Skip to content

Practical guide · October 30, 2026

Manage e-mail and follow-ups with an assistant that asks before sending

How freelancers, consultants and agents can let an AI assistant find the replies they owe and draft them, while every message still waits for their approval. Setup, the approval flow and its limits.

The follow-up problem#

If you answer your own e-mail, the expensive failure is not a slow reply. It is the reply you forgot: the quote you promised on Friday, the viewing you said you would confirm, the invoice query that sat for a week. The messages are in your inbox, but nothing tells you which of them are waiting for you.

An assistant can find those and draft the answer. The risk is the opposite one: an assistant that sends the wrong thing from your address. Analysts see the same risk at company scale. Gartner predicted in June 2025 that over 40 percent of agentic AI projects will be cancelled by the end of 2027 because of cost, unclear value or inadequate risk controls (Gartner press release). For one person, the practical control is simple: the assistant drafts, you decide.

The goal is not to answer faster at any cost. It is to know what is waiting, to spend your time on the parts that need judgement, and to keep the final decision with you.

How Growth OX handles it#

  • Reading. Growth OX reads the e-mail of each project twice a day through the claude.ai Gmail connector. The connection is read only: it does not send, delete or label anything. Replies that wait for you appear in Today.
  • Drafting. On a card, the Do it (Yap) button asks the assistant to read the card's sources and prepare the reply.
  • Approving. The prepared action is shown in full, for example To, Subject and Text. You edit it, or ask the assistant for a change such as "make it shorter". Nothing runs until you approve.
  • Warning. Before you approve, the card says that a sent mail cannot be taken back.

If the app cannot run a step itself, it says so and offers an Open button, so you do the last step in your own mail program. See Approvals for the full flow.

Set it up#

  1. Connect Claude. Install Claude Code and sign in to it yourself; Growth OX does not sign in for you. Use the subscription sign-in, because a paid API key sign-in is not used for this. See Connect a model.
  2. Connect Gmail. Use the Gmail (with Claude) card in Connections. In claude.ai, find Gmail in the connector list and choose Connect. When the button reads Disconnect, the connection is complete.
  3. Tell it which mail belongs to which project. Growth OX matches mail to a project by the project name and website. You can add an address or word under e-mail forwarding.
  4. Turn on the suggested approval rules, so requests reach you as cards. With no rules, the assistant's requests are refused and nothing happens.
  5. Open Today each morning. Work through the cards that wait for you.

What an approval card shows#

Fields on an approval card
FieldWhat it tells you
What it acts onThe message, post or order the action concerns
Asked byThe assistant, an automation or you
Estimated costA figure in USD, or "Unknown"
What will be done, in fullThe whole content, not a summary
Asked, ExpiresWhen the request was made and when it lapses

After a decision, the Record in Approvals lists it as approved, refused or expired, so you can see later what was sent and who asked for it.

What it looks like in different jobs#

  • Consultant. A client asked for a revised scope on Tuesday. The reply waits in Today. You press Do it, read the draft, correct the figures yourself and approve.
  • Real-estate agent. A buyer asked to move a viewing. The assistant drafts a confirmation with the new time; you check the time and the address before approving.
  • Shop owner. A customer asks where an order is. The assistant drafts a short reply; you add the tracking detail it could not know, then approve.

These are examples of the flow, not results. The assistant only knows what is in the card's sources, so anything it could not have read, such as a phone call you had, you add yourself.

What to check before you approve#

  1. The recipient in the To field is the person you mean.
  2. Every figure, date and name in the text is correct.
  3. The tone fits the relationship. Edit or ask the assistant to revise it.
  4. Nothing in the text is a promise you did not decide to make.

Rules that keep it calm#

Approval rules decide which kinds of request reach you. Add the suggested rules first, then tighten them. A rule is either Ask me or Always refuse, so you can refuse a whole kind of work, such as spending, and keep only e-mail and posts as cards.

If you paired a phone with the Can approve role, the same cards are on the phone, and approving there asks for your fingerprint or screen lock. Look at the Record in Approvals now and then: it shows what was approved, refused or expired.

Where the text of your e-mail goes#

Mail is read through the claude.ai Gmail connector, so Claude's terms apply to that reading. When the assistant drafts a reply, the text of the job goes to the model you assigned to drafts: Claude, Codex, an API provider, or a local model on your Mac. Growth OX itself never receives it. If a mailbox holds client confidences, assign drafts to a local model; see Privacy.

Limits to know about#

  • Gmail works through Claude. Gmail is not supported for Codex yet.
  • Mail is read twice a day, not continuously. Do not rely on it for same-minute alerts.
  • A drafted reply is a draft. It can be wrong or have the wrong tone. Read it before you approve.
  • Where a provider is involved, the text of the job goes to that provider. A local model keeps it on your Mac but is weaker; see Privacy.
  • Some sends cannot be undone. The Undo timer on the card helps for other actions; a sent e-mail stays sent.

Questions#

Can the assistant send e-mail without asking me?
Not with the default settings. E-mails wait for your approval, and without approval rules the assistant's requests are refused. Standing rules you write yourself change this for the e-mails inside them.
Does Growth OX change my Gmail?
The connector is read only. It does not send, delete or label anything.
How often does it check my mail?
Twice a day for each project. Waiting replies appear in Today.
Can I edit the draft?
Yes. Edit any field yourself, or ask the assistant to revise it, then approve the edited version.
Does it work with Outlook or other mail?
Gmail through Claude is the supported e-mail source today. Other mail providers are not covered by this article.
Can I take back a sent e-mail?
No. The card warns you before you approve.
Can I approve from my phone?
Yes, if you paired a phone with the Can approve role. Approving on the phone asks for your fingerprint or screen lock.

Sources#