Skip to content
Dustin's AI Lab
Go back

Judging AI Risk on Two Axes: Reversibility × Environment Isolation

A common question in corporate trainings — "can I let AI do this automatically?" I answer with two axes: reversible vs irreversible, isolated vs production. This 2x2 prevents more incidents than any prompt-engineering tutorial.


In this morning’s corporate training session I flagged a small risk-judgment framework worth keeping handy.

When you find yourself asking “can I let AI handle this automatically?”, first run it through two axes:

Why these two axes

Reversibility decides the cost when things go wrong. A wrong line of code can be edited back. A sent email cannot be unsent. Inserting a database row is recoverable. Dropping a table is catastrophic.

Environment isolation decides the blast radius. Running a sketchy script inside Docker — worst case is rebuilding the container. Running it on a production database — worst case is taking the company down.

Cross the two and you get four quadrants:

Isolated envProduction env
ReversiblePlay freely, mistakes are fineBe careful but rollback is possible
IrreversibleStill careful, but only affects youRed line of red lines

The red line of red lines

A real example:

Sending an email = irreversible Sending it to an external party = directly in production

Both at once — that is the red line of red lines.

The correct workflow: draft → human review → only then send.

My own hooks enforce this: every outbound message (Gmail, Slack, LINE Bot, third-party) goes to drafts first, I confirm, then it sends. No exceptions.

How to apply this

Next time you are about to let AI “handle X automatically,” ask yourself:

  1. If this goes wrong, can I get back to the original state in 5 minutes?
  2. What does this affect — only my own machine, or external people / systems?

Both answers in the “reversible + isolated” zone? Let AI run.

Any answer touching “irreversible” or “production”? Insert a human checkpoint — draft, preview, dry-run, however you implement it. But do not let AI run end-to-end.

This framework prevents more incidents than any amount of prompt engineering.


Share this post on:

Previous Post
Jack Dorsey Cut 40% and Rebuilt: The Shape of Companies in the AI Era
Next Post
GPT-image-2 Wants Fewer Constraints — Plus a Consistency Drill