Human Review
Put a person between AI output and real-world consequences, with review that's fast for good output and careful for risky output.
7 states5 edge casesDesktop + Mobile
01The problem
AI can draft the email, categorise the transaction or propose the database change — but sending, booking or executing it has consequences. Reviewing everything defeats the point of automation; reviewing nothing lets one confident mistake reach a customer. Reviewers need to see what the AI is about to do, why, and what it's unsure about, then approve, edit or reject quickly.
02When to use
Use when
- AI drafts that are sent to customers: emails, replies, documents
- Agents that take actions: refunds, bookings, record updates, code changes
- Classification or extraction feeding financial or legal records
Avoid when
- Output is only seen by the person who asked for it — they review it naturally.
- Actions are trivially reversible and low-impact; log them and allow undo instead.
03Recommended UX
Statecraft recommendation
Show the proposed action as a concrete preview — exactly what will be sent or changed — with the AI's sources and the fields it's least certain about highlighted.
04Desktop behaviour
On desktop, review is a queue: the list of pending items on the left, the full preview and decision on the right, with keyboard shortcuts for throughput.
Desktop rules
- Preview the exact output (rendered email, record diff), not a summary of it.
- Highlight low-confidence fields and link each claim to its source.
- Approve (A), Edit (E), Reject (R) and Next (J) by keyboard for high-volume queues.
- Show why each item needed review: “Amount above $500” or “Low confidence on customer name”.
05Mobile behaviour
On mobile, reviewers handle one item at a time, often from a notification, so each card must be decidable on its own.
Mobile rules
3 mobile rules are included in Pro.
Statecraft Pro
Unlock the full pattern
Get the complete interaction flow, states, edge cases, mobile rules and implementation prompt.
- Interaction flow · 6 steps
- 3 mobile rules
- 7 states
- 5 edge cases with recommendations
- 4 microcopy comparisons
- Accessibility notes
- Developer notes
- Claude Code / Codex implementation prompt
Edge cases covered in this pattern
- Rubber-stamping
- Context changed before approval
- Reviewer lacks domain knowledge
- Action fails after approval
- Nobody reviews