to the job

Not every task needs the strongest model. Sorting notes, reformatting a table, drafting a first-pass summary: a smaller, cheaper model handles it. Analysis that has to hold up under scrutiny gets the stronger one. Routing work this way keeps costs sane without lowering the bar where it counts.

AI is my junior associate

It drafts fast. It never gets tired. It also needs supervision, and nothing leaves my desk without a human review. Every deliverable, every time. Speed comes from the tools. Judgment stays with me.

What this means for clients

If you’re weighing how to bring AI into your operation, start with the same questions I ask myself. What is the unit of work? What does the model need to see? What do the terms say about your data? Who reviews the output?

Answer those four and most of the rest gets easier.