The machine does the work, a person signs off
A pattern for putting AI into workflows with real consequences: paper mode first, approval queues, and decisions explained in plain English.
Most of the automations I build share one rule, whether they post videos, place trades or print labels: the machine does the work, and a person signs off on anything that matters.
It sounds conservative. In practice it’s what lets you automate more, because you can hand an AI system real responsibility without betting everything on it being right every time.
1. Start in paper mode
Anything that can spend money or publish something starts in a mode where it can’t. The trading research bot runs against a paper account. The video pipeline renders but doesn’t post. You get to watch the system make real decisions on real data, with nothing at stake.
Paper mode isn’t a phase you rush through. It’s how you collect the evidence that the system has earned the next step.
2. Gate the irreversible step
When something does go live, the irreversible action — the order, the post, the email — goes through a queue. A person approves it. That approval should be cheap: one glance, one click.
The design work is in making that glance enough.
3. Explain decisions in plain English
An approval queue full of raw JSON is a rubber stamp waiting to happen. Every pending action should read like a note from a colleague: what it wants to do, why, and what it costs if it’s wrong — in dollars, not percentages.
This is where language models genuinely shine. They’re very good at turning a pile of signals into a short, honest summary. Let them argue both sides, too: a “for” case and an “against” case make it much harder to approve something on autopilot.
4. Make the safe path the default
Two switches to go live, not one. Position limits so a single bad call is survivable. A kill switch that’s always one click away. Logs written for a human reading them later, not just for a parser.
Why this matters for AI-first development
AI makes it dramatically cheaper to build the work part of an automation. That shifts the interesting engineering to the boundary: where the system hands control to a person, and how it earns more autonomy over time.
Get that boundary right and you can automate things you’d never have trusted to a script before.