Choose the Level of Automation

Six levels, and the point of naming them is that most teams only know two.

Observe. The system records what it would have done and changes nothing. Almost nobody uses this and almost everybody should, because it is how you find out what the workflow would actually do at volume before it does it. Recommend. The system tells a person what it thinks should happen, in words. Queue. The system creates the task with the specifics filled in, and a person does the work. Draft. The system produces the actual artifact, the email text or the field value, and a person reviews it and commits it. Act on approval. The system performs it once a named person approves this specific instance. Act. The system performs it.

The published reference on permissions teaches a five-level version of this. This guide splits its middle rung into queue and draft, because the difference between handing someone a task and handing them a finished artifact turns out to change who is really deciding.

Blast radius, not confidence

The instinct is to set the level by confidence. Automate what the system is sure about, escalate what it is not. It reads as caution and it produces a system that automates its most confident mistakes.

The right question is what happens if the action is wrong. Four categories are enough. Internal and reversible: a field written on your own record, where wrong is a correction. Internal and hard to reverse: money spent, a record merged, history collapsed. Nobody outside sees it and you cannot simply undo it. External and reversible: something reached a person and can be walked back, though they may already have received it. External and irreversible: it left the building and there is no version of it you can take back.

The ceiling follows from that, and it is not negotiable by how good the model is. Internal and reversible can run at act. Internal and hard to reverse, and external and reversible, stop at act on approval. External and irreversible stops at draft: a person commits it, every time.

The example that proves it

A duplicate-detection system finds two kinds of match. One kind auto-merges above a confidence threshold, which is correct and saves an enormous amount of work. The other kind is detected and never auto-merged at any confidence, ever, because merging those particular records carries consequences that reach into open deals and signed contracts.

The second rule is not about confidence. The system can be completely certain and it still does not act. If you set the level by confidence, that rule is unwritable, because there is no threshold high enough to express never.