AI chief of staff: how an owner builds one on Claude
An AI chief of staff reads your inbox and calendar, prepares your meetings and waits for your approval to send.
What an AI chief of staff does for an owner
An AI chief of staff is an assistant connected to your email, calendar and files. Standing instructions tell the assistant which parts of a chief of staff’s work to do. The same setup often goes by the name AI executive assistant.
Calendar tools use that name too. Reclaim says the tool schedules tasks on your calendar by priority, and Motion says the tool plans your day with smart scheduling. An AI chief of staff on Claude also reads the mailbox, the CRM and the ERP behind each meeting.
Sorting the inbox
Each morning, the assistant reads new email and sorts each message: needs your answer, needs someone else, waiting on a reply, or for your information. For each message that needs your answer, the assistant drafts a reply and leaves the draft for you.
The Gmail connector can apply labels, so each group gets a label you choose. Through Gmail, Claude sees an attachment’s name and details but not the content, so open the attachment yourself before you act on a summary.
Preparing meetings from the CRM and the ERP
Before each meeting on the calendar, the assistant gathers what the meeting needs. That means the attendees, the last emails with each one, the notes in the CRM and the open orders in the ERP. The assistant then writes one page on who is coming, what is open and what to ask.
Connectors reach the CRM and the ERP. Claude’s connectors directory lists a HubSpot connector described as useful for pre-call prep, and an Oracle NetSuite connector for analysis of NetSuite data. For an ERP with no listed connector, a custom connector built on an MCP server does the same job.
Writing a weekly brief
A weekly brief collects what happened and what needs a decision, such as sales against plan, late orders and the follow-ups still open. The assistant writes the brief from the ERP, the CRM and the week’s email.
Claude runs the brief as a scheduled task, available on all paid Claude plans. Anthropic lists weekly reports among the common uses, and scheduled tasks run even when your computer is asleep.
Tracking follow-ups
Ask the assistant to list the promises in your sent mail. A promise is a quote you said you would send, or a call you agreed to return. Each Friday, the list shows what is still open and who is waiting. The assistant drafts each reminder, and you send each one yourself.
Start from the work that fills your inbox
Tell Derik which job takes the most hours in your week and who would attend. He will tell you what a session built on that job would cover.
Start a conversationHow to set one up on Claude
- Choose the plan. Scheduled tasks need a paid plan, and on Team or Enterprise an owner enables connectors first.
- Connect email and calendar, through Google Workspace or Microsoft 365.
- Connect the CRM and the ERP, through a listed connector or a custom one.
- Create a project for the chief of staff, and write the instructions.
- Schedule the morning sort and the weekly brief as tasks.
- Set every action that sends, posts, shares or deletes to need your approval.
For what the Team and Enterprise plans include, see Claude for business.
Connecting email and calendar
The Gmail, Google Calendar and Google Drive connectors are available to every Claude user. On Team and Enterprise plans, an owner enables the connectors for the organization first, and each person then signs in with a Google account. If the company runs Google Workspace, the Workspace admin may need to mark Claude as a trusted app.
The Microsoft 365 connector covers Outlook mail and calendar, SharePoint, OneDrive and Teams. The connector needs a work account tied to a Microsoft Entra tenant, and a Global Administrator grants consent once for the company. Sending email and changing the calendar work only after an admin turns on write tools.
Writing the instructions
Create a Claude project for the chief of staff. A project holds a separate chat history, files and project instructions, so every chat in the project starts from the same rules. Keep the project private, because connectors are only available in private projects.
The instructions say how to sort mail, what a meeting page covers and what goes into the weekly brief. Add a list of key customers and suppliers, and the names of the people who report to you. End with the rule that nothing goes out without your approval.
Part of the instructions can read: “Every Monday at 7 a.m., write the weekly brief. List late orders from the ERP export, deals in the CRM that moved, and every promise in my sent mail that is still open. Draft replies, and never send one.”
Which actions keep a human approval
Keep one rule: a named person approves every message that leaves the company. On Claude, that rule lives in the approval settings of each connector.
In Gmail, Claude asks for your approval before each send, reply or forward by default. On Team and Enterprise plans, owners decide whether members can let those actions run without asking. Keep the default.
In Microsoft 365, Teams chat messages and channel posts always ask you to confirm before Claude sends anything. Email that Claude sends through Microsoft 365 carries a header that marks the message as agent-initiated.
Anthropic’s connectors guide says that on Team and Enterprise plans, an owner sets each tool to Always allow, Needs approval or Blocked. Set every tool that sends, posts, shares or deletes to Needs approval, and block the tools the chief of staff never needs.
Scheduled tasks need the same care. Anthropic’s safety guide advises against scheduling tasks that send messages on your behalf, and recommends manual approval when a mistake would be hard to undo. The morning sort and the weekly brief read, label and draft without sending, so they can run on schedule.
The same guide says that you remain responsible for every action Claude takes on your behalf. For where approvals sit in an AI system, see human in the loop.
What the assistant can see once connected
Claude inherits each person’s permissions from the connected service. If you can open a folder, a mailbox or a CRM record, the connector can reach that item. If you cannot, the connector cannot either.
- Shared calendars and shared mailboxes come along when your account can open them.
- Claude reads connected data when you ask, and Anthropic says the Microsoft 365 connector runs no background searches.
- Data that the Google connectors retrieve is stored with the chat, so deleting the chat deletes that data.
- Anthropic says it does not train models on Gmail, Drive or Calendar connector data.
- A custom connector that uses a shared credential reaches whatever that credential can access.
Email is also where outside text reaches the assistant. Anthropic’s safety guide describes prompt injection, where an attacker’s email carries instructions that the assistant may follow while reading the mailbox. Approval on every send stops an injected instruction from sending mail in your name. For more, see prompt injection, explained plainly.
How to limit what the assistant can reach
- Turn connectors on for each chat from the + menu, under Connectors, and leave off the ones a job does not need.
- Turn off individual tools in Customize, under Connectors, in the connector’s Tool permissions.
- On Team and Enterprise plans, restrict actions for the whole organization, such as allowing Claude to summarize email but not send it.
- Fix permissions at the source. The Microsoft 365 connector cannot be limited to a list of SharePoint sites, so the connector reaches every site you can open.
- Give a custom ERP connector a separate read-only account, limited to the records that the weekly brief needs.
Connected services also process data under their own terms, so the rules of the CRM and ERP vendors still apply. Read the connector’s page in the directory before you connect it.
Can AI be your chief of staff?
AI covers the reading and drafting part of the job. An assistant sorts the inbox, writes the meeting pages and the weekly brief, and keeps the list of follow-ups.
A human chief of staff also makes judgment calls and speaks for the owner with staff and customers. Those parts stay with a person, and so does the approval of every message that leaves the company.
Building yours in a working session
ThriveAI’s training runs as working sessions on your own tasks and documents. For an owner, the task can be this setup on your own mailbox and calendar. You write the instructions for the weekly brief and set every send to need your approval.
Sessions run on site at your plant or office, in French or English, and Derik Lawlis leads every ThriveAI project. Each engagement is planned and quoted separately, around the jobs you choose and the people who take part. For the other jobs an owner hands to AI, see AI training for executives.