How an app shows AI at work, and hands its output to a manager to check.
Many Atrium apps use AI to draft replies, summarise records or suggest next steps. Managers stay in charge: the AI drafts, the manager decides. Every AI moment follows the same three steps, so managers recognise them in any app.
| Step | What the manager sees |
|---|---|
| 1. Ask | A button that names what the AI will do: "Draft reply", "Summarise". Not "Ask AI" or "Magic". |
| 2. Generating | An AI content block with the moving border and the label "Writing". Keep the rest of the screen usable. |
| 3. Check and use | The output marked "Drafted by AI", with Edit and a primary action that names the result: "Use reply". Nothing is sent or saved until the manager presses it. |
Generating
Example
Example
HTML
<div class="at-ai at-ai--generating" aria-busy="true">
<p class="at-ai__label">Writing</p>
<div class="at-stack at-stack--tight"><span class="at-skeleton"></span><span class="at-skeleton at-skeleton--medium"></span></div>
</div>Ready to check
Example
Drafted by AI
Hi Sam, the boiler engineer is booked for Thursday between 9am and 1pm. They will call 30 minutes before arriving.
Example
Drafted by AI
Hi Sam, the boiler engineer is booked for Thursday between 9am and 1pm. They will call 30 minutes before arriving.
HTML
<div class="at-ai">
<p class="at-ai__label">Drafted by AI</p>
<p>Hi Sam, the boiler engineer is booked for Thursday between 9am and 1pm. They will call 30 minutes before arriving.</p>
<div class="at-inline at-inline--end">
<button type="button" class="at-button">Discard</button>
<button type="button" class="at-button">Edit</button>
<button type="button" class="at-button at-button--primary">Use reply</button>
</div>
</div>Rules
- A person presses the button. AI output is never sent to a tenant, saved to a record or used to change money without a manager choosing to.
- Always marked. Keep the AI content marking until the manager has used or approved the output.
- Always editable. Every output has Edit and Discard.
- Honest when unsure. If the AI could be wrong, say so: "Check the dates before sending."
- Errors in plain words. "Could not draft a reply. Try again, or write one yourself." Never show model or API errors.