Why every AI task needs an approval gate
By Alejandro Rioja
There's a tempting version of AI task management where you assign everything and check nothing. It works until an agent confidently marks the wrong thing done, or ships something you'd never have approved. The fix isn't less automation — it's putting a gate on the moment that matters.
Auto-OK, or approve?
In Done, every task carries a simple flag. If it's marked auto-OK, an agent can complete it directly — good for low-stakes, reversible work. If it isn't, the agent can't close it. It can do the work and propose completion with a summary, but the task sits in pending approval until a human approves or rejects it.
The same idea applies before the work, not just after: for something substantial, an agent proposes a plan and waits for you to approve the approach before it spends any effort.
The rule that makes it trustworthy
- A gated task can only reach done through a human approval — never directly from an agent.
- Only a human can make a task auto-approvable; an agent can't lift its own gate.
- Every proposal, approval, and rejection is recorded in the task's activity, so there's always a trail.
That's what lets you delegate widely without watching every step. You decide once, per task, how much you trust the outcome — and the system holds that line for you.
Frequently asked questions
What is an approval gate for AI tasks?
It's a rule that a task marked as needing approval can't be closed by an AI agent directly. The agent proposes completion with a summary, and a human approves or rejects it before it's marked done.
Can an AI agent bypass the approval gate?
No. In Done only a human can make a task auto-approvable, and a gated task can only reach done through a human approval — an agent can never lift its own gate.