Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: Autonomy runs from suggesting (the user decides everything), to drafting for approval, to acting with an undo window, to acting within a budget or limits, to fully delegated tasks. Choose the level for each action separately, based on how costly and reversible a mistake is and how much evidence you have that the AI gets that action right. Start lower and move up as evidence accumulates.
How to approach it
Lay out the levels with one example each, then the decision rule, then how autonomy changes over time. Keep it concrete.
A strong answer
The levels, using an email assistant as the example.
| Level | What the AI does | Example |
|---|---|---|
| L0: Suggest | Offers options; user does everything | Suggests three reply phrasings |
| L1: Draft for approval | Prepares the action; user approves or edits | Drafts a reply; user clicks send |
| L2: Act with undo | Acts, notifies, user can reverse | Archives newsletters; "undo" available for a period |
| L3: Act within limits | Acts within budgets or rules without asking | Accepts meeting invites from colleagues within working hours |
| L4: Delegate | Handles a whole task, reports back | "Sort out the travel for next week's offsite" within a budget |
How to choose. Two questions for each action:
- What does a mistake cost, and can it be undone? Archiving an email: cheap, reversible, so L2 or L3. Sending an email to a client: costly, not reversible, so L1. Paying an invoice: costly, often irreversible, so L1 with strong checks.
- How much evidence do we have? Even a low-risk action starts at L1 until approval data shows the AI's proposals are reliably right.
Autonomy changes over time. Approval data is evidence. If users approve 98% of the AI's meeting-acceptance drafts unchanged over hundreds of cases, offer to move that action to L3. If undo rates rise, move it back down. Users should be able to choose a lower level for themselves at any time.
What the research and the usage data add. Feng, McDonald and Zhang's Levels of Autonomy for AI Agents (2025) define the levels by the user's role instead of the agent's: operator (the user is in charge, the agent helps on demand), collaborator (they plan and execute together), consultant (the agent leads and the user gives direction), approver (the user steps in only when the agent is blocked) and observer (the agent runs alone and the user can only watch). Their central point is that autonomy is a design decision made separately from capability: a very capable agent that must ask before every action is still low-autonomy. That is the PM's lever.
Anthropic's measurement of Claude Code usage (18 February 2026, company-reported) shows how the setting moves in practice. About 20% of new users' sessions used full auto-approve, rising to over 40% as users gained experience, while experienced users interrupted Claude in about 9% of turns against 5% for new users. Experienced users delegated more and steered more. On its public API, 0.8% of sampled agent actions appeared irreversible. Two design consequences follow. Autonomy should be a per-user setting that the product invites people to raise as they gain experience. And as approval gates fall away, a fast interrupt-and-redirect control becomes the main safety mechanism, so it needs as much design care as the approval screen.
A common mistake is treating autonomy as one setting for a whole product ("our agent is fully autonomous"). Real products mix levels: the same assistant may archive at L3, draft replies at L1, and never send payments without approval.
What interviewers probe next
"Users ask for full autonomy right away." Let them opt into higher levels for low-risk actions, with clear undo; keep high-risk actions at approval regardless.
"How does L4 work safely?" Budgets (money, time, number of actions), limits on which tools can be used, checkpoints for anything outside them, and a report of everything done.
How it's scored
Weak: Describes autonomy as "the AI does things automatically" with no levels or rule.
Hire: Lists levels and links the choice to risk.
Strong Hire: Chooses per action using cost and reversibility plus evidence, explains how autonomy moves up and down with data, and designs the interrupt control that replaces approvals as users delegate more.
Common mistakes
One autonomy level for the whole product.
Granting autonomy before collecting approval evidence.
No path back down when errors rise.
See autonomy levels.