Build case study
An AI Executive Assistant Built on Claude, Accepted in 22 Days
22 days from signing to acceptance · 11 capabilities tested live · 3 inboxes and 3 calendars in one morning page
Anonymized. Every figure dated and sourced. Published 2026-09-29, updated 2026-10-02.
Want an assistant like this set up for your business? Book a short call and I’ll tell you what it would take for your team.
The short version
The owner of a small, owner-led company was about to lose the assistant who had run the inbox, calendar and follow-ups for years. We built an AI executive assistant on Claude, a custom Claude plugin that took over that work: one morning page that leads with what the owner owes, three inboxes sorted into lanes, waiting-on and going-quiet tracking, and drafts on request that never send on their own. It ran on the owner’s own machine six days after signing and was accepted 22 days after signing, tested against eleven written capabilities.
22 days
signing to acceptance
11
capabilities in the written test
3 + 3
inboxes and calendars, one morning page
The situation
The owner runs a small company with a large network of outside relationships. Mail arrived in three inboxes, meetings sat in three calendars, and the assistant was the system: the one who knew who mattered, what was owed, and which thread had gone cold. The assistant left behind a detailed written handoff but was not available to train anyone.
The owner already used Claude daily and had written prompt workflows of their own. What was missing was the architecture: a place for the rules to live, a routine that ran without being asked, and a way to change it without calling a developer. The goal the owner set was a clean starting point going forward, not a cleanup of years of old mail.
Where this transfers. Any owner, partner or executive whose assistant is leaving, principals of small professional firms, and anyone running more than one inbox who spends the first hour of the day working out what matters.
What was built
One Claude plugin on the owner’s desktop, connected to the owner’s accounts through a single connector hub, with a small database behind a read-only dashboard the owner opens on a phone. Every account was created in the owner’s name and authorized by the owner at the owner’s own keyboard. We never saw a password.
In order: the sources, what Claude does with them, and what the owner sees
Three inboxes
two business lines and personal
Three calendars
connected accounts
Connector hub
one authorization per account
Claude plugin
lanes, roster, obligations, memory
Scheduled runs
four tasks plus an hourly refresh
Morning page
what is owed, what is due, who wrote
Phone dashboard
read-only
Drafts on request
owner approves every send
| Part | What it does | What it replaced |
|---|---|---|
| Morning page | Leads with what the owner owes and what is due, then the day’s calendar, then who reached out. Built on a schedule and on demand | The assistant’s morning rundown |
| Lanes | Sorts mail from all three inboxes into two business lanes and a personal lane | Three inboxes read separately |
| Priority roster | Contacts in four tiers (VIP, queue, later, unknown). The owner moves a person in plain English | Knowledge that lived in one person’s head |
| Waiting-on and going quiet | What the owner owes, what others owe the owner (read from sent mail), and live threads that have stalled past a threshold | Follow-ups that depended on memory |
| Drafts on request | Pulls the thread history and drafts from the correct sending identity. Never sends | Replies started from a blank screen |
| Focus mode | Pauses delivery while collection and sorting continue in the background | Turning everything off and catching up later |
| Scheduled runs | Four scheduled tasks plus an hourly refresh | The assistant checking in through the day |
| Dashboard | Read-only view of lanes, obligations and calendar, readable on a phone | Asking someone what happened today |
| Memory and system map | Business background, lanes, projects and people that persist between sessions, plus a prompt that surveys the workspace and writes its own map | A binder nobody updates |
A day in the life
Early morning. A scheduled run pulls overnight mail and calendar changes from every connected account, sorts them into lanes, updates obligations, and writes the morning page. The owner reads it on the desktop or on the dashboard from a phone.
Through the day. The hourly refresh keeps the dashboard current. When a thread needs an answer, the owner names it in plain English and gets a draft from the right address with the history attached, then edits and sends it personally.
When the owner needs quiet. Focus mode goes on. Mail keeps being sorted and tracked, and nothing interrupts until it goes off.
When something changes. The owner moves a person between tiers, adds a project, or changes the going-quiet threshold by asking Claude to do it. The change is written to a file the owner can open and read.
