Writing

AI should earn its autonomy

How I decide how much an AI feature should do on its own, and the patterns that keep people in control. Drawn from designing TIP, an AI tool that drafts LinkedIn Learning course content.

Autonomy is a ladder

Every AI feature sits somewhere between suggesting and acting on its own. I make that choice on purpose, rung by rung.

1. Level 1SuggestSurfaces an insight. A person does the work.
2. Level 2DraftCreates content for a person to review. A person decides what ships.
3. Level 3Act with approvalPrepares an action and waits for a yes.
4. Level 4Act and reportActs, then reports back, with undo.

Six principles I design by

1. 01
   Pick the rung by the stakes

   The more a mistake costs, the lower the rung. TIP’s content goes out under LinkedIn’s name, so it stays at Draft: the AI writes, and an editor decides what publishes.

2. 02
   Show where it might be wrong

   A confidence score on every item, and a reason when it’s low, so people know where to look instead of rereading everything. “The transcripts cover roads, not intersections” is more useful than “58%”.

3. 03
   Fix it where it is

   Edit inline, or regenerate just the one item. If fixing the AI’s work means starting over somewhere else, people stop checking.

4. 04
   Make every step reversible

   Deletes come with undo, and any single item can be regenerated without touching the rest. Reversibility is what makes it safe to move fast.

5. 05
   Be clear about who’s responsible

   In TIP, Legal asked for a notice at the start: use AI with caution and double-check everything. I raised the cost in friction; Legal decided the caution was worth it and set the wording, and editors can turn it off and reopen it. It makes the handoff explicit: the editor owns what ships.

6. 06
   Aim for calibrated trust, not maximum trust

   Enough trust that people don’t redo the AI’s work, not so much that they stop checking. That balance is the design goal.

What it added up to

In TIP, these patterns let editors trust AI drafts enough to use them. Course production went from 6–8 weeks to 2, and content capacity went up 150%. Speed came from trust, and trust came from control.

Read the TIP case study →