How much should your AI agent do without asking? Design it action by action, with four levels of control and the controls your users need.
If your product includes an AI agent, one of the first design decisions is how much it may do without asking a person. You don’t set one level of control for “the agent”. You set it for each action it can take. The more an action can change, the harder it is to undo, and the less the agent knows at that moment, the more control a person should keep.
Why This Matters for Your Product
Give an agent too little room, and it adds steps instead of saving time: your users approve every click and soon stop reading what they approve. Give it too much, and its mistakes reach your customers before anyone can catch them.
How much makes sense depends on what your product touches. An agent that tidies a user’s own notes can run with little friction, though deleting someone’s notes still needs care. One that moves money or writes to your customers needs a person to say yes for those actions.
One Agent, Different Actions
The same agent in your product can do some things on its own and need a person for others. For example:
- A customer support agent can look up the refund policy and draft a reply on its own. Issuing the refund moves money, so it waits for someone on the team to approve it, unless it is a small amount within a clear policy.
- A sales assistant can sort a user’s leads and draft follow-up emails on its own. Sending those emails should wait for the user’s yes, because the client may read one and act on it before anyone can recall it.
- A coding agent can try a fix and test it on a copy of a project on its own. Putting that fix live should wait for the project owner to decide.
Four Levels of Control
Give each action one of four levels. Each one hands the agent a little more. “A person” here is whoever answers for the action: your user, or someone on your team.
- Suggest. The agent offers options. A person picks one.
- Prepare. The agent drafts the change. A person makes it.
- Act with approval. The agent shows the change. A person says yes.
- Act and report. The agent makes the change. A person gets a report.
How to Set the Level
Three things decide the level for each action. When they point different ways, choose the level where a person stays more involved.
- What it changes, and who it reaches. Renaming a private draft is minor. Messaging customers, moving money, or editing thousands of records at once is not. The further a mistake would spread, the more control a person keeps.
- Whether it can be undone. A moved calendar block can be put back. A sent email may already be read and acted on. Build an undo for changes that can be reversed, and put a preview and an approval step before the ones that can’t.
- Whether the agent has the facts. An action needs the right details, a current source, and no conflicting instructions. When something is missing, the agent should ask, or stop at a draft.
Start each new action with more human control. Give it more room once your record of errors and corrections shows it is ready, and look again whenever the model, the data, or the tools change.
Build the Controls Into Your Product
Your team sets the limits. Within them, users can see and adjust how much the agent does, and some limits stay fixed for everyone. Someone new to your product needs to understand what the agent does before trusting it with more. Someone using a screen reader or a keyboard may need more time and a clear confirmation at each step. Someone who does the task every day may want it to get on with the work. Design for all three.
- Settings per action. People pick each action’s level, within your limits.
- “Ask me first.” One switch sends any action back to approval.
- Previews. Show what will change and who it reaches.
- Stop. Pause the agent mid-task and see where it stopped.
- History. Each action, when it ran, and how to undo it.
A setting on screen is only a promise. The product itself has to enforce each limit, so the agent cannot act beyond it, whatever its instructions say.
Start new users with the agent asking first, and let them give it more room as their understanding grows.
Check What the Law Requires
Some rules require a person in the loop. GDPR gives people rights around certain fully automated decisions that significantly affect them, and the EU AI Act requires human oversight for high-risk AI systems. Neither means approving every action. Check what applies to your product before you settle the levels.
We Design Agent Authority at Creative Glue Lab
At Creative Glue Lab, we help product teams decide what an agent may do, where it needs a human decision, and what people need to see before they let it act. Our services cover AI Product Design and AI Development, so strategy, UX, guardrails, and technical enforcement can be reviewed as one system.
If you are deciding how much authority to give an agent, we can create the action map, set the level for each action, define the approval rules, and test the controls in the real workflow. Book a clarity call.





