Field Guide · Judgment and ops
When not to use AI
Skipping the model is a skill. Here is when enthusiasm should wait.
The skill is pause
Using a model well includes knowing when not to open the tab. Pause is not Luddism. It is judgment under enthusiasm.
This chapter is a decision aid for the first thirty seconds after someone says “just ask the AI.”
Pause when
Use this checklist before you paste or prompt:
- You cannot spot a bad answer. If you lack enough domain sense to reject nonsense, the model is guessing in a language you cannot audit.
- The task is irreversible and high stakes without a second checker (money moved, customer notified, policy shipped, medical or legal advice implied).
- You are avoiding a skill you need to own. If the job requires you to write, calculate, or decide unaided next week, outsourcing today can dig a hole.
- Secrets or regulated data. Credentials, PHI, customer dumps, unpublished deals, or anything your policy forbids leaving the trust boundary.
- You need a primary source, not a paraphrase. Citations, statutes, vendor contracts, and incident timelines belong in the original document first.
- Emotions are hot. Angry emails, performance feedback, and public posts get worse when fluency arrives before cooling off.
- The ask is a decision dressed as a draft. “Should we fire the vendor?” is not a drafting task. Separate brainstorm from decide (Pulse when relevant).
If two or more boxes apply, stop. Do the replacement activity first.
Replacement activities
Pick one that fits the risk:
| Situation | Do this instead |
|---|---|
| Domain fog | Sketch the problem on paper; name what “good” looks like in your words |
| High stakes | Ask a colleague who owns the outcome; require a second pair of eyes |
| Primary source needed | Open the document, statute, or ticket; quote before you summarize |
| Skill you must own | Do a short unaided attempt; only then compare with a model draft |
| Secrets | Use approved tools only, or redact until nothing sensitive remains |
| Hot emotion | Sleep on it, or write a private note you will not send |
The goal is better judgment, not maximum tokens.
A thirty-second script
Say out loud (or write in the ticket):
- What decision or artifact am I producing?
- Who gets hurt if this is wrong?
- Can I check the risky parts without the model?
- If yes to pause: what is the smallest non-AI next step?
Then either prompt with a tight brief or step away.
Team signal
When a teammate reaches for a model on a pause case, ask one question: “What would make you trust the answer enough to ship?” If they cannot answer, the pause was correct.
Practice
Take one real task from this week. Run the pause checklist. Document whether you used a model and why. Share one sentence with your team so the norm is visible.