July 6, 2026
Prompt around and find out
Not everything should cost a token: the case for deterministic AI
Why are people paying AI to do chores a tiny script could do for free?
TLDR: The article says companies are wasting money by using AI for repetitive tasks that simple software can do faster, cheaper, and more reliably. Commenters were savage about it, joking that “deterministic AI” is just regular programming and mocking teams for paying extra to make basic chores worse.
The article’s big warning is almost painfully simple: stop paying chatbot-style AI to do boring robot chores. The example that set people off was a team using an AI system every morning to pull numbers from an app, rearrange the data, and save it in a table — a job commenters said a basic script could do perfectly, cheaply, and without random mistakes. The writer’s point was that AI should be saved for fuzzy human-like judgment, not repetitive housekeeping. But in the comments, readers didn’t just agree — they pounced.
The loudest reaction was pure disbelief. One commenter basically spat out, who on earth uses AI to format data? Another boiled the whole idea of “deterministic AI” down to a brutal punchline: “Aka, a computer program.” Ouch. The mood was part roast, part intervention. Several people mocked the now-common habit of treating AI like a magic machine for everything, with one commenter incredulous that an entire team didn’t just ask the AI to write the simple script once and move on.
And then came the comedy gold. One workshop attendee shared that people gave AI agents a fancy coffee-order challenge — only for the agents to immediately reverse-engineer the system and build boring, reliable parsers instead. The crowd seemed delighted by the irony: even the bots know not everything should be a prompt. If there was a community consensus, it was this: when the smart new thing starts replacing perfectly good boring tools, the invoice — and the ridicule — arrive right on schedule.
Key Points
- •The article uses a daily metrics API workflow to illustrate the cost and reliability problems of routing deterministic tasks through a language model.
- •It states that deterministic tasks handled by an LLM become non-deterministic, slower, and metered on every run.
- •The article identifies two main failure modes: token costs tied to mechanical work volume and context-window bloat caused by loading raw data into prompts.
- •It proposes separating work into judgment-based tasks for agents and exact, repeatable tasks for conventional applications or scheduled jobs.
- •The article also warns against using an agent’s memory or notes to store structured operational data as if it were a database.