Home/Agents/Inbox Triage
Smart Inbox · Production · v4.2

Your inbox, triaged in 8 seconds — drafts ready, CRM updated.

Watches Gmail or Outlook. Claude Haiku classifies every email (support / sales / invoice / urgent / noise). Claude Sonnet drafts a reply in your voice. WhatsApp shows the draft — reply 1 to send. CRM gets updated automatically. $0.10 per email resolved. Noise is free.

  • $0.10per email resolved
  • $0.50per lead routed to CRM
  • $2per support ticket closed
  • $0for filtered noise
Deploy this agent →See full pricing
How it runs

The loop, end to end.

01

Watch

Gmail push notifications or Outlook webhooks deliver new mail in <2s. Agent decides: noise (newsletter, unsubscribe, internal cc) → drop; signal → continue.

02

Classify

Claude Haiku (cheap, fast) labels the email: support, sales lead, invoice/AP, partnership, billing question, urgent. Confidence below 80% → human review.

03

Draft

Claude Sonnet drafts a reply in your voice, pulling context from past threads, your CRM, and your knowledge base (RAG). Stays under 150 words by default. No corporate-speak.

04

Send + log

WhatsApp shows draft preview. Reply 1 = send, 2 = edit, 3 = skip. CRM auto-updated with classification + outcome. Audit chain records the send.

Inside the agent loop

The tools Claude actually calls.

→ tool_use: gmail.watch_subscription()
← message_id 'gm_482c' delivered
→ tool_use: claude_haiku.classify(subject=..., body=...)
← {label: 'sales_lead', confidence: 0.92, urgency: 'medium'}
→ tool_use: hubspot.lookup_contact(email='sender@acme.com')
← existing contact, last touch 47 days ago
→ tool_use: rag.knowledge_lookup(query='pricing tier comparison')
← 3 docs returned, prior threads with sender included
→ tool_use: claude_sonnet.draft_reply(context=..., voice_sample=...)
← draft generated, 118 words, cost $0.011
→ tool_use: twilio.whatsapp_approval(to='+1...', preview=draft)
← user reply '1' received
→ tool_use: gmail.send_reply(thread='gm_482c', body=draft)
→ tool_use: hubspot.update_contact(id='c_99', last_touch=now, status='engaged')
→ tool_use: outcome_ledger.record(event='lead_routed', value=0.50)

Real tool-call sequence from a recent production run. Names redacted.

Integrations included

Native MCP. No custom wrappers.

  • Gmail / Google Workspace
  • Outlook / Microsoft 365
  • HubSpot
  • Salesforce
  • Pipedrive
  • Zoho CRM
  • Intercom (support routing)
  • Zendesk (ticket creation)
  • Slack (escalations)
  • WhatsApp Business API
  • Notion (knowledge base)
  • Custom CRM via webhook
What this agent will NOT do
  • — Marketing email blasts (use Mailchimp/Customer.io)
  • — Calendar scheduling (handled by recruiter agent or your scheduling tool)
  • — Contract redlining (manual review territory)
Before · After

Your inbox-agent process, redrawn.

Before

Manual triage: 5 steps, hours per day

  1. 01
    Read every subject line, decide priority
    👤 Owner · ⏱ 30 min
  2. 02
    Open, read, classify each signal email
    👤 Owner · ⏱ 45 min
  3. 03
    Draft reply from scratch or template
    👤 Owner · ⏱ 30 min
  4. 04
    Forward or route sales leads to CRM manually
    👤 Owner · ⏱ 10 min
  5. 05
    File / archive / snooze remaining
    👤 Owner · ⏱ 5 min
Total time
2 hrs / day
Handoffs
3
Humans
2
After · with rpa-automate

Agentic triage: 4 steps, 8 seconds each

  1. 01
    Email arrives → Claude Haiku classifies
    🤖 Inbox agent · ⏱ 1.5 sec
  2. 02
    Below 80% confidence → surface for human review
    🤖 Inbox agent · ⏱ instant
  3. 03
    Claude Sonnet drafts reply in your voice + RAG context
    🤖 Inbox agent · ⏱ 3 sec
  4. 04
    WhatsApp preview → reply 1 to send → CRM auto-updates
    🤖 You + agent · ⏱ 1 tap
Total time
8 sec / email
Handoffs
1
Humans
1
ROI · inbox-agent

Do the math with your numbers.

emails / mo
$
In-house monthly cost
$3,010
1,800 × $1.67
With rpa-automate
$180
1,800 × $0.10 / email resolved
You save
$2,830
94% cheaper / month

Assumes linear scaling and no ramp-up. Real deployments hit steady-state around week 3–4.

Also see: Automate Inbox Triage → — the same math from a process-first angle, for buyers thinking "we need to fix our inbox triage process" before shopping for an agent.

FAQ

Will it auto-send without my approval?

Only if you explicitly enable autonomous mode for a specific email category (e.g., 'auto-reply to known-vendor invoice acknowledgements'). Default is human-in-the-loop for every send. Every send is hash-chained for audit.

What about email I don't want it to touch?

Set excluded senders (your boss, your spouse) or excluded threads (the offer letter negotiation). The agent skips those and shows them in your normal inbox unchanged.

How does it learn my voice?

On day 1 we point Claude at your last 60 days of sent mail (Gmail or Outlook). It builds a voice fingerprint — sentence length, common phrases, salutation style, sign-off patterns. You can correct it by replying 2 (edit) — the edits feed back into the voice sample.

Is my email data sent to Anthropic for training?

No. We use the Anthropic API with the no-training contractual default. Your email content is sent for inference only, not retained, not used for training. See /compliance for full posture.

How does this compare to Superhuman AI or Shortwave?

Superhuman and Shortwave are AI email clients — you still operate the inbox. This is an agent — it works your inbox while you do other things, and surfaces only what needs your judgment. Outcome-priced instead of seat-priced. See /vs/lindy for the closest agentic comparison.

Want to see it run on your data?

A $0 outcome audit takes 3 minutes. We'll show you the outcome math, the integrations, and what week-one looks like — with your real systems.

Book the audit →

Pair this agent with

Recruiter →Bookkeeping →Voice →Why Claude →Compliance →Exceptions →By process →