AI-assisted workflow surfaces
AI-assisted workflow surfaces are the points where a model’s output enters a business process: a suggested reorder, a drafted document, a classified alert, a review comment.
The design work concerns how that output becomes a record a person is accountable for.
Key facts
- Model output enters a process in three modes: suggestion, draft or action.
- The mode decides who decides: the user, the user after review, or the model with review afterwards.
- Stored values should show whether a person or a model created them, and whether a person reviewed them.
- Model output written into a field without a mark becomes indistinguishable from typed input.
What it covers
| Mode | Who decides | Interface needs |
|---|---|---|
| Suggestion | User | Reasoning and one-step accept or reject |
| Draft | User approves | Editing, comparison with the draft, approval |
| Action | Model executes | Review after execution, undo, audit |
Examples in business software: reorder suggestions, schedule changes, data extracted from documents into a form, drafted reports, classified alerts, review comments on code.
Who uses it
| Role | What they need from the surface |
|---|---|
| Everyday user | To see the suggestion where the decision is made, and accept or correct it quickly |
| Reviewer or approver | Evidence for each output and a clear record of the decision |
| Auditor | Which values a model produced and who reviewed them |
| Product team | Data on acceptance, edits and rejections to tune the feature |
What makes it hard to design
Evidence. Each output should show what it is based on. Generic explanations that point to no data give the reviewer nothing to check.
Authorship. Stored values should show whether a person or a model created them, and whether a person reviewed them.
Fitting the existing workflow. Suggestions placed inside the screens where decisions are made reach users at the moment of the decision. A separate panel asks users to leave their task.
Models return partial, low-confidence or empty results. The surface needs a designed state for each, and the manual path has to keep working.
Trade-offs
- More autonomy saves clicks, and errors travel further before anyone sees them. Each extra review step keeps control and eats into the time saving that justified the feature.
Common mistakes
- Model output written directly into a field and indistinguishable from typed input. Months later nobody knows which values were reviewed.
- Hiding low confidence to keep the screen clean
- One confirmation dialog for every action, whatever its consequence
- No manual path when the model fails
Related experience