The right to change your mind

Ask a calendar assistant to clear two hours on Thursday afternoon. It moves three meetings and sends updated invitations. Only then do you notice that one of them was the conversation you needed before a difficult decision.

A preview might have helped. You might also have approved the plan and regretted it later. Seeing a proposal can clarify what matters to you, including something you left out of the request.

Now try to correct it. Can you stop the assistant, keep that meeting in place, and find out which invitations have already gone out? Changing your mind should be an ordinary part of using the product.

Agree on the scope

In 1999, Eric Horvitz argued that automated assistance should account for uncertainty about people's goals, allow people to stop it, and narrow its actions when confidence was low. His paper¹ used a meeting assistant to illustrate the principles. Today's agents face the same decisions about when to act and when to ask.

The calendar task needs a small agreement to start. The assistant can inspect Thursday's schedule and move personal focus blocks. Meetings with other people go into a proposal before invitations are sent. Show this scope beside the work, let the person change it there, and end the permission when the task finishes.

The system must enforce the scope the interface describes. A July 2026 survey by Alexandra E. Michael and Franziska Roesner² examined 21 agent permission proposals and five commercial agents. They found gaps between permission interfaces, enforcement, and ongoing user control. The preprint covers a particular set of systems, but its concern is useful to any team building this flow: can the assistant do something the person was told it could not do?

Make the proposal editable

Show the original and proposed times together, who will be notified, and the reason for each move. The person can inspect that list and change one item. If the preview only says the schedule has been optimized, they have to check the calendar from scratch.

Amershi and colleagues' 2019 guidelines³ address dismissal, correction, and explaining the consequences of user actions. Their evaluation involved 49 design practitioners inspecting 20 AI products. The guidelines offer a starting point; the controls still need testing in the product where people will use them.

Keeping the important meeting in place should preserve the useful parts of the plan. Let the assistant find another arrangement around it. Repeating every constraint in a new conversation makes a small correction feel like beginning again.

One confirmation can cover the agreed changes: these meetings will move, and these people will receive invitations. Reading availability and calculating alternatives can continue within the existing scope. Ask again when the action would change that agreement or introduce a consequence the proposal left out.

Every question demands attention. In a 2021 experiment with 199 participants⁴, Zana Buçinca and colleagues tested a food substitution task using simulated AI. Interfaces that encouraged more deliberate thought reduced reliance on incorrect suggestions compared with simple explanations. They did not significantly improve overall performance over those explanations, and the most effective designs received less favorable subjective ratings. That narrow result is a reason to test what an extra prompt achieves before adding it.

In its February 2026 analysis of agent use⁵, Anthropic found that more experienced Claude Code users enabled automatic approval more often and also interrupted more often. These observations come from one provider, largely in coding, and do not establish that experience caused either behavior. They do show people delegating work while continuing to intervene in it.

Make recovery understandable

If the person says, “Keep the one-to-one where it is,” pause pending changes and show what has already happened. Revise the remaining plan from there. If an invitation is already being sent and cannot be canceled, keep its status visible until the result is known.

A calendar entry can be moved back, but people may already have read the invitation and arranged their day around it. Explain the recovery available: restore the original time and send another update. Calling the whole thing undone would hide what other people have already seen.

Recovery also has to account for newer work. If a colleague has since proposed a different time, restoring an old snapshot could erase their proposal. Preserve it and show the conflict so the person can decide which time to use.

Keeping this Thursday's meeting in place should not silently become a permanent rule against moving one-to-ones. If the assistant offers to remember the preference, show exactly what it will remember and let the person edit it. An exception made in a busy week may only belong to that week.

To test the workflow, give someone a reason to change their mind halfway through. Include a partially completed action. Watch whether they notice the change, correct it, and understand what is still in motion. Completion time alone misses the work of getting back to an acceptable result.

The assistant may have followed the original request exactly. The person still needs room to reconsider it. They should be able to keep the one-to-one and let the assistant arrange the rest.