Honest AI

No token tax: when a plain script beats calling AI every time

July 7, 2026 · 5 min read

Here's something the AI industry has no incentive to tell you: for a lot of everyday business tasks, the best use of AI is to help you build a tool that never calls AI again.

First, a translation

When software "calls an AI model," it sends your request to a company's servers, and you're billed for the amount of text going in and out — measured in units called tokens. Each call costs a fraction of a cent to a few cents. Sounds tiny. But a task you run five hundred times a day, every day, turns "tiny" into a permanent line item. I call it the token tax: a toll you pay every single time the work runs.

And you're not just paying money. AI models are probabilistic — ask the same thing twice and you can get slightly different answers. For creative work, that's a feature. For "rename these files the same way we always do," it's a bug you're paying for.

The alternative: deterministic scripts

A deterministic script is ordinary code that follows fixed rules: same input, same output, every time, forever. It runs on your own computer, instantly, for free. No subscription, no toll, no surprises.

Some of the most valuable systems I've built are exactly this — local Python scripts with no per-use AI cost at all. The AI's role was huge, but it happened once: it helped me write the script. The AI cooks the tool; the tool then works for free, forever.

A concrete example

When I processed over a gigabyte of NHTSA vehicle-complaint text (that story is here), the naive approach would have been to feed every complaint to an AI model and ask "what part failed?" Hundreds of thousands of records, every record a paid call, re-running the whole bill any time I tweaked something.

Instead, I used AI to help me write Python extraction code — and then ran that code locally as many times as I wanted. Refining it cost nothing. Re-running it cost nothing. And because the rules were fixed, I could trust that a change in results came from my changes, not from the model's mood.

How to tell which one you need

Ask one question about the task: could you write down the rules?

• If yes — script it. "Move every invoice PDF into a folder named for its month." "Flag any order over $500 missing a PO number." Fixed rules, fixed logic. Pay the AI once to help build it; never pay a toll to run it.
• If no — that's AI's territory. "Summarize what this angry customer email is actually about." "Draft a first response." When the input is messy human language and the rules can't be written down, a model call genuinely earns its cost.
• Most real workflows are a mix. The winning design is usually a script for the repetitive spine of the task, with an AI call only at the messy joints — the smallest possible token tax, placed exactly where it buys something.

The skill isn't "using AI everywhere." It's knowing where the rules end and the judgment begins — and only paying for intelligence on the judgment side of that line.