Human-on-the-loop: where automation should stop
A simple framework for deciding which steps run alone and which always wait for a person to approve.
The question we get on every intro call, in some form, is: "how do I know it won't do something stupid?" It's the right question. The answer isn't "trust the AI" — it's a design decision you make step by step, and there's a simple framework for it.
The one-question test
Every step in a workflow lands in one of three zones:
- Run alone. Cheap to get wrong, easy to undo, or purely internal. Classifying an email, drafting text, pulling data, filing a receipt. If a mistake means "a draft was slightly off", automate it fully.
- Run, then report. Actions with real but bounded consequences. Sending a routine reply, adjusting a bid within a set range, scheduling posts. Let it run inside hard limits, and post everything it did to a channel you actually read.
- Wait for approval. Irreversible, expensive, or reputation-touching: refunds above a threshold, anything legal, first contact with an important person, publishing under your name. The workflow prepares everything and stops. A human taps yes.
Approval should be one tap
The reason most "review steps" fail is that they're heavier than doing the task. If approving a refund means opening three tools, people stop reviewing and start rubber-stamping. We design approvals to arrive with full context — what's requested, why the AI recommends it, what the policy says — so the human decision takes ten seconds and stays a real decision.
Loosen over time, on evidence
Start conservative. Every workflow we ship logs each decision and its outcome, so after a month you have data: this step was approved unchanged 240 times out of 241. That's when you move it from "wait for approval" to "run, then report" — deliberately, based on a track record, not on optimism.
The mistake in both directions
Too little automation and you've built a fancy to-do list; too much and one bad edge case erases months of saved hours. The framework keeps you honest in both directions — and it's why every workflow on our site lists its escalation rules as a feature, not a footnote.