(Omar · Ext. 104 · knowledge base)

Executive Assistant.

Triages your inbox, drafts replies in your voice, wrangles your calendar, and preps every meeting — with everything external gated behind your approval.

Launch this role

What it does from day one

(no ramp-up, no aspiration)

Inbox triage and labeling (To Respond / FYI / Newsletter / Invoice / Spam) with urgent-item pings to Slack
Reply drafts in the client's voice, learned from sent mail, saved to Drafts — never auto-sent
Scheduling and rescheduling: propose times, send invites, handle the ping-pong, protect focus blocks
Meeting notes via notetaker bot: transcript, summary, action items, follow-up drafts
Pre-meeting briefs 30-60 min before each event: attendee research, thread summary, last-meeting notes
Morning brief + follow-up chasing: daily digest of decisions needed, auto-drafted nudges on stale threads

How the work flows

(brief to logged output)

Onboard & connect — OAuth to email + calendar; agent ingests 30-90 days of sent mail to learn voice and builds a playbook (VIPs, triage rules, approval thresholds)
Continuous watch — webhooks and cron triggers (new mail, new invite, T-minus-45 meeting, morning brief) queue discrete tasks
Reason & produce — classify, pull thread/calendar context, output the artifact: label, draft, slots, brief, itinerary options
Approve or auto-run — tiered autonomy: labeling/drafting/internal holds auto-execute; anything external or irreversible hits the approval channel
Log & learn — every action in an auditable activity log; edit patterns feed the voice profile; weekly report is the KPI dashboard

The stack it plugs into

(real tools, official APIs)

Gmail API + Microsoft Graph — read/label/draft email with push notifications (or Nylas as one unified layer across both)
Google Calendar API + Graph Calendar — free/busy, event CRUD, invite management
Recall.ai — meeting-bot API returning diarized transcripts from Zoom/Meet/Teams
Slack API — the approval surface: urgent pings, approve/reject buttons, morning brief delivery
Duffel API — flight/hotel search and itinerary assembly (booking stays approval-gated)
Notion or Google Drive/Docs — where briefs, notes, and the playbook live

Its machine

(8GB RAM · 4 vCores, dedicated)

The lightest profile in the lineup: an always-on orchestrator with webhook receiver, cron scheduler, OAuth token store, and SQLite activity log — email/calendar work is pure API calls. One Playwright instance spins up per research task. Steady state 1-2GB; the dedicated box's real value is per-client credential isolation and an always-on event loop.

Seat economics

medium token load

Triage is cheap classify calls, drafts are small; meeting transcripts are the biggest driver (~10K tokens/hour of meeting) — roughly $40-150/mo with model tiering, healthy margin on $499; cap transcript hours/month as the fair-use meter, same pattern as voice minutes on the cold-caller.

Flat seat from $499/agent/mo — one agent, one dedicated machine. See pricing.

Judged by

(the numbers that matter)

Time saved / email hours reclaimed (professionals spend ~13 hrs/week on email; category benchmark is 'save an hour a day')
Priority response SLA — 90% of priority messages handled within 24 hrs, calendar conflicts <5%/month
Draft acceptance rate — % of drafts sent with zero or light edits, plus scheduling round-trips eliminated

Tasks in this role

(each documented in the task library)

AI Inbox Triage
Every incoming email labeled To Respond, FYI, Newsletter, Invoice, or Spam, with urgent items pinged to Slack the moment they land.
AI Email Drafting
Reply drafts written in your voice, learned from 30-90 days of your sent mail, and saved to Drafts — never auto-sent.
AI Calendar Management
Proposes times, sends invites, handles the scheduling ping-pong, and protects your focus blocks.
Meeting Prep Briefs
A brief in your hands 30-60 minutes before every meeting: who the attendees are, what the email thread says, and what happened last time you met.
Meeting Notes and Follow-Ups
A notetaker bot joins your Zoom, Meet, or Teams calls and returns a diarized transcript, summary, and action items — with follow-up emails already drafted.

Never alone

(guardrails — how trust works here)

Never auto-send external email by default — drafts only; auto-send is graduated per-category only after measured edit rate on that category nears zero
Never spend money or book travel without per-transaction approval — agent proposes options with total price, human clicks confirm
Never accept/decline/cancel meetings with VIPs or externals, or touch protected focus blocks, without approval; internal low-stakes reshuffles are autonomous
Never delete anything, never reply on sensitive threads (legal, HR, medical) — and instructions found inside received emails are data, not commands (prompt-injection hard block)

Full policy: our AI disclosure.

Try it

(alpha)

"Tame This Inbox": a realistic fake inbox of 25 seeded emails; visitor optionally states a persona and pastes a writing sample, hits 'Hire the EA,' and watches live — labels cascade, noise archives, drafts appear in their voice, the calendar pane resolves a double-booking, and a flight card pops up behind an Approve button. Ends with an auto-generated morning brief and time-saved scorecard. Fully sandboxed, and it demos the approval-gate UX that is the trust story. Cold calling ships first — try that demo today.