When you open an inbound email, up to five context-aware chips appear below the body showing the most likely reply actions for that specific email. Labels are generated from the email content (e.g. “Accept the meeting”, “Push back on timeline”, “Send the proposal”). Two universal chips (Quick thanks and Custom…) are always appended to the end of the row.

Clicking a chip opens the inline reply and streams an AI-generated draft character-by-character. Edit the draft, or click Send.

A tone strip sits above the chip row, so you pick the voice first and then fire generation by clicking a chip. Tone applies to the chip you click. The strip scrolls horizontally at every screen width (it never wraps to a second row), and your own default tone is always the first pill, so the tone you use most is the one under your thumb.

After your default come the plain tones (Customer-friendly, Match sender, Warm, Professional, Direct, Formal, Casual, Brief), then the three that depend on extra context: Match my style, Match recipient, and Brand. Brand only appears once an admin has set a brand voice for your workspace. At the end of the strip, Custom… opens a short text box for a one-off tone such as “terse and friendly”.

The AI analyzes your last 30 sent emails (from the connected Gmail/Outlook account) to extract your voice — preferred greetings, signoffs, formality, and tone. Replies are written to match. Refresh or delete the profile anytime from Settings → My preferences → Profile → Data controls.

• Regenerate chips: click the circular refresh icon at the end of the chip row to ask the AI for a fresh set of suggestions for the same email.
• Custom prompt: click Custom… to open a text input for a free-form instruction (for example “Reply in German with an apology for the delay”). Uses the same tone pill you have selected.
• Pop out to full compose: if you need attachments, CC tweaks, or more editing room, pop the streamed draft out of the inline reply into the full compose window — the AI-written body is preserved.

Press 1–5 to trigger the first five chips without reaching for the mouse. Press Alt+T to cycle through tones. Press Esc during generation to stop — partial text is preserved so you can finish editing without losing what streamed. Shortcuts only fire when chips are visible and you are not typing in an input.

The AI sees the last 3 messages of the thread verbatim plus an AI-generated summary of up to 7 older messages. For long threads, this means replies can reference context from much earlier in the conversation.

• Calendar invites: chip set swaps to Accept / Decline politely / Propose new time / Ask for agenda. These are text replies — they don’t create real calendar invites.
• Marketing / newsletters: suggestions are not generated automatically (writing five reply options for a newsletter nobody intends to answer is wasted AI spend). Instead a Generate AI suggestions button appears in place of the chip row. Click it and you get the full chip set for that email, exactly as you would anywhere else. Replying is never blocked: a chip you click, a custom prompt, and the reply editor itself all work normally on a promotional email.
• Forwards: the AI knows the reply goes to the forwarder, not the original sender.
• Unmatched sender: chips still appear but with generic context. An Add contact link appears below the chip row so future replies can use CRM data.
• Outbound emails: chips do not appear on emails you sent. They are only shown for inbound messages.
• Bulk-select mode: when you are selecting multiple emails, chips are hidden to keep the list action bar clean.

• AI is not configured: an “Enable AI reply suggestions” card replaces the chip row. Admins can click through to the provider setup; non-admins see a note to contact an admin.
• Rate limit reached: an amber “AI rate limit reached. Try again later.” banner appears. See the rate limits below.
• Empty email: an email with no body and no preview text gets no chip row at all. There is nothing for the AI to read.
• Marketing email: the chip row is replaced by a Generate AI suggestions button. This is a cost guard on the automatic fetch, not a refusal. Click the button and the suggestions arrive.
• Generation failed: if the LLM returns no usable chips, the email falls back to a static set (Acknowledge, Ask for details, Propose a time, Provide info, Decline politely) plus Quick thanks and Custom. A “Using default suggestions (AI unavailable)” note appears.

• Email body content is stored in your CRM to enable search and fast AI replies.
• Bodies are encrypted at rest and purged when you delete emails.
• When you click a chip, the email body + CRM context is sent to our AI provider to generate the reply.
• Full data controls: Data controls. Admin controls: Data privacy.

Each AI endpoint has its own rolling-hour bucket per user — hitting the chip cap does not consume your chat, agent-run, or draft allowance. Buckets are distributed via Redis so scaling up serverless instances doesn't multiply the effective limit.

• Chip generation: 60 per rolling hour per user
• Reply generation: 20 per rolling hour per user
• Writing style refresh: 3 per rolling hour per user

When a bucket is exhausted, the endpoint returns an HTTP 429 with a Retry-After header (seconds until the next slot opens) and an error body { error: '…', code: 'RATE_LIMITED' }. The UI surfaces an amber “AI rate limit reached. Try again in N seconds.” banner. Usage also counts against your org's AI cost budget — if the budget is reached you'll see a separate “Budget reached. Turn on Extra AI Credits or upgrade” message with a Top up or Upgrade CTA.