What it took and how it was accepted
Nine working calls over about four weeks, roughly 14 hours in total, with the owner driving the screen in almost all of them. A developer built the database, the dashboard and the upgrade path behind the plugin. The owner supplied the written handoff, authorized each connection, and set the rules.
| Measure | Result | When |
|---|---|---|
| Signing to running on the owner’s own machine | 6 days | August 2026 |
| Signing to acceptance | 22 days | September 2026 |
| Written acceptance standard | 11 capabilities, each with a live test | Agreed before the build |
| Owner’s own 14-point check on final demo day | 9 clean, 4 partial, 1 open | September 2026 |
| Version 2 upgrade, run by the owner in place | 59-file backup first, zero patch conflicts, no re-setup | August 2026 |
| Obligations list after the upgrade | About a third smaller, because version 1 was counting noise | August 2026 |
What made it work
The finish line was written before the build. Eleven capabilities, each with a pass test, laid out on the headings of the owner’s own handoff document. The owner then wrote a personal checklist from it and graded the final demo.
Lead with what you owe, not with what arrived. The first design sorted mail. The owner’s day is driven by obligations, so the morning page was rebuilt to open with what is owed and what is due, and lanes became a filter across it.
The owner drives. In the sessions the owner’s hands were on the keyboard, not ours. By the end the owner could explain approval tiers, when a change needs a plugin update, and where each rule lives.
Owned from day one. Accounts in the owner’s name, connections authorized by the owner, no passwords shared, and a system map the owner can regenerate at any time.
Upgrade in place. When version 1 proved it was storing summaries instead of full messages, the fix shipped as one command the owner ran, with a backup first, rather than a rebuild.
Honest limits
Drafting was the last capability demonstrated and is still on request only; learned voice and proactive drafting were out of scope. The owner’s live personal calendar sits with a second provider, so the unified view covers the connected calendars until that history is imported. At close-out the scheduled tasks were found paused; they were re-enabled on the call and the cause is still being tested. At the time of the build, scheduled tasks ran on the owner’s computer, so it had to stay awake with Claude open. One mailbox arrives through a relay that hides the original sender from the API; the fix is connecting the source account directly, which is new work. The architecture was re-planned twice in the first week, and the sessions ran longer than scheduled. The import of older mail archives is being handled by the owner.
Similar situation?
If your assistant is leaving, or your day starts with three inboxes and a guess, this is what a custom Claude build can look like. Bring the handoff notes or one ordinary week of mail and we will start there.
Book a short call →Questions people ask
Can Claude be an AI executive assistant?
Yes, for the reading, sorting, tracking and drafting that fill most of an assistant’s day. In this build Claude sorts email from three inboxes into lanes, tracks what the owner owes and is owed, flags threads that have gone quiet, and drafts replies on request from the right address. It does not send anything or make judgment calls on its own; the owner approves every message.
How do you build an AI executive assistant with Claude?
Write down what done looks like first, then build to it. Here that meant eleven capabilities with a pass test each, one connector hub linking every email and calendar account, and a custom Claude plugin holding the lanes, priority roster, rules and memory. Scheduled tasks produce the morning page, and the owner changes the rules in plain English. A developer added the database and phone dashboard; a simpler version can run on the plugin and connectors alone.
Can a Claude AI executive assistant manage email and calendars across multiple accounts?
Yes. This AI email and calendar assistant covers three inboxes and three calendars in one morning page, each account authorized by the owner through one connector hub. Every service is in the owner’s name and on the owner’s billing, the data sits in the owner’s own database, and the phone dashboard is read-only.
Keep reading
Timesheet and Invoicing Automation With Claude: 159 of 159 Matched on the First Full RunNew
AI Permit Tracking for an Engineering Firm: Every Agency Portal, One Workbook, 617 Documents FiledRelated
Inside a Custom Claude Plugin Built for a Small Business OwnerRelated
Custom Claude Plugin DevelopmentRelated
Claude Workspace Setup, Done for YouRelated
Inside a Custom Claude Plugin Built for a Small Business OwnerRelated
All case studies