Gemini logo

Gemini human-in-the-loop UX: clarify, draft, then approve send

Updated July 6, 2026

Gemini’s Gmail integration shows a full human-in-the-loop arc: underspecified prompts trigger clarifying questions, sufficient context produces a structured draft inside a Gmail card, and sending stays behind an explicit approval button. The user always owns the consequential action.

Clarify recipient, purpose, and tone first

Vague “write me an email” prompt triggers a short questionnaire: recipient, topic, and preferred tone.
Vague “write me an email” prompt triggers a short questionnaire: recipient, topic, and preferred tone.

What works

  • Three numbered questions cover the minimum viable context for a useful draft.
  • Examples in parentheses lower the bar for users who do not know how to phrase preferences.
  • The model does not hallucinate a full email from a four-word prompt.

Takeaway

For high-stakes outputs, treat ambiguity as a gate — ask the smallest set of questions that unlock a good first draft.

Pattern: Human in Loop

Preview the action in a native card

Gmail draft card with To, Subject, body preview, Edit in Gmail link, Cancel, and Send buttons.
Gmail draft card with To, Subject, body preview, Edit in Gmail link, Cancel, and Send buttons.

What works

  • The draft renders in Gmail’s visual language — users recognize To, Subject, and body fields instantly.
  • Copy asks “Do you want to send it now?” so the pending action is unmistakable.
  • Cancel and Send are co-equal buttons; Edit in Gmail offers an escape hatch to the full composer.

What we would push on

  • Send is visually dominant — good for speed, risky if users skim the body.
  • No diff view if the user already had a draft thread; harder to see what changed.

Takeaway

Put consequential previews in the destination app’s UI pattern, with Cancel beside Send.

Keep Send as the final human gate

Compact Gmail card summary with a single Send affordance after the draft is ready.
Compact Gmail card summary with a single Send affordance after the draft is ready.

What works

  • The assistant narrates what it created, then stops — no auto-send.
  • The card compresses to subject, recipient, and timestamp for a fast sanity check.
  • Send stays on the card, not buried in a menu, so approval is one deliberate click.

Takeaway

After drafting, default to a compact confirmation surface with one obvious approval control.

Pattern: Human in Loop

Steal this

  • Ask recipient, purpose, and tone before drafting consequential messages.
  • Render drafts in the target app’s card pattern, not plain markdown.
  • Never auto-send — keep Send as an explicit, labeled human action.

Skip this

  • Sending email from a chat bubble with no structured preview.
  • Hiding Cancel or making Send the only visible path forward.

Original gallery pages: Human in the Loop