1. Moving from AI approval to AI review workflow
The assumption behind this design is that document editing itself is already a mature experience. Users are already familiar with creating and editing documents through tools like Notion and Google Docs. The challenge is not creating another editing workflow, but helping users understand and manage changes introduced by an AI agent.
Instead of requiring users to approve every AI-generated change before it is applied, I designed the experience around an AI review workflow. The agent completes the editing task first, and users review the updated document afterward. This allows AI to make larger edits efficiently while giving users the ability to keep, modify, or revert changes based on their intent.
We originally asked users to approve every AI-generated change before applying it. That slowed the flow and limited the size of edits the agent could make. Shifting to an AI review workflow lets the agent finish the work up front while users decide afterward what to keep, modify, or revert.
2. Making the current document the primary review surface
A traditional side-by-side diff view is useful for code review, but document editing requires more context. Users need to understand how the updated document reads as a whole, not only compare individual changes.
Therefore, I prioritized the current document view over a permanent comparison layout. The latest version remains the main canvas, while previous versions are available when users need additional context. This helps users evaluate the final outcome first and investigate specific changes only when necessary.
Nothing is lost by keeping the current version in front. Every earlier version of a paragraph stays one click away in the comparison panel, and reverting puts it back exactly as it was written.
3. Revealing change details progressively
Showing every AI change, explanation, and comparison state upfront can make the experience feel complex and distract users from the document itself. The design uses progressive disclosure to balance simplicity and transparency.
The document remains unclutteredclean by default, while users can identify that changes exist and open detailed comparisons when they need more information. This supports both technical users who want to inspect changes and non-technical users who prefer a calmersimpler experience.
4. Treating AI as a collaborative editor instead of reinventing the editing experience
The goal of this experience is not to create a completely new AI-specific workflow, but to make AI collaboration feel like a natural extension of existing document editing behaviors.
Users already understand the mental model of working with documents: editing content, reviewing updates, making adjustments, and saving the final version. Instead of introducing unfamiliar concepts around AI approvals or agent actions, I treated the AI agent as another editor contributing to the document.
By building on existing editing patterns, the experience feels familiar for both technical users and users who are comfortable with tools like Notion and Google Docs. AI enhances the existing workflow rather than requiring users to learn a new way of working.
5. Keeping review, editing, and recovery in the same experience
The document remains editable throughout the workflow, allowing users to review AI changes, modify content directly, and save the final version without switching contexts.
Because AI edits can affect multiple parts of a document, the experience also supports reversible version switching before saving. Users can explore changes, switch between versions, and continue editing with confidence. This keeps AI collaboration flexible while preserving user ownership of the final document.