Story
When an AI agent (an AI that can take actions for me) does the same kinds of jobs for me every day, I want to choose which ones it may do alone and which ones wait for me. Then I am not asked about small things and not surprised by big ones.
Amara, who sells fabric online from Lagos, answering customers on her phone at the market.
- 01 Setting
Amara lets the shop assistant handle customer messages - replies and refunds.
- 02 Trigger
3 messages come in: 1 question, 1 small refund and 1 big refund.
- 03 Without it
The agent asks about all 3. By the 3rd tap she approves without reading - or she turns asking off for everything.
- 04 With it
The dial is set to "Ask for big ones". The reply and the small refund go through on their own. Only ₦45,000 waits.
- 05 Outcome
She approves 1 refund and sees which 2 the agent did alone.
Live demo
This demo follows a script, so it plays the same every time - there is no real AI behind it. Change a setting to see the same moment play out differently.
What happens, step by step
| State | The person sees | The AI is | When | Screen reader |
|---|---|---|---|---|
| 1 | The current level and its rule in words: "Refunds over ₦20,000 wait for you." | Idle. | set once, change any time | - |
| 2 | Each job, tagged "Done alone" or "Waiting for you". | Acts below the line, stops above it. | - | “Working” |
| 3 | Only the jobs above the line, with the action named on the button. | Paused on those jobs only. | when a job crosses the line | “Needs your approval” |
| 4 | The person approves or says not now. | Paused. | on tap | - |
| 5 | A count of what was done alone and with you: "3 done · 2 alone, 1 with you". | Idle. | when done | “Done” |
On each device
- phone
- The dial is a 3-step control at the top of the agent's screen. The rule for the current level sits under it in 1 line.
- tablet
- As on desktop.
- desktop
- The dial sits in a column beside the work, so people see the setting and its effect together. More detail - 1 dial per kind of job - can open in settings.
When to use it
- The agent does the same kinds of jobs again and again, and some are riskier than others.
- People's trust grows over time - they want to start careful and loosen up.
- There is a clear line people understand: an amount, "to people outside my team", "can it be undone".
- The action must always ask, whatever the dial says - paying someone new, deleting for good. Use Approval gate instead.
- Every job is safe and undoable. Let it act, and show what it did. Use Activity log and undo instead.
- It is a 1-off task. Agree this run's steps instead of a lasting setting. Use Intent preview and plan instead.
The parts of the pattern
- ADial2 to 4 levels, from ask every time to just do it. The current 1 is filled.
- BRule in wordsWhat the level means for this agent, with the real line: ₦20,000.
- CDone aloneA tag on each job the agent did without asking.
- DWaitingOnly the jobs above the line stop here.
- EApproveNames the action and the amount.
Do and Avoid
How much can it do alone?
Ask every timeAsk for big onesJust do it
Refunds over ₦20,000 wait for you.
Autonomy
LowMediumHigh
Medium
✓ Reply to Chidi · Done alone
✓ Refund ₦8,000 · Done alone
Refund ₦45,000? · Waiting for you
All done ✓
3 messages handled
Writing and accessibility
- Name levels by what happens: "Ask every time", "Ask for big ones", "Just do it". Not "Low", "Medium", "High".
- Put the real line in the rule: "Refunds over ₦20,000", "Emails to people outside the team".
- Tag results "Done alone" or "Approved by you", so people can check the setting.
- Say the level can change: 1 short line and a link to the dial.
- Build the dial as a radio group (a set of options where only 1 can be picked) with a visible label, not as a slider with no steps.
- Each option's name says what it does, so it reads well on its own.
- "Done alone" and "Waiting for you" are words, not only colour dots.
- Announce "Needs your approval" when a job stops, and the count when there are several.
Seen in real products
- Claude in ChromeChrome extension · Nov 2026A drop-down on the chat input sets the mode: "Manually approve", "Automatically approve" (a safety check reviews each action) or "Skip all approvals". Some actions ask in every mode.Claude Help: Chrome permissions guide ↗
- ChatGPT WorkWeb, apps and desktop · Jul 2026People decide what the agent can access, when it needs approval before acting, and when it should check in.OpenAI: ChatGPT for your most ambitious work (9 Jul 2026) ↗
Each example comes from the product’s help pages, release notes or teardowns, with the month it was checked. Products change every month - tell us when 1 is out of date.
Check yourself
Seen this pattern done better, or a product that changed? Press Suggest next to any heading, or tell us about the whole page.